[dev.simd] cmd/compile: add masked merging ops and optimizations

This CL generates optimizations for masked variant of AVX512 instructions for patterns: x.Op(y).Merge(z, mask) => OpMasked(z, x, y mask), where OpMasked is resultInArg0. Change-Id: Ife7ccc9ddbf76ae921a085bd6a42b965da9bc179 Reviewed-on: https://go-review.googlesource.com/c/go/+/718160 Reviewed-by: David Chase <drchase@google.com> TryBot-Bypass: Junyang Shao <shaojunyang@google.com>
author: Junyang Shao <shaojunyang@google.com> 2025-11-05 19:25:00 +0000
committer: Junyang Shao <shaojunyang@google.com> 2025-11-11 13:34:39 -0800
commit: 86b4fe31d9b7fe4b249a3a8007290305eaa4f16a (patch)
tree: 23f7b2fed426e41c055c39d864835012e602119c /src/cmd/compile
parent: 771a1dc216ff02dd23c78ada35a207a363690d11 (diff)
download: go-86b4fe31d9b7fe4b249a3a8007290305eaa4f16a.tar.xz
7 files changed, 17198 insertions, 594 deletions
diff --git a/src/cmd/compile/internal/amd64/simdssa.go b/src/cmd/compile/internal/amd64/simdssa.go
index 4f5cacea02..0abcd95e37 100644
--- a/src/cmd/compile/internal/amd64/simdssa.go
+++ b/src/cmd/compile/internal/amd64/simdssa.go
@@ -914,12 +914,6 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VSQRTPDMasked128,
 		ssa.OpAMD64VSQRTPDMasked256,
 		ssa.OpAMD64VSQRTPDMasked512,
-		ssa.OpAMD64VMOVUPSMasked128,
-		ssa.OpAMD64VMOVUPSMasked256,
-		ssa.OpAMD64VMOVUPSMasked512,
-		ssa.OpAMD64VMOVUPDMasked128,
-		ssa.OpAMD64VMOVUPDMasked256,
-		ssa.OpAMD64VMOVUPDMasked512,
 		ssa.OpAMD64VMOVDQU8Masked128,
 		ssa.OpAMD64VMOVDQU8Masked256,
 		ssa.OpAMD64VMOVDQU8Masked512,
@@ -1225,6 +1219,129 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPDPBUSDSMasked128,
 		ssa.OpAMD64VPDPBUSDSMasked256,
 		ssa.OpAMD64VPDPBUSDSMasked512,
+		ssa.OpAMD64VADDPSMasked128Merging,
+		ssa.OpAMD64VADDPSMasked256Merging,
+		ssa.OpAMD64VADDPSMasked512Merging,
+		ssa.OpAMD64VADDPDMasked128Merging,
+		ssa.OpAMD64VADDPDMasked256Merging,
+		ssa.OpAMD64VADDPDMasked512Merging,
+		ssa.OpAMD64VPADDBMasked128Merging,
+		ssa.OpAMD64VPADDBMasked256Merging,
+		ssa.OpAMD64VPADDBMasked512Merging,
+		ssa.OpAMD64VPADDWMasked128Merging,
+		ssa.OpAMD64VPADDWMasked256Merging,
+		ssa.OpAMD64VPADDWMasked512Merging,
+		ssa.OpAMD64VPADDDMasked128Merging,
+		ssa.OpAMD64VPADDDMasked256Merging,
+		ssa.OpAMD64VPADDDMasked512Merging,
+		ssa.OpAMD64VPADDQMasked128Merging,
+		ssa.OpAMD64VPADDQMasked256Merging,
+		ssa.OpAMD64VPADDQMasked512Merging,
+		ssa.OpAMD64VPADDSBMasked128Merging,
+		ssa.OpAMD64VPADDSBMasked256Merging,
+		ssa.OpAMD64VPADDSBMasked512Merging,
+		ssa.OpAMD64VPADDSWMasked128Merging,
+		ssa.OpAMD64VPADDSWMasked256Merging,
+		ssa.OpAMD64VPADDSWMasked512Merging,
+		ssa.OpAMD64VPADDUSBMasked128Merging,
+		ssa.OpAMD64VPADDUSBMasked256Merging,
+		ssa.OpAMD64VPADDUSBMasked512Merging,
+		ssa.OpAMD64VPADDUSWMasked128Merging,
+		ssa.OpAMD64VPADDUSWMasked256Merging,
+		ssa.OpAMD64VPADDUSWMasked512Merging,
+		ssa.OpAMD64VPANDDMasked128Merging,
+		ssa.OpAMD64VPANDDMasked256Merging,
+		ssa.OpAMD64VPANDDMasked512Merging,
+		ssa.OpAMD64VPANDQMasked128Merging,
+		ssa.OpAMD64VPANDQMasked256Merging,
+		ssa.OpAMD64VPANDQMasked512Merging,
+		ssa.OpAMD64VPAVGBMasked128Merging,
+		ssa.OpAMD64VPAVGBMasked256Merging,
+		ssa.OpAMD64VPAVGBMasked512Merging,
+		ssa.OpAMD64VPAVGWMasked128Merging,
+		ssa.OpAMD64VPAVGWMasked256Merging,
+		ssa.OpAMD64VPAVGWMasked512Merging,
+		ssa.OpAMD64VPACKSSDWMasked128Merging,
+		ssa.OpAMD64VPACKSSDWMasked256Merging,
+		ssa.OpAMD64VPACKSSDWMasked512Merging,
+		ssa.OpAMD64VPACKUSDWMasked128Merging,
+		ssa.OpAMD64VPACKUSDWMasked256Merging,
+		ssa.OpAMD64VPACKUSDWMasked512Merging,
+		ssa.OpAMD64VDIVPSMasked128Merging,
+		ssa.OpAMD64VDIVPSMasked256Merging,
+		ssa.OpAMD64VDIVPSMasked512Merging,
+		ssa.OpAMD64VDIVPDMasked128Merging,
+		ssa.OpAMD64VDIVPDMasked256Merging,
+		ssa.OpAMD64VDIVPDMasked512Merging,
+		ssa.OpAMD64VPMADDWDMasked128Merging,
+		ssa.OpAMD64VPMADDWDMasked256Merging,
+		ssa.OpAMD64VPMADDWDMasked512Merging,
+		ssa.OpAMD64VPMADDUBSWMasked128Merging,
+		ssa.OpAMD64VPMADDUBSWMasked256Merging,
+		ssa.OpAMD64VPMADDUBSWMasked512Merging,
+		ssa.OpAMD64VGF2P8MULBMasked128Merging,
+		ssa.OpAMD64VGF2P8MULBMasked256Merging,
+		ssa.OpAMD64VGF2P8MULBMasked512Merging,
+		ssa.OpAMD64VMAXPSMasked128Merging,
+		ssa.OpAMD64VMAXPSMasked256Merging,
+		ssa.OpAMD64VMAXPSMasked512Merging,
+		ssa.OpAMD64VMAXPDMasked128Merging,
+		ssa.OpAMD64VMAXPDMasked256Merging,
+		ssa.OpAMD64VMAXPDMasked512Merging,
+		ssa.OpAMD64VPMAXSBMasked128Merging,
+		ssa.OpAMD64VPMAXSBMasked256Merging,
+		ssa.OpAMD64VPMAXSBMasked512Merging,
+		ssa.OpAMD64VPMAXSWMasked128Merging,
+		ssa.OpAMD64VPMAXSWMasked256Merging,
+		ssa.OpAMD64VPMAXSWMasked512Merging,
+		ssa.OpAMD64VPMAXSDMasked128Merging,
+		ssa.OpAMD64VPMAXSDMasked256Merging,
+		ssa.OpAMD64VPMAXSDMasked512Merging,
+		ssa.OpAMD64VPMAXSQMasked128Merging,
+		ssa.OpAMD64VPMAXSQMasked256Merging,
+		ssa.OpAMD64VPMAXSQMasked512Merging,
+		ssa.OpAMD64VPMAXUBMasked128Merging,
+		ssa.OpAMD64VPMAXUBMasked256Merging,
+		ssa.OpAMD64VPMAXUBMasked512Merging,
+		ssa.OpAMD64VPMAXUWMasked128Merging,
+		ssa.OpAMD64VPMAXUWMasked256Merging,
+		ssa.OpAMD64VPMAXUWMasked512Merging,
+		ssa.OpAMD64VPMAXUDMasked128Merging,
+		ssa.OpAMD64VPMAXUDMasked256Merging,
+		ssa.OpAMD64VPMAXUDMasked512Merging,
+		ssa.OpAMD64VPMAXUQMasked128Merging,
+		ssa.OpAMD64VPMAXUQMasked256Merging,
+		ssa.OpAMD64VPMAXUQMasked512Merging,
+		ssa.OpAMD64VMINPSMasked128Merging,
+		ssa.OpAMD64VMINPSMasked256Merging,
+		ssa.OpAMD64VMINPSMasked512Merging,
+		ssa.OpAMD64VMINPDMasked128Merging,
+		ssa.OpAMD64VMINPDMasked256Merging,
+		ssa.OpAMD64VMINPDMasked512Merging,
+		ssa.OpAMD64VPMINSBMasked128Merging,
+		ssa.OpAMD64VPMINSBMasked256Merging,
+		ssa.OpAMD64VPMINSBMasked512Merging,
+		ssa.OpAMD64VPMINSWMasked128Merging,
+		ssa.OpAMD64VPMINSWMasked256Merging,
+		ssa.OpAMD64VPMINSWMasked512Merging,
+		ssa.OpAMD64VPMINSDMasked128Merging,
+		ssa.OpAMD64VPMINSDMasked256Merging,
+		ssa.OpAMD64VPMINSDMasked512Merging,
+		ssa.OpAMD64VPMINSQMasked128Merging,
+		ssa.OpAMD64VPMINSQMasked256Merging,
+		ssa.OpAMD64VPMINSQMasked512Merging,
+		ssa.OpAMD64VPMINUBMasked128Merging,
+		ssa.OpAMD64VPMINUBMasked256Merging,
+		ssa.OpAMD64VPMINUBMasked512Merging,
+		ssa.OpAMD64VPMINUWMasked128Merging,
+		ssa.OpAMD64VPMINUWMasked256Merging,
+		ssa.OpAMD64VPMINUWMasked512Merging,
+		ssa.OpAMD64VPMINUDMasked128Merging,
+		ssa.OpAMD64VPMINUDMasked256Merging,
+		ssa.OpAMD64VPMINUDMasked512Merging,
+		ssa.OpAMD64VPMINUQMasked128Merging,
+		ssa.OpAMD64VPMINUQMasked256Merging,
+		ssa.OpAMD64VPMINUQMasked512Merging,
 		ssa.OpAMD64VFMADD213PSMasked128,
 		ssa.OpAMD64VFMADD213PSMasked256,
 		ssa.OpAMD64VFMADD213PSMasked512,
@@ -1237,12 +1354,39 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VFMADDSUB213PDMasked128,
 		ssa.OpAMD64VFMADDSUB213PDMasked256,
 		ssa.OpAMD64VFMADDSUB213PDMasked512,
+		ssa.OpAMD64VPMULHWMasked128Merging,
+		ssa.OpAMD64VPMULHWMasked256Merging,
+		ssa.OpAMD64VPMULHWMasked512Merging,
+		ssa.OpAMD64VPMULHUWMasked128Merging,
+		ssa.OpAMD64VPMULHUWMasked256Merging,
+		ssa.OpAMD64VPMULHUWMasked512Merging,
+		ssa.OpAMD64VMULPSMasked128Merging,
+		ssa.OpAMD64VMULPSMasked256Merging,
+		ssa.OpAMD64VMULPSMasked512Merging,
+		ssa.OpAMD64VMULPDMasked128Merging,
+		ssa.OpAMD64VMULPDMasked256Merging,
+		ssa.OpAMD64VMULPDMasked512Merging,
+		ssa.OpAMD64VPMULLWMasked128Merging,
+		ssa.OpAMD64VPMULLWMasked256Merging,
+		ssa.OpAMD64VPMULLWMasked512Merging,
+		ssa.OpAMD64VPMULLDMasked128Merging,
+		ssa.OpAMD64VPMULLDMasked256Merging,
+		ssa.OpAMD64VPMULLDMasked512Merging,
+		ssa.OpAMD64VPMULLQMasked128Merging,
+		ssa.OpAMD64VPMULLQMasked256Merging,
+		ssa.OpAMD64VPMULLQMasked512Merging,
 		ssa.OpAMD64VFMSUBADD213PSMasked128,
 		ssa.OpAMD64VFMSUBADD213PSMasked256,
 		ssa.OpAMD64VFMSUBADD213PSMasked512,
 		ssa.OpAMD64VFMSUBADD213PDMasked128,
 		ssa.OpAMD64VFMSUBADD213PDMasked256,
 		ssa.OpAMD64VFMSUBADD213PDMasked512,
+		ssa.OpAMD64VPORDMasked128Merging,
+		ssa.OpAMD64VPORDMasked256Merging,
+		ssa.OpAMD64VPORDMasked512Merging,
+		ssa.OpAMD64VPORQMasked128Merging,
+		ssa.OpAMD64VPORQMasked256Merging,
+		ssa.OpAMD64VPORQMasked512Merging,
 		ssa.OpAMD64VPERMI2BMasked128,
 		ssa.OpAMD64VPERMI2BMasked256,
 		ssa.OpAMD64VPERMI2BMasked512,
@@ -1261,6 +1405,45 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPERMI2QMasked256,
 		ssa.OpAMD64VPERMI2PDMasked512,
 		ssa.OpAMD64VPERMI2QMasked512,
+		ssa.OpAMD64VPSHUFBMasked256Merging,
+		ssa.OpAMD64VPSHUFBMasked512Merging,
+		ssa.OpAMD64VPSHUFBMasked128Merging,
+		ssa.OpAMD64VPROLVDMasked128Merging,
+		ssa.OpAMD64VPROLVDMasked256Merging,
+		ssa.OpAMD64VPROLVDMasked512Merging,
+		ssa.OpAMD64VPROLVQMasked128Merging,
+		ssa.OpAMD64VPROLVQMasked256Merging,
+		ssa.OpAMD64VPROLVQMasked512Merging,
+		ssa.OpAMD64VPRORVDMasked128Merging,
+		ssa.OpAMD64VPRORVDMasked256Merging,
+		ssa.OpAMD64VPRORVDMasked512Merging,
+		ssa.OpAMD64VPRORVQMasked128Merging,
+		ssa.OpAMD64VPRORVQMasked256Merging,
+		ssa.OpAMD64VPRORVQMasked512Merging,
+		ssa.OpAMD64VSCALEFPSMasked128Merging,
+		ssa.OpAMD64VSCALEFPSMasked256Merging,
+		ssa.OpAMD64VSCALEFPSMasked512Merging,
+		ssa.OpAMD64VSCALEFPDMasked128Merging,
+		ssa.OpAMD64VSCALEFPDMasked256Merging,
+		ssa.OpAMD64VSCALEFPDMasked512Merging,
+		ssa.OpAMD64VPSHLDWMasked128Merging,
+		ssa.OpAMD64VPSHLDWMasked256Merging,
+		ssa.OpAMD64VPSHLDWMasked512Merging,
+		ssa.OpAMD64VPSHLDDMasked128Merging,
+		ssa.OpAMD64VPSHLDDMasked256Merging,
+		ssa.OpAMD64VPSHLDDMasked512Merging,
+		ssa.OpAMD64VPSHLDQMasked128Merging,
+		ssa.OpAMD64VPSHLDQMasked256Merging,
+		ssa.OpAMD64VPSHLDQMasked512Merging,
+		ssa.OpAMD64VPSHRDWMasked128Merging,
+		ssa.OpAMD64VPSHRDWMasked256Merging,
+		ssa.OpAMD64VPSHRDWMasked512Merging,
+		ssa.OpAMD64VPSHRDDMasked128Merging,
+		ssa.OpAMD64VPSHRDDMasked256Merging,
+		ssa.OpAMD64VPSHRDDMasked512Merging,
+		ssa.OpAMD64VPSHRDQMasked128Merging,
+		ssa.OpAMD64VPSHRDQMasked256Merging,
+		ssa.OpAMD64VPSHRDQMasked512Merging,
 		ssa.OpAMD64VPSHLDVWMasked128,
 		ssa.OpAMD64VPSHLDVWMasked256,
 		ssa.OpAMD64VPSHLDVWMasked512,
@@ -1270,6 +1453,15 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPSHLDVQMasked128,
 		ssa.OpAMD64VPSHLDVQMasked256,
 		ssa.OpAMD64VPSHLDVQMasked512,
+		ssa.OpAMD64VPSLLVWMasked128Merging,
+		ssa.OpAMD64VPSLLVWMasked256Merging,
+		ssa.OpAMD64VPSLLVWMasked512Merging,
+		ssa.OpAMD64VPSLLVDMasked128Merging,
+		ssa.OpAMD64VPSLLVDMasked256Merging,
+		ssa.OpAMD64VPSLLVDMasked512Merging,
+		ssa.OpAMD64VPSLLVQMasked128Merging,
+		ssa.OpAMD64VPSLLVQMasked256Merging,
+		ssa.OpAMD64VPSLLVQMasked512Merging,
 		ssa.OpAMD64VPSHRDVWMasked128,
 		ssa.OpAMD64VPSHRDVWMasked256,
 		ssa.OpAMD64VPSHRDVWMasked512,
@@ -1278,7 +1470,61 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPSHRDVDMasked512,
 		ssa.OpAMD64VPSHRDVQMasked128,
 		ssa.OpAMD64VPSHRDVQMasked256,
-		ssa.OpAMD64VPSHRDVQMasked512:
+		ssa.OpAMD64VPSHRDVQMasked512,
+		ssa.OpAMD64VPSRAVWMasked128Merging,
+		ssa.OpAMD64VPSRAVWMasked256Merging,
+		ssa.OpAMD64VPSRAVWMasked512Merging,
+		ssa.OpAMD64VPSRAVDMasked128Merging,
+		ssa.OpAMD64VPSRAVDMasked256Merging,
+		ssa.OpAMD64VPSRAVDMasked512Merging,
+		ssa.OpAMD64VPSRAVQMasked128Merging,
+		ssa.OpAMD64VPSRAVQMasked256Merging,
+		ssa.OpAMD64VPSRAVQMasked512Merging,
+		ssa.OpAMD64VPSRLVWMasked128Merging,
+		ssa.OpAMD64VPSRLVWMasked256Merging,
+		ssa.OpAMD64VPSRLVWMasked512Merging,
+		ssa.OpAMD64VPSRLVDMasked128Merging,
+		ssa.OpAMD64VPSRLVDMasked256Merging,
+		ssa.OpAMD64VPSRLVDMasked512Merging,
+		ssa.OpAMD64VPSRLVQMasked128Merging,
+		ssa.OpAMD64VPSRLVQMasked256Merging,
+		ssa.OpAMD64VPSRLVQMasked512Merging,
+		ssa.OpAMD64VSUBPSMasked128Merging,
+		ssa.OpAMD64VSUBPSMasked256Merging,
+		ssa.OpAMD64VSUBPSMasked512Merging,
+		ssa.OpAMD64VSUBPDMasked128Merging,
+		ssa.OpAMD64VSUBPDMasked256Merging,
+		ssa.OpAMD64VSUBPDMasked512Merging,
+		ssa.OpAMD64VPSUBBMasked128Merging,
+		ssa.OpAMD64VPSUBBMasked256Merging,
+		ssa.OpAMD64VPSUBBMasked512Merging,
+		ssa.OpAMD64VPSUBWMasked128Merging,
+		ssa.OpAMD64VPSUBWMasked256Merging,
+		ssa.OpAMD64VPSUBWMasked512Merging,
+		ssa.OpAMD64VPSUBDMasked128Merging,
+		ssa.OpAMD64VPSUBDMasked256Merging,
+		ssa.OpAMD64VPSUBDMasked512Merging,
+		ssa.OpAMD64VPSUBQMasked128Merging,
+		ssa.OpAMD64VPSUBQMasked256Merging,
+		ssa.OpAMD64VPSUBQMasked512Merging,
+		ssa.OpAMD64VPSUBSBMasked128Merging,
+		ssa.OpAMD64VPSUBSBMasked256Merging,
+		ssa.OpAMD64VPSUBSBMasked512Merging,
+		ssa.OpAMD64VPSUBSWMasked128Merging,
+		ssa.OpAMD64VPSUBSWMasked256Merging,
+		ssa.OpAMD64VPSUBSWMasked512Merging,
+		ssa.OpAMD64VPSUBUSBMasked128Merging,
+		ssa.OpAMD64VPSUBUSBMasked256Merging,
+		ssa.OpAMD64VPSUBUSBMasked512Merging,
+		ssa.OpAMD64VPSUBUSWMasked128Merging,
+		ssa.OpAMD64VPSUBUSWMasked256Merging,
+		ssa.OpAMD64VPSUBUSWMasked512Merging,
+		ssa.OpAMD64VPXORDMasked128Merging,
+		ssa.OpAMD64VPXORDMasked256Merging,
+		ssa.OpAMD64VPXORDMasked512Merging,
+		ssa.OpAMD64VPXORQMasked128Merging,
+		ssa.OpAMD64VPXORQMasked256Merging,
+		ssa.OpAMD64VPXORQMasked512Merging:
 		p = simdV3kvResultInArg0(s, v)
 
 	case ssa.OpAMD64VPSLLW128,
@@ -1979,6 +2225,199 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 	case ssa.OpAMD64SHA256RNDS2128:
 		p = simdV31x0AtIn2ResultInArg0(s, v)
 
+	case ssa.OpAMD64VPABSBMasked128Merging,
+		ssa.OpAMD64VPABSBMasked256Merging,
+		ssa.OpAMD64VPABSBMasked512Merging,
+		ssa.OpAMD64VPABSWMasked128Merging,
+		ssa.OpAMD64VPABSWMasked256Merging,
+		ssa.OpAMD64VPABSWMasked512Merging,
+		ssa.OpAMD64VPABSDMasked128Merging,
+		ssa.OpAMD64VPABSDMasked256Merging,
+		ssa.OpAMD64VPABSDMasked512Merging,
+		ssa.OpAMD64VPABSQMasked128Merging,
+		ssa.OpAMD64VPABSQMasked256Merging,
+		ssa.OpAMD64VPABSQMasked512Merging,
+		ssa.OpAMD64VBROADCASTSSMasked128Merging,
+		ssa.OpAMD64VPBROADCASTQMasked128Merging,
+		ssa.OpAMD64VPBROADCASTBMasked128Merging,
+		ssa.OpAMD64VPBROADCASTWMasked128Merging,
+		ssa.OpAMD64VPBROADCASTDMasked128Merging,
+		ssa.OpAMD64VBROADCASTSSMasked256Merging,
+		ssa.OpAMD64VBROADCASTSDMasked256Merging,
+		ssa.OpAMD64VPBROADCASTBMasked256Merging,
+		ssa.OpAMD64VPBROADCASTWMasked256Merging,
+		ssa.OpAMD64VPBROADCASTDMasked256Merging,
+		ssa.OpAMD64VPBROADCASTQMasked256Merging,
+		ssa.OpAMD64VBROADCASTSSMasked512Merging,
+		ssa.OpAMD64VBROADCASTSDMasked512Merging,
+		ssa.OpAMD64VPBROADCASTBMasked512Merging,
+		ssa.OpAMD64VPBROADCASTWMasked512Merging,
+		ssa.OpAMD64VPBROADCASTDMasked512Merging,
+		ssa.OpAMD64VPBROADCASTQMasked512Merging,
+		ssa.OpAMD64VRNDSCALEPSMasked128Merging,
+		ssa.OpAMD64VRNDSCALEPSMasked256Merging,
+		ssa.OpAMD64VRNDSCALEPSMasked512Merging,
+		ssa.OpAMD64VRNDSCALEPDMasked128Merging,
+		ssa.OpAMD64VRNDSCALEPDMasked256Merging,
+		ssa.OpAMD64VRNDSCALEPDMasked512Merging,
+		ssa.OpAMD64VREDUCEPSMasked128Merging,
+		ssa.OpAMD64VREDUCEPSMasked256Merging,
+		ssa.OpAMD64VREDUCEPSMasked512Merging,
+		ssa.OpAMD64VREDUCEPDMasked128Merging,
+		ssa.OpAMD64VREDUCEPDMasked256Merging,
+		ssa.OpAMD64VREDUCEPDMasked512Merging,
+		ssa.OpAMD64VPMOVWBMasked128Merging,
+		ssa.OpAMD64VPMOVWBMasked256Merging,
+		ssa.OpAMD64VPMOVDBMasked128Merging,
+		ssa.OpAMD64VPMOVQBMasked128Merging,
+		ssa.OpAMD64VPMOVSWBMasked128Merging,
+		ssa.OpAMD64VPMOVSWBMasked256Merging,
+		ssa.OpAMD64VPMOVSDBMasked128Merging,
+		ssa.OpAMD64VPMOVSQBMasked128Merging,
+		ssa.OpAMD64VPMOVSXBWMasked256Merging,
+		ssa.OpAMD64VPMOVSXBWMasked512Merging,
+		ssa.OpAMD64VPMOVDWMasked128Merging,
+		ssa.OpAMD64VPMOVDWMasked256Merging,
+		ssa.OpAMD64VPMOVQWMasked128Merging,
+		ssa.OpAMD64VPMOVSDWMasked128Merging,
+		ssa.OpAMD64VPMOVSDWMasked256Merging,
+		ssa.OpAMD64VPMOVSQWMasked128Merging,
+		ssa.OpAMD64VPMOVSXBWMasked128Merging,
+		ssa.OpAMD64VCVTTPS2DQMasked128Merging,
+		ssa.OpAMD64VCVTTPS2DQMasked256Merging,
+		ssa.OpAMD64VCVTTPS2DQMasked512Merging,
+		ssa.OpAMD64VPMOVSXBDMasked512Merging,
+		ssa.OpAMD64VPMOVSXWDMasked256Merging,
+		ssa.OpAMD64VPMOVSXWDMasked512Merging,
+		ssa.OpAMD64VPMOVQDMasked128Merging,
+		ssa.OpAMD64VPMOVQDMasked256Merging,
+		ssa.OpAMD64VPMOVSQDMasked128Merging,
+		ssa.OpAMD64VPMOVSQDMasked256Merging,
+		ssa.OpAMD64VPMOVSXBDMasked128Merging,
+		ssa.OpAMD64VPMOVSXWDMasked128Merging,
+		ssa.OpAMD64VPMOVSXBDMasked256Merging,
+		ssa.OpAMD64VPMOVSXWQMasked512Merging,
+		ssa.OpAMD64VPMOVSXDQMasked256Merging,
+		ssa.OpAMD64VPMOVSXDQMasked512Merging,
+		ssa.OpAMD64VPMOVSXBQMasked128Merging,
+		ssa.OpAMD64VPMOVSXWQMasked128Merging,
+		ssa.OpAMD64VPMOVSXDQMasked128Merging,
+		ssa.OpAMD64VPMOVSXBQMasked256Merging,
+		ssa.OpAMD64VPMOVSXBQMasked512Merging,
+		ssa.OpAMD64VPMOVUSWBMasked128Merging,
+		ssa.OpAMD64VPMOVUSWBMasked256Merging,
+		ssa.OpAMD64VPMOVUSDBMasked128Merging,
+		ssa.OpAMD64VPMOVUSQBMasked128Merging,
+		ssa.OpAMD64VPMOVZXBWMasked256Merging,
+		ssa.OpAMD64VPMOVZXBWMasked512Merging,
+		ssa.OpAMD64VPMOVUSDWMasked128Merging,
+		ssa.OpAMD64VPMOVUSDWMasked256Merging,
+		ssa.OpAMD64VPMOVUSQWMasked128Merging,
+		ssa.OpAMD64VPMOVZXBWMasked128Merging,
+		ssa.OpAMD64VCVTPS2UDQMasked128Merging,
+		ssa.OpAMD64VCVTPS2UDQMasked256Merging,
+		ssa.OpAMD64VCVTPS2UDQMasked512Merging,
+		ssa.OpAMD64VPMOVZXBDMasked512Merging,
+		ssa.OpAMD64VPMOVZXWDMasked256Merging,
+		ssa.OpAMD64VPMOVZXWDMasked512Merging,
+		ssa.OpAMD64VPMOVUSQDMasked128Merging,
+		ssa.OpAMD64VPMOVUSQDMasked256Merging,
+		ssa.OpAMD64VPMOVZXBDMasked128Merging,
+		ssa.OpAMD64VPMOVZXWDMasked128Merging,
+		ssa.OpAMD64VPMOVZXBDMasked256Merging,
+		ssa.OpAMD64VPMOVZXWQMasked512Merging,
+		ssa.OpAMD64VPMOVZXDQMasked256Merging,
+		ssa.OpAMD64VPMOVZXDQMasked512Merging,
+		ssa.OpAMD64VPMOVZXBQMasked128Merging,
+		ssa.OpAMD64VPMOVZXWQMasked128Merging,
+		ssa.OpAMD64VPMOVZXDQMasked128Merging,
+		ssa.OpAMD64VPMOVSXWQMasked256Merging,
+		ssa.OpAMD64VPMOVZXBQMasked256Merging,
+		ssa.OpAMD64VPMOVZXWQMasked256Merging,
+		ssa.OpAMD64VPMOVZXBQMasked512Merging,
+		ssa.OpAMD64VPLZCNTDMasked128Merging,
+		ssa.OpAMD64VPLZCNTDMasked256Merging,
+		ssa.OpAMD64VPLZCNTDMasked512Merging,
+		ssa.OpAMD64VPLZCNTQMasked128Merging,
+		ssa.OpAMD64VPLZCNTQMasked256Merging,
+		ssa.OpAMD64VPLZCNTQMasked512Merging,
+		ssa.OpAMD64VPOPCNTBMasked128Merging,
+		ssa.OpAMD64VPOPCNTBMasked256Merging,
+		ssa.OpAMD64VPOPCNTBMasked512Merging,
+		ssa.OpAMD64VPOPCNTWMasked128Merging,
+		ssa.OpAMD64VPOPCNTWMasked256Merging,
+		ssa.OpAMD64VPOPCNTWMasked512Merging,
+		ssa.OpAMD64VPOPCNTDMasked128Merging,
+		ssa.OpAMD64VPOPCNTDMasked256Merging,
+		ssa.OpAMD64VPOPCNTDMasked512Merging,
+		ssa.OpAMD64VPOPCNTQMasked128Merging,
+		ssa.OpAMD64VPOPCNTQMasked256Merging,
+		ssa.OpAMD64VPOPCNTQMasked512Merging,
+		ssa.OpAMD64VPSHUFDMasked256Merging,
+		ssa.OpAMD64VPSHUFDMasked512Merging,
+		ssa.OpAMD64VPSHUFHWMasked256Merging,
+		ssa.OpAMD64VPSHUFHWMasked512Merging,
+		ssa.OpAMD64VPSHUFHWMasked128Merging,
+		ssa.OpAMD64VPSHUFDMasked128Merging,
+		ssa.OpAMD64VRCP14PSMasked128Merging,
+		ssa.OpAMD64VRCP14PSMasked256Merging,
+		ssa.OpAMD64VRCP14PSMasked512Merging,
+		ssa.OpAMD64VRCP14PDMasked128Merging,
+		ssa.OpAMD64VRCP14PDMasked256Merging,
+		ssa.OpAMD64VRCP14PDMasked512Merging,
+		ssa.OpAMD64VRSQRT14PSMasked128Merging,
+		ssa.OpAMD64VRSQRT14PSMasked256Merging,
+		ssa.OpAMD64VRSQRT14PSMasked512Merging,
+		ssa.OpAMD64VRSQRT14PDMasked128Merging,
+		ssa.OpAMD64VRSQRT14PDMasked256Merging,
+		ssa.OpAMD64VRSQRT14PDMasked512Merging,
+		ssa.OpAMD64VPROLDMasked128Merging,
+		ssa.OpAMD64VPROLDMasked256Merging,
+		ssa.OpAMD64VPROLDMasked512Merging,
+		ssa.OpAMD64VPROLQMasked128Merging,
+		ssa.OpAMD64VPROLQMasked256Merging,
+		ssa.OpAMD64VPROLQMasked512Merging,
+		ssa.OpAMD64VPRORDMasked128Merging,
+		ssa.OpAMD64VPRORDMasked256Merging,
+		ssa.OpAMD64VPRORDMasked512Merging,
+		ssa.OpAMD64VPRORQMasked128Merging,
+		ssa.OpAMD64VPRORQMasked256Merging,
+		ssa.OpAMD64VPRORQMasked512Merging,
+		ssa.OpAMD64VSQRTPSMasked128Merging,
+		ssa.OpAMD64VSQRTPSMasked256Merging,
+		ssa.OpAMD64VSQRTPSMasked512Merging,
+		ssa.OpAMD64VSQRTPDMasked128Merging,
+		ssa.OpAMD64VSQRTPDMasked256Merging,
+		ssa.OpAMD64VSQRTPDMasked512Merging,
+		ssa.OpAMD64VPSLLWMasked128constMerging,
+		ssa.OpAMD64VPSLLWMasked256constMerging,
+		ssa.OpAMD64VPSLLWMasked512constMerging,
+		ssa.OpAMD64VPSLLDMasked128constMerging,
+		ssa.OpAMD64VPSLLDMasked256constMerging,
+		ssa.OpAMD64VPSLLDMasked512constMerging,
+		ssa.OpAMD64VPSLLQMasked128constMerging,
+		ssa.OpAMD64VPSLLQMasked256constMerging,
+		ssa.OpAMD64VPSLLQMasked512constMerging,
+		ssa.OpAMD64VPSRLWMasked128constMerging,
+		ssa.OpAMD64VPSRLWMasked256constMerging,
+		ssa.OpAMD64VPSRLWMasked512constMerging,
+		ssa.OpAMD64VPSRLDMasked128constMerging,
+		ssa.OpAMD64VPSRLDMasked256constMerging,
+		ssa.OpAMD64VPSRLDMasked512constMerging,
+		ssa.OpAMD64VPSRLQMasked128constMerging,
+		ssa.OpAMD64VPSRLQMasked256constMerging,
+		ssa.OpAMD64VPSRLQMasked512constMerging,
+		ssa.OpAMD64VPSRAWMasked128constMerging,
+		ssa.OpAMD64VPSRAWMasked256constMerging,
+		ssa.OpAMD64VPSRAWMasked512constMerging,
+		ssa.OpAMD64VPSRADMasked128constMerging,
+		ssa.OpAMD64VPSRADMasked256constMerging,
+		ssa.OpAMD64VPSRADMasked512constMerging,
+		ssa.OpAMD64VPSRAQMasked128constMerging,
+		ssa.OpAMD64VPSRAQMasked256constMerging,
+		ssa.OpAMD64VPSRAQMasked512constMerging:
+		p = simdV2kvResultInArg0(s, v)
+
 	default:
 		// Unknown reg shape
 		return false
@@ -2843,12 +3282,6 @@ func ssaGenSIMDValue(s *ssagen.State, v *ssa.Value) bool {
 		ssa.OpAMD64VPXORQMasked256load,
 		ssa.OpAMD64VPXORQMasked512,
 		ssa.OpAMD64VPXORQMasked512load,
-		ssa.OpAMD64VMOVUPSMasked128,
-		ssa.OpAMD64VMOVUPSMasked256,
-		ssa.OpAMD64VMOVUPSMasked512,
-		ssa.OpAMD64VMOVUPDMasked128,
-		ssa.OpAMD64VMOVUPDMasked256,
-		ssa.OpAMD64VMOVUPDMasked512,
 		ssa.OpAMD64VMOVDQU8Masked128,
 		ssa.OpAMD64VMOVDQU8Masked256,
 		ssa.OpAMD64VMOVDQU8Masked512,
diff --git a/src/cmd/compile/internal/amd64/ssa.go b/src/cmd/compile/internal/amd64/ssa.go
index dfc282608a..a4676cd0a9 100644
--- a/src/cmd/compile/internal/amd64/ssa.go
+++ b/src/cmd/compile/internal/amd64/ssa.go
@@ -1963,6 +1963,22 @@ func simdV2kv(s *ssagen.State, v *ssa.Value) *obj.Prog {
 	return p
 }
 
+// Example instruction: VPABSB X1, X2, K3 (masking merging)
+func simdV2kvResultInArg0(s *ssagen.State, v *ssa.Value) *obj.Prog {
+	p := s.Prog(v.Op.Asm())
+	p.From.Type = obj.TYPE_REG
+	p.From.Reg = simdReg(v.Args[1])
+	// These "simd*" series of functions assumes:
+	// Any "K" register that serves as the write-mask
+	// or "predicate" for "predicated AVX512 instructions"
+	// sits right at the end of the operand list.
+	// TODO: verify this assumption.
+	p.AddRestSourceReg(maskReg(v.Args[2]))
+	p.To.Type = obj.TYPE_REG
+	p.To.Reg = simdReg(v)
+	return p
+}
+
 // This function is to accustomize the shifts.
 // The 2nd arg is an XMM, and this function merely checks that.
 // Example instruction: VPSLLQ Z1, X1, K1, Z2
diff --git a/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go b/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
index dcf452f183..1e9eb0146e 100644
--- a/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
+++ b/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
@@ -213,7 +213,7 @@ func init() {
 		vloadk  = regInfo{inputs: []regMask{gpspsb, mask, 0}, outputs: vonly}
 		vstorek = regInfo{inputs: []regMask{gpspsb, mask, v, 0}}
 
-		v11     = regInfo{inputs: vzonly, outputs: vonly}
+		v11     = regInfo{inputs: vonly, outputs: vonly}            // used in resultInArg0 ops, arg0 must not be x15
 		v21     = regInfo{inputs: []regMask{v, vz}, outputs: vonly} // used in resultInArg0 ops, arg0 must not be x15
 		vk      = regInfo{inputs: vzonly, outputs: maskonly}
 		kv      = regInfo{inputs: maskonly, outputs: vonly}
@@ -231,13 +231,13 @@ func init() {
 		gpv     = regInfo{inputs: []regMask{gp}, outputs: vonly}
 		v2flags = regInfo{inputs: []regMask{vz, vz}}
 
-		w11   = regInfo{inputs: wzonly, outputs: wonly}
+		w11   = regInfo{inputs: wonly, outputs: wonly} // used in resultInArg0 ops, arg0 must not be x15
 		w21   = regInfo{inputs: []regMask{wz, wz}, outputs: wonly}
 		wk    = regInfo{inputs: wzonly, outputs: maskonly}
 		kw    = regInfo{inputs: maskonly, outputs: wonly}
 		w2k   = regInfo{inputs: []regMask{wz, wz}, outputs: maskonly}
 		wkw   = regInfo{inputs: []regMask{wz, mask}, outputs: wonly}
-		w2kw  = regInfo{inputs: []regMask{wz, wz, mask}, outputs: wonly}
+		w2kw  = regInfo{inputs: []regMask{w, wz, mask}, outputs: wonly} // used in resultInArg0 ops, arg0 must not be x15
 		w2kk  = regInfo{inputs: []regMask{wz, wz, mask}, outputs: maskonly}
 		w31   = regInfo{inputs: []regMask{w, wz, wz}, outputs: wonly}       // used in resultInArg0 ops, arg0 must not be x15
 		w3kw  = regInfo{inputs: []regMask{w, wz, wz, mask}, outputs: wonly} // used in resultInArg0 ops, arg0 must not be x15
diff --git a/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules b/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
index b48aeecdd1..8332af2738 100644
--- a/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
@@ -1862,6 +1862,424 @@
 (VMOVDQU64Masked128 (VPSRAQ128const [a] x) mask) => (VPSRAQMasked128const [a] x mask)
 (VMOVDQU64Masked256 (VPSRAQ256const [a] x) mask) => (VPSRAQMasked256const [a] x mask)
 (VMOVDQU64Masked512 (VPSRAQ512const [a] x) mask) => (VPSRAQMasked512const [a] x mask)
+(VPBLENDMQMasked512 dst (VPSLLQ512const [a] x) mask) => (VPSLLQMasked512constMerging dst [a] x mask)
+(VPBLENDVB256 dst (VPMOVSXBW512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBWMasked512Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMOVSDW256 x) mask) => (VPMOVSDWMasked256Merging dst x mask)
+(VPBLENDMDMasked512 dst (VPLZCNTD512 x) mask) => (VPLZCNTDMasked512Merging dst x mask)
+(VPBLENDMWMasked512 dst (VPMAXSW512 x y) mask) => (VPMAXSWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMINUD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMULHW512 x y) mask) => (VPMULHWMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPMULLD512 x y) mask) => (VPMULLDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPROLQ128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLQMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMADDUBSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMADDUBSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDUSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDUSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VBROADCASTSS256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VBROADCASTSSMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXBW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBWMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINSQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VMULPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMULPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VGF2P8MULB512 x y) mask) => (VGF2P8MULBMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VMAXPS512 x y) mask) => (VMAXPSMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPOPCNTB256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTBMasked256Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VSUBPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSUBPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPSUBQ512 x y) mask) => (VPSUBQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSUBUSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBUSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXBQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBQMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMOVUSDB128 x) mask) => (VPMOVUSDBMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VPMAXUQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VRSQRT14PS512 x) mask) => (VRSQRT14PSMasked512Merging dst x mask)
+(VPBLENDVB256 dst (VPROLD256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPROLQ512 [a] x) mask) => (VPROLQMasked512Merging dst [a] x mask)
+(VPBLENDMQMasked512 dst (VPSLLVQ512 x y) mask) => (VPSLLVQMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSRAVD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VADDPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VADDPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVSXDQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXDQMasked512Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVUSWB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVUSWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXWQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWQMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMULLW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPOPCNTB512 x) mask) => (VPOPCNTBMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPSHLDQ128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDQMasked128Merging dst [a] x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRAQ256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAQMasked256constMerging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMOVDW256 x) mask) => (VPMOVDWMasked256Merging dst x mask)
+(VPBLENDMQMasked512 dst (VPMOVUSQB128 x) mask) => (VPMOVUSQBMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VCVTPS2UDQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VCVTPS2UDQMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXBQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBQMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXSQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPOPCNTW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VRCP14PS512 x) mask) => (VRCP14PSMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPBROADCASTW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMOVWB256 x) mask) => (VPMOVWBMasked256Merging dst x mask)
+(VPBLENDVB128 dst (VPRORVD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSHLDD256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDDMasked256Merging dst [a] x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLVW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRLVQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSUBUSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBUSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VREDUCEPS512 [a] x) mask) => (VREDUCEPSMasked512Merging dst [a] x mask)
+(VPBLENDVB256 dst (VPMAXSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VMINPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMINPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPADDQ512 x y) mask) => (VPADDQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VBROADCASTSD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VBROADCASTSDMasked256Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VRNDSCALEPD512 [a] x) mask) => (VRNDSCALEPDMasked512Merging dst [a] x mask)
+(VPBLENDVB128 dst (VPMOVZXDQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINSD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRAQ128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAQMasked128constMerging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VRNDSCALEPS256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRNDSCALEPSMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPACKUSDW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPACKUSDWMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMADDUBSW512 x y) mask) => (VPMADDUBSWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPLZCNTD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPLZCNTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXUD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPOPCNTB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPROLVQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPABSQ512 x) mask) => (VPABSQMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VBROADCASTSD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VBROADCASTSDMasked512Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMINPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMINPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMULHW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULHWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSHLDW512 [a] x y) mask) => (VPSHLDWMasked512Merging dst [a] x y mask)
+(VPBLENDVB128 dst (VPSHRDW128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDWMasked128Merging dst [a] x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VADDPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VADDPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXWD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWDMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXWQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWQMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VDIVPS512 x y) mask) => (VDIVPSMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VDIVPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VDIVPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPLZCNTQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPLZCNTQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSUBSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VREDUCEPD128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VREDUCEPDMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVUSQD256 x) mask) => (VPMOVUSQDMasked256Merging dst x mask)
+(VPBLENDVB128 dst (VPMOVZXBD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBDMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMULHUW512 x y) mask) => (VPMULHUWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPRORQ128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORQMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSLLVW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRLVD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPSUBSB512 x y) mask) => (VPSUBSBMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPADDD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXBW256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBWMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVSDW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINSD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VADDPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VADDPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VADDPD512 x y) mask) => (VADDPDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVSXBD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBDMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXDQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMOVUSWB256 x) mask) => (VPMOVUSWBMasked256Merging dst x mask)
+(VPBLENDVB256 dst (VPOPCNTD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPROLVD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRLVQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDUSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDUSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXSD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINUB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMULLQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VSQRTPD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSQRTPDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSUBD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VREDUCEPS256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VREDUCEPSMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMINSW512 x y) mask) => (VPMINSWMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VRCP14PD512 x) mask) => (VRCP14PDMasked512Merging dst x mask)
+(VPBLENDMWMasked512 dst (VPSRAVW512 x y) mask) => (VPSRAVWMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPSRLVD512 x y) mask) => (VPSRLVDMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPSUBD512 x y) mask) => (VPSUBDMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSUBQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTDMasked512Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVSXWD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWDMasked512Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMADDWD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMADDWDMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VGF2P8MULB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VGF2P8MULBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPROLD128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLVD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRAD128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRADMasked128constMerging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRLVW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSUBUSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBUSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDUSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDUSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXBW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBWMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXDQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXDQMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPROLVQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPROLD512 [a] x) mask) => (VPROLDMasked512Merging dst [a] x mask)
+(VPBLENDMQMasked512 dst (VPSRLVQ512 x y) mask) => (VPSRLVQMasked512Merging dst x y mask)
+(VPBLENDMBMasked512 dst (VPSUBB512 x y) mask) => (VPSUBBMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPADDW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDUSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDUSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VRNDSCALEPS128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRNDSCALEPSMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VREDUCEPD256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VREDUCEPDMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINUW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPORD512 x y) mask) => (VPORDMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VRNDSCALEPD256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRNDSCALEPDMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMULLD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSHUFB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPRORD128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPRORVD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPRORVQ512 x y) mask) => (VPRORVQMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSHLDW256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDWMasked256Merging dst [a] x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VCVTTPS2DQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VCVTTPS2DQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VCVTTPS2DQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VCVTTPS2DQMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMINPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMINPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPSHLDD512 [a] x y) mask) => (VPSHLDDMasked512Merging dst [a] x y mask)
+(VPBLENDMQMasked512 dst (VPSRAVQ512 x y) mask) => (VPSRAVQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VSUBPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSUBPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VSUBPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSUBPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSUBD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPADDW512 x y) mask) => (VPADDWMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPANDQ512 x y) mask) => (VPANDQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPBROADCASTB512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTBMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPACKUSDW512 x y) mask) => (VPACKUSDWMasked512Merging dst x y mask)
+(VPBLENDMWMasked512 dst (VPSHUFHW512 [a] x) mask) => (VPSHUFHWMasked512Merging dst [a] x mask)
+(VPBLENDVB128 dst (VRCP14PD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRCP14PDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSHRDW512 [a] x y) mask) => (VPSHRDWMasked512Merging dst [a] x y mask)
+(VPBLENDVB256 dst (VSQRTPS256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSQRTPSMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSUBSW512 x y) mask) => (VPSUBSWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVSXWD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWDMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTW256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTWMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTDMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVQB128 x) mask) => (VPMOVQBMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VPACKUSDW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPACKUSDWMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPMINSB512 x y) mask) => (VPMINSBMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPMULLD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPADDB512 x y) mask) => (VPADDBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPADDD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVWB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMADDWD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMADDWDMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMAXSD512 x y) mask) => (VPMAXSDMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPSHLDQ512 [a] x y) mask) => (VPSHLDQMasked512Merging dst [a] x y mask)
+(VPBLENDVB128 dst (VBROADCASTSS128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VBROADCASTSSMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVQD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXDQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXDQMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VDIVPD512 x y) mask) => (VDIVPDMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VADDPS512 x y) mask) => (VADDPSMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVSXBD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBDMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMOVUSDW256 x) mask) => (VPMOVUSDWMasked256Merging dst x mask)
+(VPBLENDVB256 dst (VPMULHUW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULHUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMULLQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPROLVD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPROLVQ512 x y) mask) => (VPROLVQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSHLDW128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDWMasked128Merging dst [a] x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVUSDW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVUSDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXUQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMULLW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULLWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPRORD256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPRORQ512 [a] x) mask) => (VPRORQMasked512Merging dst [a] x mask)
+(VPBLENDVB128 dst (VPSHLDD128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDDMasked128Merging dst [a] x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRAVW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VSUBPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSUBPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINUD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSHUFD256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPRORVQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLVQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSUBUSW512 x y) mask) => (VPSUBUSWMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPMOVSDB128 x) mask) => (VPMOVSDBMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VPMOVUSQD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVUSQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPMAXUB512 x y) mask) => (VPMAXUBMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPMINSQ512 x y) mask) => (VPMINSQMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VSQRTPD512 x) mask) => (VSQRTPDMasked512Merging dst x mask)
+(VPBLENDMDMasked512 dst (VSUBPS512 x y) mask) => (VSUBPSMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSUBUSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBUSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMAXUD512 x y) mask) => (VPMAXUDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VBROADCASTSS512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VBROADCASTSSMasked512Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVSQD256 x) mask) => (VPMOVSQDMasked256Merging dst x mask)
+(VPBLENDVB128 dst (VPMOVZXBD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBDMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXBQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBQMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VRSQRT14PD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRSQRT14PDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPRORD512 [a] x) mask) => (VPRORDMasked512Merging dst [a] x mask)
+(VPBLENDMWMasked512 dst (VPSUBW512 x y) mask) => (VPSUBWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPABSW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPADDUSB512 x y) mask) => (VPADDUSBMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPMOVZXWD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWDMasked512Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VMINPD512 x y) mask) => (VMINPDMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPMULLQ512 x y) mask) => (VPMULLQMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPROLVD512 x y) mask) => (VPROLVDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSUBW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VCVTTPS2DQ512 x) mask) => (VCVTTPS2DQMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPMOVZXWQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWQMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMADDWD512 x y) mask) => (VPMADDWDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VGF2P8MULB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VGF2P8MULBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPROLQ256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPROLQMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSLLVW512 x y) mask) => (VPSLLVWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPABSD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPAVGB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPAVGBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPAVGB512 x y) mask) => (VPAVGBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPBROADCASTB256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTBMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMAXPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMAXPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPMINUB512 x y) mask) => (VPMINUBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMINUQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMULPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMULPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VMAXPD512 x y) mask) => (VMAXPDMasked512Merging dst x y mask)
+(VPBLENDMBMasked512 dst (VPMAXSB512 x y) mask) => (VPMAXSBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMULHUW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULHUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMULPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMULPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPRORVQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSUBB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPACKSSDW512 x y) mask) => (VPACKSSDWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VCVTPS2UDQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VCVTPS2UDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVZXDQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXDQMasked512Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINUB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPRORVD512 x y) mask) => (VPRORVDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VSCALEFPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSCALEFPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSLLVQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLW256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLWMasked256constMerging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPABSW512 x) mask) => (VPABSWMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPMOVSXBQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBQMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VSCALEFPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSCALEFPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLQ256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLQMasked256constMerging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPADDW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VMULPD512 x y) mask) => (VMULPDMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPORQ512 x y) mask) => (VPORQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVSXWD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWDMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVUSQW128 x) mask) => (VPMOVUSQWMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VPMINSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VRSQRT14PD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRSQRT14PDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRAW128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAWMasked128constMerging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPABSQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VREDUCEPD512 [a] x) mask) => (VREDUCEPDMasked512Merging dst [a] x mask)
+(VPBLENDVB128 dst (VPMULHW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMULHWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSHUFHW256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFHWMasked256Merging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSRAW512const [a] x) mask) => (VPSRAWMasked512constMerging dst [a] x mask)
+(VPBLENDMDMasked512 dst (VPADDD512 x y) mask) => (VPADDDMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPOPCNTQ512 x) mask) => (VPOPCNTQMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPSHRDD128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDDMasked128Merging dst [a] x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSUBB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSUBSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPSUBUSB512 x y) mask) => (VPSUBUSBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPADDSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPADDUSW512 x y) mask) => (VPADDUSWMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VMAXPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMAXPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMAXSD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMINSB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VMULPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMULPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VRNDSCALEPS512 [a] x) mask) => (VRNDSCALEPSMasked512Merging dst [a] x mask)
+(VPBLENDMDMasked512 dst (VCVTPS2UDQ512 x) mask) => (VCVTPS2UDQMasked512Merging dst x mask)
+(VPBLENDVB256 dst (VDIVPS256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VDIVPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMAXSQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VMINPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMINPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSHUFD128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMBMasked512 dst (VPSHUFB512 x y) mask) => (VPSHUFBMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSHLDQ256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHLDQMasked256Merging dst [a] x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTQMasked512Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VREDUCEPS128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VREDUCEPSMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXWQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWQMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VSCALEFPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSCALEFPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPSHRDD512 [a] x y) mask) => (VPSHRDDMasked512Merging dst [a] x y mask)
+(VPBLENDVB128 dst (VPSRAVW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VSQRTPD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSQRTPDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPXORQ512 x y) mask) => (VPXORQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPAVGW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPAVGWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVSWB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VDIVPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VDIVPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VDIVPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VDIVPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINSQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINSQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPOPCNTW512 x) mask) => (VPOPCNTWMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPOPCNTD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPOPCNTD512 x) mask) => (VPOPCNTDMasked512Merging dst x mask)
+(VPBLENDVB256 dst (VPABSD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPBROADCASTQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTQMasked256Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VRNDSCALEPD128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRNDSCALEPDMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMOVDB128 x) mask) => (VPMOVDBMasked128Merging dst x mask)
+(VPBLENDVB128 dst (VPMOVSXWQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWQMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMINUW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMINUW512 x y) mask) => (VPMINUWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPOPCNTQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVQD256 x) mask) => (VPMOVQDMasked256Merging dst x mask)
+(VPBLENDVB256 dst (VPSHRDW256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDWMasked256Merging dst [a] x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPSRAD512const [a] x) mask) => (VPSRADMasked512constMerging dst [a] x mask)
+(VPBLENDVB128 dst (VPAVGB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPAVGBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPAVGW512 x y) mask) => (VPAVGWMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVSXBQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBQMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVZXBW512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBWMasked512Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXSW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMAXUD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMAXUQ512 x y) mask) => (VPMAXUQMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VMINPS512 x y) mask) => (VMINPSMasked512Merging dst x y mask)
+(VPBLENDMBMasked512 dst (VPABSB512 x) mask) => (VPABSBMasked512Merging dst x mask)
+(VPBLENDMDMasked512 dst (VPANDD512 x y) mask) => (VPANDDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMOVZXBW256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBWMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXBD512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBDMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMAXSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPSHUFD512 [a] x) mask) => (VPSHUFDMasked512Merging dst [a] x mask)
+(VPBLENDVB128 dst (VPSHUFHW128 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFHWMasked128Merging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSHRDQ256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDQMasked256Merging dst [a] x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMADDUBSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMADDUBSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPMINSD512 x y) mask) => (VPMINSDMasked512Merging dst x y mask)
+(VPBLENDMDMasked512 dst (VPSRAVD512 x y) mask) => (VPSRAVDMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VSUBPD512 x y) mask) => (VSUBPDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSLLW128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLWMasked128constMerging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSLLD256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLDMasked256constMerging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPMOVSWB256 x) mask) => (VPMOVSWBMasked256Merging dst x mask)
+(VPBLENDMQMasked512 dst (VPMOVQW128 x) mask) => (VPMOVQWMasked128Merging dst x mask)
+(VPBLENDVB256 dst (VPMINUQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMINUQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VRCP14PD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VRCP14PDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSHRDD256 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDDMasked256Merging dst [a] x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPSHRDQ512 [a] x y) mask) => (VPSHRDQMasked512Merging dst [a] x y mask)
+(VPBLENDVB128 dst (VPSLLVD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRLVD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPADDQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPADDQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXWQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXWQMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPLZCNTQ512 x) mask) => (VPLZCNTQMasked512Merging dst x mask)
+(VPBLENDVB256 dst (VPMAXUB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPRORQ256 [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPRORQMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VSCALEFPD512 x y) mask) => (VSCALEFPDMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSUBQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSLLD128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLDMasked128constMerging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VADDPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VADDPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMOVSQW128 x) mask) => (VPMOVSQWMasked128Merging dst x mask)
+(VPBLENDMWMasked512 dst (VPMAXUW512 x y) mask) => (VPMAXUWMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSHUFB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHUFBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRLVW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRLVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSLLQ128const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSLLQMasked128constMerging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRAD256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRADMasked256constMerging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMINUQ512 x y) mask) => (VPMINUQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPSRAVD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDMWMasked512 dst (VPSRLVW512 x y) mask) => (VPSRLVWMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPSUBW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRAW256const [a] x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAWMasked256constMerging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPABSW256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSWMasked256Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPACKSSDW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPACKSSDWMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVSQD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVSXBD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVSXBDMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXBQ512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXBQMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPLZCNTD256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPLZCNTDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPLZCNTQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPLZCNTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VMAXPD256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMAXPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPAVGW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPAVGWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPACKSSDW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPACKSSDWMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMOVZXWD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVZXWDMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPOPCNTQ256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSRAVQ128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSUBSW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPXORD512 x y) mask) => (VPXORDMasked512Merging dst x y mask)
+(VPBLENDMBMasked512 dst (VPADDSB512 x y) mask) => (VPADDSBMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPBROADCASTD128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VMAXPS128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VMAXPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMAXUW256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPSHRDQ128 [a] x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSHRDQMasked128Merging dst [a] x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPSLLVD512 x y) mask) => (VPSLLVDMasked512Merging dst x y mask)
+(VPBLENDMWMasked512 dst (VPSLLW512const [a] x) mask) => (VPSLLWMasked512constMerging dst [a] x mask)
+(VPBLENDMDMasked512 dst (VPSLLD512const [a] x) mask) => (VPSLLDMasked512constMerging dst [a] x mask)
+(VPBLENDMWMasked512 dst (VPADDSW512 x y) mask) => (VPADDSWMasked512Merging dst x y mask)
+(VPBLENDMQMasked512 dst (VPMOVSQB128 x) mask) => (VPMOVSQBMasked128Merging dst x mask)
+(VPBLENDMDMasked512 dst (VPMINUD512 x y) mask) => (VPMINUDMasked512Merging dst x y mask)
+(VPBLENDVB256 dst (VPOPCNTW256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPOPCNTWMasked256Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VRSQRT14PD512 x) mask) => (VRSQRT14PDMasked512Merging dst x mask)
+(VPBLENDMDMasked512 dst (VSCALEFPS512 x y) mask) => (VSCALEFPSMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VPMAXUW128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSRAVQ256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSRAVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VSQRTPS512 x) mask) => (VSQRTPSMasked512Merging dst x mask)
+(VPBLENDMQMasked512 dst (VPSRAQ512const [a] x) mask) => (VPSRAQMasked512constMerging dst [a] x mask)
+(VPBLENDVB128 dst (VPABSB128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPABSB256 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSBMasked256Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPABSQ128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPABSQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPMOVDW128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMOVDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(VPBLENDMQMasked512 dst (VPMAXSQ512 x y) mask) => (VPMAXSQMasked512Merging dst x y mask)
+(VPBLENDVB128 dst (VSCALEFPD128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSCALEFPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VSQRTPS128 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VSQRTPSMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+(VPBLENDVB256 dst (VPSUBSB256 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPSUBSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VPABSD512 x) mask) => (VPABSDMasked512Merging dst x mask)
+(VPBLENDVB128 dst (VPBROADCASTW512 x) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPBROADCASTWMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+(VPBLENDVB128 dst (VPMAXUB128 x y) mask) && v.Block.CPUfeatures.hasFeature(CPUavx512) => (VPMAXUBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(VPBLENDMDMasked512 dst (VMULPS512 x y) mask) => (VMULPSMasked512Merging dst x y mask)
+(VPBLENDMWMasked512 dst (VPMULLW512 x y) mask) => (VPMULLWMasked512Merging dst x y mask)
 (VPABSD512 l:(VMOVDQUload512 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSD512load {sym} [off] ptr mem)
 (VPABSQ128 l:(VMOVDQUload128 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ128load {sym} [off] ptr mem)
 (VPABSQ256 l:(VMOVDQUload256 {sym} [off] ptr mem)) && canMergeLoad(v, l) && clobber(l) => (VPABSQ256load {sym} [off] ptr mem)
diff --git a/src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go b/src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go
index f867c6e315..4f22d8582b 100644
--- a/src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go
+++ b/src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go
@@ -167,12 +167,6 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
 		{name: "VMOVDQU64Masked128", argLength: 2, reg: wkw, asm: "VMOVDQU64", commutative: false, typ: "Vec128", resultInArg0: false},
 		{name: "VMOVDQU64Masked256", argLength: 2, reg: wkw, asm: "VMOVDQU64", commutative: false, typ: "Vec256", resultInArg0: false},
 		{name: "VMOVDQU64Masked512", argLength: 2, reg: wkw, asm: "VMOVDQU64", commutative: false, typ: "Vec512", resultInArg0: false},
-		{name: "VMOVUPDMasked128", argLength: 2, reg: wkw, asm: "VMOVUPD", commutative: false, typ: "Vec128", resultInArg0: false},
-		{name: "VMOVUPDMasked256", argLength: 2, reg: wkw, asm: "VMOVUPD", commutative: false, typ: "Vec256", resultInArg0: false},
-		{name: "VMOVUPDMasked512", argLength: 2, reg: wkw, asm: "VMOVUPD", commutative: false, typ: "Vec512", resultInArg0: false},
-		{name: "VMOVUPSMasked128", argLength: 2, reg: wkw, asm: "VMOVUPS", commutative: false, typ: "Vec128", resultInArg0: false},
-		{name: "VMOVUPSMasked256", argLength: 2, reg: wkw, asm: "VMOVUPS", commutative: false, typ: "Vec256", resultInArg0: false},
-		{name: "VMOVUPSMasked512", argLength: 2, reg: wkw, asm: "VMOVUPS", commutative: false, typ: "Vec512", resultInArg0: false},
 		{name: "VMULPD128", argLength: 2, reg: v21, asm: "VMULPD", commutative: true, typ: "Vec128", resultInArg0: false},
 		{name: "VMULPD256", argLength: 2, reg: v21, asm: "VMULPD", commutative: true, typ: "Vec256", resultInArg0: false},
 		{name: "VMULPD512", argLength: 2, reg: w21, asm: "VMULPD", commutative: true, typ: "Vec512", resultInArg0: false},
@@ -1900,5 +1894,448 @@ func simdAMD64Ops(v11, v21, v2k, vkv, v2kv, v2kk, v31, v3kv, vgpv, vgp, vfpv, vf
 		{name: "VRNDSCALEPSMasked512load", argLength: 3, reg: wkwload, asm: "VRNDSCALEPS", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
 		{name: "VSHUFPD512load", argLength: 3, reg: w21load, asm: "VSHUFPD", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
 		{name: "VSHUFPS512load", argLength: 3, reg: w21load, asm: "VSHUFPS", commutative: false, typ: "Vec512", aux: "SymValAndOff", symEffect: "Read", resultInArg0: false},
+		{name: "VADDPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VADDPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VADDPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VADDPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VADDPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VADDPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VADDPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VADDPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VADDPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VADDPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VADDPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VADDPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VBROADCASTSDMasked256Merging", argLength: 3, reg: w2kw, asm: "VBROADCASTSD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VBROADCASTSDMasked512Merging", argLength: 3, reg: w2kw, asm: "VBROADCASTSD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VBROADCASTSSMasked128Merging", argLength: 3, reg: w2kw, asm: "VBROADCASTSS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VBROADCASTSSMasked256Merging", argLength: 3, reg: w2kw, asm: "VBROADCASTSS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VBROADCASTSSMasked512Merging", argLength: 3, reg: w2kw, asm: "VBROADCASTSS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VCVTPS2UDQMasked128Merging", argLength: 3, reg: w2kw, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VCVTPS2UDQMasked256Merging", argLength: 3, reg: w2kw, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VCVTPS2UDQMasked512Merging", argLength: 3, reg: w2kw, asm: "VCVTPS2UDQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VCVTTPS2DQMasked128Merging", argLength: 3, reg: w2kw, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VCVTTPS2DQMasked256Merging", argLength: 3, reg: w2kw, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VCVTTPS2DQMasked512Merging", argLength: 3, reg: w2kw, asm: "VCVTTPS2DQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VDIVPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VDIVPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VDIVPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VDIVPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VDIVPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VDIVPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VDIVPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VDIVPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VDIVPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VDIVPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VDIVPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VDIVPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VGF2P8MULBMasked128Merging", argLength: 4, reg: w3kw, asm: "VGF2P8MULB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VGF2P8MULBMasked256Merging", argLength: 4, reg: w3kw, asm: "VGF2P8MULB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VGF2P8MULBMasked512Merging", argLength: 4, reg: w3kw, asm: "VGF2P8MULB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMAXPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VMAXPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMAXPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VMAXPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMAXPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VMAXPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMAXPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VMAXPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMAXPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VMAXPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMAXPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VMAXPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMINPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VMINPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMINPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VMINPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMINPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VMINPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMINPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VMINPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMINPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VMINPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMINPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VMINPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMULPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VMULPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMULPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VMULPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMULPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VMULPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VMULPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VMULPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VMULPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VMULPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VMULPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VMULPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPABSBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPABSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPABSBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPABSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPABSBMasked512Merging", argLength: 3, reg: w2kw, asm: "VPABSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPABSDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPABSD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPABSDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPABSD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPABSDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPABSD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPABSQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPABSQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPABSQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPABSQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPABSQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPABSQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPABSWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPABSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPABSWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPABSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPABSWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPABSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPACKSSDWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPACKSSDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPACKSSDWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPACKSSDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPACKSSDWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPACKSSDW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPACKUSDWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPACKUSDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPACKUSDWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPACKUSDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPACKUSDWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPACKUSDW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDUSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDUSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDUSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDUSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDUSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDUSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDUSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDUSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDUSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDUSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDUSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDUSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPADDWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPADDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPADDWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPADDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPADDWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPADDW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPANDDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPANDD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPANDDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPANDD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPANDDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPANDD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPANDQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPANDQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPANDQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPANDQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPANDQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPANDQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPAVGBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPAVGB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPAVGBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPAVGB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPAVGBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPAVGB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPAVGWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPAVGW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPAVGWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPAVGW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPAVGWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPAVGW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPBROADCASTBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPBROADCASTBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPBROADCASTBMasked512Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPBROADCASTDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPBROADCASTDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPBROADCASTDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPBROADCASTQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPBROADCASTQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPBROADCASTQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPBROADCASTWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPBROADCASTWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPBROADCASTWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPBROADCASTW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPLZCNTDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPLZCNTDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPLZCNTDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPLZCNTQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPLZCNTQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPLZCNTQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPLZCNTQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMADDUBSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMADDUBSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMADDUBSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMADDUBSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMADDUBSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMADDUBSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMADDWDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMADDWD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMADDWDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMADDWD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMADDWDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMADDWD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXSDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXSD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXSDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXSD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXSDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXSD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXSQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXSQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXSQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXSQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXSQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXSQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXUBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXUB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXUBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXUB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXUBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXUB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXUDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXUD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXUDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXUD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXUDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXUD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXUQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXUQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXUQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXUQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXUQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXUQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMAXUWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMAXUW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMAXUWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMAXUW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMAXUWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMAXUW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINSDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINSD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINSDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINSD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINSDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINSD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINSQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINSQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINSQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINSQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINSQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINSQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINUBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINUB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINUBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINUB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINUBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINUB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINUDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINUD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINUDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINUD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINUDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINUD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINUQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINUQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINUQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINUQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINUQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINUQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMINUWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMINUW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMINUWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMINUW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMINUWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMINUW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVDBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVDB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVDWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVDWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVQBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVQB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVQDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVQD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVQDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVQD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVQWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVQW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSDBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSDB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSDWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSDWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSQBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSQB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSQDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSQD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSQDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSQD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSQWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSQW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSWBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSWB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSWBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSWB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXBDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXBDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXBDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVSXBQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXBQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXBQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVSXBWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXBWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXBWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXBW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVSXDQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXDQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXDQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXDQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXDQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXDQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVSXWDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXWDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXWDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVSXWQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVSXWQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVSXWQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVSXWQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVUSDBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSDB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSDWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSDW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSDWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSDW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVUSQBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSQB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSQDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSQD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSQDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSQD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVUSQWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSQW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSWBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSWB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVUSWBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVUSWB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVWBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVWB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVWBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVWB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXBDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXBDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXBDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVZXBQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXBQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXBQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVZXBWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXBWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXBWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXBW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVZXDQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXDQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXDQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXDQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXDQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXDQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVZXWDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXWDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXWDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMOVZXWQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMOVZXWQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMOVZXWQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPMOVZXWQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMULHUWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMULHUW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMULHUWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMULHUW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMULHUWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMULHUW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMULHWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMULHW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMULHWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMULHW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMULHWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMULHW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMULLDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMULLD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMULLDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMULLD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMULLDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMULLD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMULLQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMULLQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMULLQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMULLQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMULLQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMULLQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPMULLWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPMULLW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPMULLWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPMULLW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPMULLWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPMULLW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPOPCNTBMasked128Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPOPCNTBMasked256Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPOPCNTBMasked512Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPOPCNTDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPOPCNTDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPOPCNTDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPOPCNTQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPOPCNTQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPOPCNTQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPOPCNTWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPOPCNTWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPOPCNTWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPOPCNTW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPORDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPORD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPORDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPORD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPORDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPORD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPORQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPORQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPORQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPORQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPORQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPORQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPROLVDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPROLVD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPROLVDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPROLVD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPROLVDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPROLVD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPROLVQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPROLVQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPROLVQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPROLVQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPROLVQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPROLVQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPRORVDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPRORVD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPRORVDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPRORVD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPRORVDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPRORVD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPRORVQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPRORVQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPRORVQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPRORVQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPRORVQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPRORVQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHUFBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHUFB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHUFBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHUFB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHUFBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHUFB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLVDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSLLVD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLVDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSLLVD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLVDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSLLVD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLVQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSLLVQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLVQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSLLVQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLVQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSLLVQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLVWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSLLVW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLVWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSLLVW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLVWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSLLVW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRAVDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRAVD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRAVDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRAVD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRAVDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRAVD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRAVQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRAVQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRAVQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRAVQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRAVQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRAVQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRAVWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRAVW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRAVWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRAVW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRAVWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRAVW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLVDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRLVD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLVDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRLVD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLVDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRLVD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLVQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRLVQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLVQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRLVQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLVQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRLVQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLVWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSRLVW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLVWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSRLVW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLVWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSRLVW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBUSBMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSB", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBUSBMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSB", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBUSBMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSB", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBUSWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBUSWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBUSWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBUSW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSUBWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSUBW", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSUBWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSUBW", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSUBWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSUBW", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPXORDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPXORD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPXORDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPXORD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPXORDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPXORD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPXORQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPXORQ", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPXORQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPXORQ", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPXORQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPXORQ", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRCP14PDMasked128Merging", argLength: 3, reg: w2kw, asm: "VRCP14PD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRCP14PDMasked256Merging", argLength: 3, reg: w2kw, asm: "VRCP14PD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRCP14PDMasked512Merging", argLength: 3, reg: w2kw, asm: "VRCP14PD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRCP14PSMasked128Merging", argLength: 3, reg: w2kw, asm: "VRCP14PS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRCP14PSMasked256Merging", argLength: 3, reg: w2kw, asm: "VRCP14PS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRCP14PSMasked512Merging", argLength: 3, reg: w2kw, asm: "VRCP14PS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRSQRT14PDMasked128Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRSQRT14PDMasked256Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRSQRT14PDMasked512Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRSQRT14PSMasked128Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRSQRT14PSMasked256Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRSQRT14PSMasked512Merging", argLength: 3, reg: w2kw, asm: "VRSQRT14PS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSCALEFPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSCALEFPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSCALEFPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSCALEFPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSCALEFPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSCALEFPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VSCALEFPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSQRTPDMasked128Merging", argLength: 3, reg: w2kw, asm: "VSQRTPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSQRTPDMasked256Merging", argLength: 3, reg: w2kw, asm: "VSQRTPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSQRTPDMasked512Merging", argLength: 3, reg: w2kw, asm: "VSQRTPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSQRTPSMasked128Merging", argLength: 3, reg: w2kw, asm: "VSQRTPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSQRTPSMasked256Merging", argLength: 3, reg: w2kw, asm: "VSQRTPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSQRTPSMasked512Merging", argLength: 3, reg: w2kw, asm: "VSQRTPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSUBPDMasked128Merging", argLength: 4, reg: w3kw, asm: "VSUBPD", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSUBPDMasked256Merging", argLength: 4, reg: w3kw, asm: "VSUBPD", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSUBPDMasked512Merging", argLength: 4, reg: w3kw, asm: "VSUBPD", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VSUBPSMasked128Merging", argLength: 4, reg: w3kw, asm: "VSUBPS", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VSUBPSMasked256Merging", argLength: 4, reg: w3kw, asm: "VSUBPS", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VSUBPSMasked512Merging", argLength: 4, reg: w3kw, asm: "VSUBPS", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPROLDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPROLD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPROLDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPROLD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPROLDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPROLD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPROLQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPROLQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPROLQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPROLQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPROLQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPROLQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPRORDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPRORD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPRORDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPRORD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPRORDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPRORD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPRORQMasked128Merging", argLength: 3, reg: w2kw, asm: "VPRORQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPRORQMasked256Merging", argLength: 3, reg: w2kw, asm: "VPRORQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPRORQMasked512Merging", argLength: 3, reg: w2kw, asm: "VPRORQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHLDDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHLDD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHLDDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHLDD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHLDDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHLDD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHLDQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHLDQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHLDQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHLDQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHLDQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHLDQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHLDWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHLDW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHLDWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHLDW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHLDWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHLDW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHRDDMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHRDD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHRDDMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHRDD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHRDDMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHRDD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHRDQMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHRDQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHRDQMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHRDQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHRDQMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHRDQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHRDWMasked128Merging", argLength: 4, reg: w3kw, asm: "VPSHRDW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHRDWMasked256Merging", argLength: 4, reg: w3kw, asm: "VPSHRDW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHRDWMasked512Merging", argLength: 4, reg: w3kw, asm: "VPSHRDW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHUFDMasked128Merging", argLength: 3, reg: w2kw, asm: "VPSHUFD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHUFDMasked256Merging", argLength: 3, reg: w2kw, asm: "VPSHUFD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHUFDMasked512Merging", argLength: 3, reg: w2kw, asm: "VPSHUFD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSHUFHWMasked128Merging", argLength: 3, reg: w2kw, asm: "VPSHUFHW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSHUFHWMasked256Merging", argLength: 3, reg: w2kw, asm: "VPSHUFHW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSHUFHWMasked512Merging", argLength: 3, reg: w2kw, asm: "VPSHUFHW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLDMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSLLD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLDMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSLLD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLDMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSLLD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLQMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSLLQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLQMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSLLQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLQMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSLLQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSLLWMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSLLW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSLLWMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSLLW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSLLWMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSLLW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRADMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRAD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRADMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRAD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRADMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRAD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRAQMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRAQMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRAQMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRAQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRAWMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRAW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRAWMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRAW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRAWMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRAW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLDMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRLD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLDMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRLD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLDMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRLD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLQMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRLQ", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLQMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRLQ", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLQMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRLQ", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VPSRLWMasked128constMerging", argLength: 3, reg: w2kw, asm: "VPSRLW", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VPSRLWMasked256constMerging", argLength: 3, reg: w2kw, asm: "VPSRLW", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VPSRLWMasked512constMerging", argLength: 3, reg: w2kw, asm: "VPSRLW", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VREDUCEPDMasked128Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VREDUCEPDMasked256Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VREDUCEPDMasked512Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VREDUCEPSMasked128Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPS", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VREDUCEPSMasked256Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPS", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VREDUCEPSMasked512Merging", argLength: 3, reg: w2kw, asm: "VREDUCEPS", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRNDSCALEPDMasked128Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPD", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRNDSCALEPDMasked256Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPD", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRNDSCALEPDMasked512Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPD", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
+		{name: "VRNDSCALEPSMasked128Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPS", aux: "UInt8", commutative: false, typ: "Vec128", resultInArg0: true},
+		{name: "VRNDSCALEPSMasked256Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPS", aux: "UInt8", commutative: false, typ: "Vec256", resultInArg0: true},
+		{name: "VRNDSCALEPSMasked512Merging", argLength: 3, reg: w2kw, asm: "VRNDSCALEPS", aux: "UInt8", commutative: false, typ: "Vec512", resultInArg0: true},
 	}
 }
diff --git a/src/cmd/compile/internal/ssa/opGen.go b/src/cmd/compile/internal/ssa/opGen.go
index 68bfe68eb4..9d7ee4bea8 100644
--- a/src/cmd/compile/internal/ssa/opGen.go
+++ b/src/cmd/compile/internal/ssa/opGen.go
@@ -1407,12 +1407,6 @@ const (
 	OpAMD64VMOVDQU64Masked128
 	OpAMD64VMOVDQU64Masked256
 	OpAMD64VMOVDQU64Masked512
-	OpAMD64VMOVUPDMasked128
-	OpAMD64VMOVUPDMasked256
-	OpAMD64VMOVUPDMasked512
-	OpAMD64VMOVUPSMasked128
-	OpAMD64VMOVUPSMasked256
-	OpAMD64VMOVUPSMasked512
 	OpAMD64VMULPD128
 	OpAMD64VMULPD256
 	OpAMD64VMULPD512
@@ -3140,6 +3134,449 @@ const (
 	OpAMD64VRNDSCALEPSMasked512load
 	OpAMD64VSHUFPD512load
 	OpAMD64VSHUFPS512load
+	OpAMD64VADDPDMasked128Merging
+	OpAMD64VADDPDMasked256Merging
+	OpAMD64VADDPDMasked512Merging
+	OpAMD64VADDPSMasked128Merging
+	OpAMD64VADDPSMasked256Merging
+	OpAMD64VADDPSMasked512Merging
+	OpAMD64VBROADCASTSDMasked256Merging
+	OpAMD64VBROADCASTSDMasked512Merging
+	OpAMD64VBROADCASTSSMasked128Merging
+	OpAMD64VBROADCASTSSMasked256Merging
+	OpAMD64VBROADCASTSSMasked512Merging
+	OpAMD64VCVTPS2UDQMasked128Merging
+	OpAMD64VCVTPS2UDQMasked256Merging
+	OpAMD64VCVTPS2UDQMasked512Merging
+	OpAMD64VCVTTPS2DQMasked128Merging
+	OpAMD64VCVTTPS2DQMasked256Merging
+	OpAMD64VCVTTPS2DQMasked512Merging
+	OpAMD64VDIVPDMasked128Merging
+	OpAMD64VDIVPDMasked256Merging
+	OpAMD64VDIVPDMasked512Merging
+	OpAMD64VDIVPSMasked128Merging
+	OpAMD64VDIVPSMasked256Merging
+	OpAMD64VDIVPSMasked512Merging
+	OpAMD64VGF2P8MULBMasked128Merging
+	OpAMD64VGF2P8MULBMasked256Merging
+	OpAMD64VGF2P8MULBMasked512Merging
+	OpAMD64VMAXPDMasked128Merging
+	OpAMD64VMAXPDMasked256Merging
+	OpAMD64VMAXPDMasked512Merging
+	OpAMD64VMAXPSMasked128Merging
+	OpAMD64VMAXPSMasked256Merging
+	OpAMD64VMAXPSMasked512Merging
+	OpAMD64VMINPDMasked128Merging
+	OpAMD64VMINPDMasked256Merging
+	OpAMD64VMINPDMasked512Merging
+	OpAMD64VMINPSMasked128Merging
+	OpAMD64VMINPSMasked256Merging
+	OpAMD64VMINPSMasked512Merging
+	OpAMD64VMULPDMasked128Merging
+	OpAMD64VMULPDMasked256Merging
+	OpAMD64VMULPDMasked512Merging
+	OpAMD64VMULPSMasked128Merging
+	OpAMD64VMULPSMasked256Merging
+	OpAMD64VMULPSMasked512Merging
+	OpAMD64VPABSBMasked128Merging
+	OpAMD64VPABSBMasked256Merging
+	OpAMD64VPABSBMasked512Merging
+	OpAMD64VPABSDMasked128Merging
+	OpAMD64VPABSDMasked256Merging
+	OpAMD64VPABSDMasked512Merging
+	OpAMD64VPABSQMasked128Merging
+	OpAMD64VPABSQMasked256Merging
+	OpAMD64VPABSQMasked512Merging
+	OpAMD64VPABSWMasked128Merging
+	OpAMD64VPABSWMasked256Merging
+	OpAMD64VPABSWMasked512Merging
+	OpAMD64VPACKSSDWMasked128Merging
+	OpAMD64VPACKSSDWMasked256Merging
+	OpAMD64VPACKSSDWMasked512Merging
+	OpAMD64VPACKUSDWMasked128Merging
+	OpAMD64VPACKUSDWMasked256Merging
+	OpAMD64VPACKUSDWMasked512Merging
+	OpAMD64VPADDBMasked128Merging
+	OpAMD64VPADDBMasked256Merging
+	OpAMD64VPADDBMasked512Merging
+	OpAMD64VPADDDMasked128Merging
+	OpAMD64VPADDDMasked256Merging
+	OpAMD64VPADDDMasked512Merging
+	OpAMD64VPADDQMasked128Merging
+	OpAMD64VPADDQMasked256Merging
+	OpAMD64VPADDQMasked512Merging
+	OpAMD64VPADDSBMasked128Merging
+	OpAMD64VPADDSBMasked256Merging
+	OpAMD64VPADDSBMasked512Merging
+	OpAMD64VPADDSWMasked128Merging
+	OpAMD64VPADDSWMasked256Merging
+	OpAMD64VPADDSWMasked512Merging
+	OpAMD64VPADDUSBMasked128Merging
+	OpAMD64VPADDUSBMasked256Merging
+	OpAMD64VPADDUSBMasked512Merging
+	OpAMD64VPADDUSWMasked128Merging
+	OpAMD64VPADDUSWMasked256Merging
+	OpAMD64VPADDUSWMasked512Merging
+	OpAMD64VPADDWMasked128Merging
+	OpAMD64VPADDWMasked256Merging
+	OpAMD64VPADDWMasked512Merging
+	OpAMD64VPANDDMasked128Merging
+	OpAMD64VPANDDMasked256Merging
+	OpAMD64VPANDDMasked512Merging
+	OpAMD64VPANDQMasked128Merging
+	OpAMD64VPANDQMasked256Merging
+	OpAMD64VPANDQMasked512Merging
+	OpAMD64VPAVGBMasked128Merging
+	OpAMD64VPAVGBMasked256Merging
+	OpAMD64VPAVGBMasked512Merging
+	OpAMD64VPAVGWMasked128Merging
+	OpAMD64VPAVGWMasked256Merging
+	OpAMD64VPAVGWMasked512Merging
+	OpAMD64VPBROADCASTBMasked128Merging
+	OpAMD64VPBROADCASTBMasked256Merging
+	OpAMD64VPBROADCASTBMasked512Merging
+	OpAMD64VPBROADCASTDMasked128Merging
+	OpAMD64VPBROADCASTDMasked256Merging
+	OpAMD64VPBROADCASTDMasked512Merging
+	OpAMD64VPBROADCASTQMasked128Merging
+	OpAMD64VPBROADCASTQMasked256Merging
+	OpAMD64VPBROADCASTQMasked512Merging
+	OpAMD64VPBROADCASTWMasked128Merging
+	OpAMD64VPBROADCASTWMasked256Merging
+	OpAMD64VPBROADCASTWMasked512Merging
+	OpAMD64VPLZCNTDMasked128Merging
+	OpAMD64VPLZCNTDMasked256Merging
+	OpAMD64VPLZCNTDMasked512Merging
+	OpAMD64VPLZCNTQMasked128Merging
+	OpAMD64VPLZCNTQMasked256Merging
+	OpAMD64VPLZCNTQMasked512Merging
+	OpAMD64VPMADDUBSWMasked128Merging
+	OpAMD64VPMADDUBSWMasked256Merging
+	OpAMD64VPMADDUBSWMasked512Merging
+	OpAMD64VPMADDWDMasked128Merging
+	OpAMD64VPMADDWDMasked256Merging
+	OpAMD64VPMADDWDMasked512Merging
+	OpAMD64VPMAXSBMasked128Merging
+	OpAMD64VPMAXSBMasked256Merging
+	OpAMD64VPMAXSBMasked512Merging
+	OpAMD64VPMAXSDMasked128Merging
+	OpAMD64VPMAXSDMasked256Merging
+	OpAMD64VPMAXSDMasked512Merging
+	OpAMD64VPMAXSQMasked128Merging
+	OpAMD64VPMAXSQMasked256Merging
+	OpAMD64VPMAXSQMasked512Merging
+	OpAMD64VPMAXSWMasked128Merging
+	OpAMD64VPMAXSWMasked256Merging
+	OpAMD64VPMAXSWMasked512Merging
+	OpAMD64VPMAXUBMasked128Merging
+	OpAMD64VPMAXUBMasked256Merging
+	OpAMD64VPMAXUBMasked512Merging
+	OpAMD64VPMAXUDMasked128Merging
+	OpAMD64VPMAXUDMasked256Merging
+	OpAMD64VPMAXUDMasked512Merging
+	OpAMD64VPMAXUQMasked128Merging
+	OpAMD64VPMAXUQMasked256Merging
+	OpAMD64VPMAXUQMasked512Merging
+	OpAMD64VPMAXUWMasked128Merging
+	OpAMD64VPMAXUWMasked256Merging
+	OpAMD64VPMAXUWMasked512Merging
+	OpAMD64VPMINSBMasked128Merging
+	OpAMD64VPMINSBMasked256Merging
+	OpAMD64VPMINSBMasked512Merging
+	OpAMD64VPMINSDMasked128Merging
+	OpAMD64VPMINSDMasked256Merging
+	OpAMD64VPMINSDMasked512Merging
+	OpAMD64VPMINSQMasked128Merging
+	OpAMD64VPMINSQMasked256Merging
+	OpAMD64VPMINSQMasked512Merging
+	OpAMD64VPMINSWMasked128Merging
+	OpAMD64VPMINSWMasked256Merging
+	OpAMD64VPMINSWMasked512Merging
+	OpAMD64VPMINUBMasked128Merging
+	OpAMD64VPMINUBMasked256Merging
+	OpAMD64VPMINUBMasked512Merging
+	OpAMD64VPMINUDMasked128Merging
+	OpAMD64VPMINUDMasked256Merging
+	OpAMD64VPMINUDMasked512Merging
+	OpAMD64VPMINUQMasked128Merging
+	OpAMD64VPMINUQMasked256Merging
+	OpAMD64VPMINUQMasked512Merging
+	OpAMD64VPMINUWMasked128Merging
+	OpAMD64VPMINUWMasked256Merging
+	OpAMD64VPMINUWMasked512Merging
+	OpAMD64VPMOVDBMasked128Merging
+	OpAMD64VPMOVDWMasked128Merging
+	OpAMD64VPMOVDWMasked256Merging
+	OpAMD64VPMOVQBMasked128Merging
+	OpAMD64VPMOVQDMasked128Merging
+	OpAMD64VPMOVQDMasked256Merging
+	OpAMD64VPMOVQWMasked128Merging
+	OpAMD64VPMOVSDBMasked128Merging
+	OpAMD64VPMOVSDWMasked128Merging
+	OpAMD64VPMOVSDWMasked256Merging
+	OpAMD64VPMOVSQBMasked128Merging
+	OpAMD64VPMOVSQDMasked128Merging
+	OpAMD64VPMOVSQDMasked256Merging
+	OpAMD64VPMOVSQWMasked128Merging
+	OpAMD64VPMOVSWBMasked128Merging
+	OpAMD64VPMOVSWBMasked256Merging
+	OpAMD64VPMOVSXBDMasked128Merging
+	OpAMD64VPMOVSXBDMasked256Merging
+	OpAMD64VPMOVSXBDMasked512Merging
+	OpAMD64VPMOVSXBQMasked128Merging
+	OpAMD64VPMOVSXBQMasked256Merging
+	OpAMD64VPMOVSXBQMasked512Merging
+	OpAMD64VPMOVSXBWMasked128Merging
+	OpAMD64VPMOVSXBWMasked256Merging
+	OpAMD64VPMOVSXBWMasked512Merging
+	OpAMD64VPMOVSXDQMasked128Merging
+	OpAMD64VPMOVSXDQMasked256Merging
+	OpAMD64VPMOVSXDQMasked512Merging
+	OpAMD64VPMOVSXWDMasked128Merging
+	OpAMD64VPMOVSXWDMasked256Merging
+	OpAMD64VPMOVSXWDMasked512Merging
+	OpAMD64VPMOVSXWQMasked128Merging
+	OpAMD64VPMOVSXWQMasked256Merging
+	OpAMD64VPMOVSXWQMasked512Merging
+	OpAMD64VPMOVUSDBMasked128Merging
+	OpAMD64VPMOVUSDWMasked128Merging
+	OpAMD64VPMOVUSDWMasked256Merging
+	OpAMD64VPMOVUSQBMasked128Merging
+	OpAMD64VPMOVUSQDMasked128Merging
+	OpAMD64VPMOVUSQDMasked256Merging
+	OpAMD64VPMOVUSQWMasked128Merging
+	OpAMD64VPMOVUSWBMasked128Merging
+	OpAMD64VPMOVUSWBMasked256Merging
+	OpAMD64VPMOVWBMasked128Merging
+	OpAMD64VPMOVWBMasked256Merging
+	OpAMD64VPMOVZXBDMasked128Merging
+	OpAMD64VPMOVZXBDMasked256Merging
+	OpAMD64VPMOVZXBDMasked512Merging
+	OpAMD64VPMOVZXBQMasked128Merging
+	OpAMD64VPMOVZXBQMasked256Merging
+	OpAMD64VPMOVZXBQMasked512Merging
+	OpAMD64VPMOVZXBWMasked128Merging
+	OpAMD64VPMOVZXBWMasked256Merging
+	OpAMD64VPMOVZXBWMasked512Merging
+	OpAMD64VPMOVZXDQMasked128Merging
+	OpAMD64VPMOVZXDQMasked256Merging
+	OpAMD64VPMOVZXDQMasked512Merging
+	OpAMD64VPMOVZXWDMasked128Merging
+	OpAMD64VPMOVZXWDMasked256Merging
+	OpAMD64VPMOVZXWDMasked512Merging
+	OpAMD64VPMOVZXWQMasked128Merging
+	OpAMD64VPMOVZXWQMasked256Merging
+	OpAMD64VPMOVZXWQMasked512Merging
+	OpAMD64VPMULHUWMasked128Merging
+	OpAMD64VPMULHUWMasked256Merging
+	OpAMD64VPMULHUWMasked512Merging
+	OpAMD64VPMULHWMasked128Merging
+	OpAMD64VPMULHWMasked256Merging
+	OpAMD64VPMULHWMasked512Merging
+	OpAMD64VPMULLDMasked128Merging
+	OpAMD64VPMULLDMasked256Merging
+	OpAMD64VPMULLDMasked512Merging
+	OpAMD64VPMULLQMasked128Merging
+	OpAMD64VPMULLQMasked256Merging
+	OpAMD64VPMULLQMasked512Merging
+	OpAMD64VPMULLWMasked128Merging
+	OpAMD64VPMULLWMasked256Merging
+	OpAMD64VPMULLWMasked512Merging
+	OpAMD64VPOPCNTBMasked128Merging
+	OpAMD64VPOPCNTBMasked256Merging
+	OpAMD64VPOPCNTBMasked512Merging
+	OpAMD64VPOPCNTDMasked128Merging
+	OpAMD64VPOPCNTDMasked256Merging
+	OpAMD64VPOPCNTDMasked512Merging
+	OpAMD64VPOPCNTQMasked128Merging
+	OpAMD64VPOPCNTQMasked256Merging
+	OpAMD64VPOPCNTQMasked512Merging
+	OpAMD64VPOPCNTWMasked128Merging
+	OpAMD64VPOPCNTWMasked256Merging
+	OpAMD64VPOPCNTWMasked512Merging
+	OpAMD64VPORDMasked128Merging
+	OpAMD64VPORDMasked256Merging
+	OpAMD64VPORDMasked512Merging
+	OpAMD64VPORQMasked128Merging
+	OpAMD64VPORQMasked256Merging
+	OpAMD64VPORQMasked512Merging
+	OpAMD64VPROLVDMasked128Merging
+	OpAMD64VPROLVDMasked256Merging
+	OpAMD64VPROLVDMasked512Merging
+	OpAMD64VPROLVQMasked128Merging
+	OpAMD64VPROLVQMasked256Merging
+	OpAMD64VPROLVQMasked512Merging
+	OpAMD64VPRORVDMasked128Merging
+	OpAMD64VPRORVDMasked256Merging
+	OpAMD64VPRORVDMasked512Merging
+	OpAMD64VPRORVQMasked128Merging
+	OpAMD64VPRORVQMasked256Merging
+	OpAMD64VPRORVQMasked512Merging
+	OpAMD64VPSHUFBMasked128Merging
+	OpAMD64VPSHUFBMasked256Merging
+	OpAMD64VPSHUFBMasked512Merging
+	OpAMD64VPSLLVDMasked128Merging
+	OpAMD64VPSLLVDMasked256Merging
+	OpAMD64VPSLLVDMasked512Merging
+	OpAMD64VPSLLVQMasked128Merging
+	OpAMD64VPSLLVQMasked256Merging
+	OpAMD64VPSLLVQMasked512Merging
+	OpAMD64VPSLLVWMasked128Merging
+	OpAMD64VPSLLVWMasked256Merging
+	OpAMD64VPSLLVWMasked512Merging
+	OpAMD64VPSRAVDMasked128Merging
+	OpAMD64VPSRAVDMasked256Merging
+	OpAMD64VPSRAVDMasked512Merging
+	OpAMD64VPSRAVQMasked128Merging
+	OpAMD64VPSRAVQMasked256Merging
+	OpAMD64VPSRAVQMasked512Merging
+	OpAMD64VPSRAVWMasked128Merging
+	OpAMD64VPSRAVWMasked256Merging
+	OpAMD64VPSRAVWMasked512Merging
+	OpAMD64VPSRLVDMasked128Merging
+	OpAMD64VPSRLVDMasked256Merging
+	OpAMD64VPSRLVDMasked512Merging
+	OpAMD64VPSRLVQMasked128Merging
+	OpAMD64VPSRLVQMasked256Merging
+	OpAMD64VPSRLVQMasked512Merging
+	OpAMD64VPSRLVWMasked128Merging
+	OpAMD64VPSRLVWMasked256Merging
+	OpAMD64VPSRLVWMasked512Merging
+	OpAMD64VPSUBBMasked128Merging
+	OpAMD64VPSUBBMasked256Merging
+	OpAMD64VPSUBBMasked512Merging
+	OpAMD64VPSUBDMasked128Merging
+	OpAMD64VPSUBDMasked256Merging
+	OpAMD64VPSUBDMasked512Merging
+	OpAMD64VPSUBQMasked128Merging
+	OpAMD64VPSUBQMasked256Merging
+	OpAMD64VPSUBQMasked512Merging
+	OpAMD64VPSUBSBMasked128Merging
+	OpAMD64VPSUBSBMasked256Merging
+	OpAMD64VPSUBSBMasked512Merging
+	OpAMD64VPSUBSWMasked128Merging
+	OpAMD64VPSUBSWMasked256Merging
+	OpAMD64VPSUBSWMasked512Merging
+	OpAMD64VPSUBUSBMasked128Merging
+	OpAMD64VPSUBUSBMasked256Merging
+	OpAMD64VPSUBUSBMasked512Merging
+	OpAMD64VPSUBUSWMasked128Merging
+	OpAMD64VPSUBUSWMasked256Merging
+	OpAMD64VPSUBUSWMasked512Merging
+	OpAMD64VPSUBWMasked128Merging
+	OpAMD64VPSUBWMasked256Merging
+	OpAMD64VPSUBWMasked512Merging
+	OpAMD64VPXORDMasked128Merging
+	OpAMD64VPXORDMasked256Merging
+	OpAMD64VPXORDMasked512Merging
+	OpAMD64VPXORQMasked128Merging
+	OpAMD64VPXORQMasked256Merging
+	OpAMD64VPXORQMasked512Merging
+	OpAMD64VRCP14PDMasked128Merging
+	OpAMD64VRCP14PDMasked256Merging
+	OpAMD64VRCP14PDMasked512Merging
+	OpAMD64VRCP14PSMasked128Merging
+	OpAMD64VRCP14PSMasked256Merging
+	OpAMD64VRCP14PSMasked512Merging
+	OpAMD64VRSQRT14PDMasked128Merging
+	OpAMD64VRSQRT14PDMasked256Merging
+	OpAMD64VRSQRT14PDMasked512Merging
+	OpAMD64VRSQRT14PSMasked128Merging
+	OpAMD64VRSQRT14PSMasked256Merging
+	OpAMD64VRSQRT14PSMasked512Merging
+	OpAMD64VSCALEFPDMasked128Merging
+	OpAMD64VSCALEFPDMasked256Merging
+	OpAMD64VSCALEFPDMasked512Merging
+	OpAMD64VSCALEFPSMasked128Merging
+	OpAMD64VSCALEFPSMasked256Merging
+	OpAMD64VSCALEFPSMasked512Merging
+	OpAMD64VSQRTPDMasked128Merging
+	OpAMD64VSQRTPDMasked256Merging
+	OpAMD64VSQRTPDMasked512Merging
+	OpAMD64VSQRTPSMasked128Merging
+	OpAMD64VSQRTPSMasked256Merging
+	OpAMD64VSQRTPSMasked512Merging
+	OpAMD64VSUBPDMasked128Merging
+	OpAMD64VSUBPDMasked256Merging
+	OpAMD64VSUBPDMasked512Merging
+	OpAMD64VSUBPSMasked128Merging
+	OpAMD64VSUBPSMasked256Merging
+	OpAMD64VSUBPSMasked512Merging
+	OpAMD64VPROLDMasked128Merging
+	OpAMD64VPROLDMasked256Merging
+	OpAMD64VPROLDMasked512Merging
+	OpAMD64VPROLQMasked128Merging
+	OpAMD64VPROLQMasked256Merging
+	OpAMD64VPROLQMasked512Merging
+	OpAMD64VPRORDMasked128Merging
+	OpAMD64VPRORDMasked256Merging
+	OpAMD64VPRORDMasked512Merging
+	OpAMD64VPRORQMasked128Merging
+	OpAMD64VPRORQMasked256Merging
+	OpAMD64VPRORQMasked512Merging
+	OpAMD64VPSHLDDMasked128Merging
+	OpAMD64VPSHLDDMasked256Merging
+	OpAMD64VPSHLDDMasked512Merging
+	OpAMD64VPSHLDQMasked128Merging
+	OpAMD64VPSHLDQMasked256Merging
+	OpAMD64VPSHLDQMasked512Merging
+	OpAMD64VPSHLDWMasked128Merging
+	OpAMD64VPSHLDWMasked256Merging
+	OpAMD64VPSHLDWMasked512Merging
+	OpAMD64VPSHRDDMasked128Merging
+	OpAMD64VPSHRDDMasked256Merging
+	OpAMD64VPSHRDDMasked512Merging
+	OpAMD64VPSHRDQMasked128Merging
+	OpAMD64VPSHRDQMasked256Merging
+	OpAMD64VPSHRDQMasked512Merging
+	OpAMD64VPSHRDWMasked128Merging
+	OpAMD64VPSHRDWMasked256Merging
+	OpAMD64VPSHRDWMasked512Merging
+	OpAMD64VPSHUFDMasked128Merging
+	OpAMD64VPSHUFDMasked256Merging
+	OpAMD64VPSHUFDMasked512Merging
+	OpAMD64VPSHUFHWMasked128Merging
+	OpAMD64VPSHUFHWMasked256Merging
+	OpAMD64VPSHUFHWMasked512Merging
+	OpAMD64VPSLLDMasked128constMerging
+	OpAMD64VPSLLDMasked256constMerging
+	OpAMD64VPSLLDMasked512constMerging
+	OpAMD64VPSLLQMasked128constMerging
+	OpAMD64VPSLLQMasked256constMerging
+	OpAMD64VPSLLQMasked512constMerging
+	OpAMD64VPSLLWMasked128constMerging
+	OpAMD64VPSLLWMasked256constMerging
+	OpAMD64VPSLLWMasked512constMerging
+	OpAMD64VPSRADMasked128constMerging
+	OpAMD64VPSRADMasked256constMerging
+	OpAMD64VPSRADMasked512constMerging
+	OpAMD64VPSRAQMasked128constMerging
+	OpAMD64VPSRAQMasked256constMerging
+	OpAMD64VPSRAQMasked512constMerging
+	OpAMD64VPSRAWMasked128constMerging
+	OpAMD64VPSRAWMasked256constMerging
+	OpAMD64VPSRAWMasked512constMerging
+	OpAMD64VPSRLDMasked128constMerging
+	OpAMD64VPSRLDMasked256constMerging
+	OpAMD64VPSRLDMasked512constMerging
+	OpAMD64VPSRLQMasked128constMerging
+	OpAMD64VPSRLQMasked256constMerging
+	OpAMD64VPSRLQMasked512constMerging
+	OpAMD64VPSRLWMasked128constMerging
+	OpAMD64VPSRLWMasked256constMerging
+	OpAMD64VPSRLWMasked512constMerging
+	OpAMD64VREDUCEPDMasked128Merging
+	OpAMD64VREDUCEPDMasked256Merging
+	OpAMD64VREDUCEPDMasked512Merging
+	OpAMD64VREDUCEPSMasked128Merging
+	OpAMD64VREDUCEPSMasked256Merging
+	OpAMD64VREDUCEPSMasked512Merging
+	OpAMD64VRNDSCALEPDMasked128Merging
+	OpAMD64VRNDSCALEPDMasked256Merging
+	OpAMD64VRNDSCALEPDMasked512Merging
+	OpAMD64VRNDSCALEPSMasked128Merging
+	OpAMD64VRNDSCALEPSMasked256Merging
+	OpAMD64VRNDSCALEPSMasked512Merging
 
 	OpARMADD
 	OpARMADDconst
@@ -20102,7 +20539,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20118,7 +20555,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20134,7 +20571,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20195,7 +20632,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20211,7 +20648,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20227,7 +20664,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20409,7 +20846,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVAESIMC,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20422,7 +20859,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVBROADCASTSD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20435,7 +20872,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVBROADCASTSD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20476,7 +20913,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVBROADCASTSS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20489,7 +20926,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVBROADCASTSS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20502,7 +20939,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVBROADCASTSS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20641,7 +21078,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTPS2UDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20654,7 +21091,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTPS2UDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20667,7 +21104,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTPS2UDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20722,7 +21159,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTTPS2DQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20735,7 +21172,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTTPS2DQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -20748,7 +21185,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVCVTTPS2DQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -20846,7 +21283,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20861,7 +21298,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20876,7 +21313,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20933,7 +21370,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20948,7 +21385,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -20963,7 +21400,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21698,7 +22135,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21713,7 +22150,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21728,7 +22165,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21901,7 +22338,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21917,7 +22354,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21933,7 +22370,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -21994,7 +22431,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22010,7 +22447,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22026,7 +22463,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22087,7 +22524,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22103,7 +22540,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22119,7 +22556,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22180,7 +22617,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22196,7 +22633,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22212,7 +22649,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22389,90 +22826,6 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
-		name:   "VMOVUPDMasked128",
-		argLen: 2,
-		asm:    x86.AVMOVUPD,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
-		name:   "VMOVUPDMasked256",
-		argLen: 2,
-		asm:    x86.AVMOVUPD,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
-		name:   "VMOVUPDMasked512",
-		argLen: 2,
-		asm:    x86.AVMOVUPD,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
-		name:   "VMOVUPSMasked128",
-		argLen: 2,
-		asm:    x86.AVMOVUPS,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
-		name:   "VMOVUPSMasked256",
-		argLen: 2,
-		asm:    x86.AVMOVUPS,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
-		name:   "VMOVUPSMasked512",
-		argLen: 2,
-		asm:    x86.AVMOVUPS,
-		reg: regInfo{
-			inputs: []inputInfo{
-				{1, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-			outputs: []outputInfo{
-				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
-			},
-		},
-	},
-	{
 		name:        "VMULPD128",
 		argLen:      2,
 		commutative: true,
@@ -22525,7 +22878,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22541,7 +22894,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22557,7 +22910,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22618,7 +22971,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22634,7 +22987,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22650,7 +23003,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -22664,7 +23017,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22677,7 +23030,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22690,7 +23043,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -22745,7 +23098,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22758,7 +23111,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22771,7 +23124,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -22826,7 +23179,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -22839,7 +23192,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -22852,7 +23205,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -22907,7 +23260,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22920,7 +23273,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -22933,7 +23286,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPABSW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -23031,7 +23384,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23046,7 +23399,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23061,7 +23414,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23118,7 +23471,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23133,7 +23486,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23148,7 +23501,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23209,7 +23562,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23225,7 +23578,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23241,7 +23594,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23302,7 +23655,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23318,7 +23671,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23334,7 +23687,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23395,7 +23748,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23411,7 +23764,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23427,7 +23780,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23488,7 +23841,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23504,7 +23857,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23520,7 +23873,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23581,7 +23934,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23597,7 +23950,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23613,7 +23966,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23674,7 +24027,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23690,7 +24043,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23706,7 +24059,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23767,7 +24120,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23783,7 +24136,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23799,7 +24152,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23860,7 +24213,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23876,7 +24229,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23892,7 +24245,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23953,7 +24306,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23969,7 +24322,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -23985,7 +24338,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24042,7 +24395,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24057,7 +24410,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24072,7 +24425,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24101,7 +24454,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24116,7 +24469,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24131,7 +24484,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24162,7 +24515,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24178,7 +24531,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24194,7 +24547,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24255,7 +24608,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24271,7 +24624,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24287,7 +24640,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24348,7 +24701,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24364,7 +24717,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24380,7 +24733,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24395,7 +24748,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24410,7 +24763,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24425,7 +24778,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24440,7 +24793,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -24484,7 +24837,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24497,7 +24850,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24510,7 +24863,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -24565,7 +24918,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24578,7 +24931,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24591,7 +24944,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -24646,7 +24999,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24659,7 +25012,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24672,7 +25025,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -24727,7 +25080,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24740,7 +25093,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -24753,7 +25106,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPBROADCASTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -25650,7 +26003,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -25665,7 +26018,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -25708,7 +26061,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -25723,7 +26076,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26360,7 +26713,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26375,7 +26728,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26418,7 +26771,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26433,7 +26786,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26476,7 +26829,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26491,7 +26844,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26548,7 +26901,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26563,7 +26916,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26578,7 +26931,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -26928,7 +27281,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -26941,7 +27294,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -26954,7 +27307,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -27009,7 +27362,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -27022,7 +27375,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -27035,7 +27388,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPLZCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -27133,7 +27486,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27148,7 +27501,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27163,7 +27516,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27220,7 +27573,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27235,7 +27588,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27250,7 +27603,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27311,7 +27664,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27327,7 +27680,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27343,7 +27696,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27404,7 +27757,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27420,7 +27773,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27436,7 +27789,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27497,7 +27850,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27513,7 +27866,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27529,7 +27882,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27590,7 +27943,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27606,7 +27959,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27622,7 +27975,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27683,7 +28036,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27699,7 +28052,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27715,7 +28068,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27776,7 +28129,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27792,7 +28145,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27808,7 +28161,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27869,7 +28222,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27885,7 +28238,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27901,7 +28254,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27962,7 +28315,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27978,7 +28331,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -27994,7 +28347,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28055,7 +28408,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28071,7 +28424,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28087,7 +28440,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28148,7 +28501,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28164,7 +28517,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28180,7 +28533,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28241,7 +28594,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28257,7 +28610,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28273,7 +28626,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28334,7 +28687,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28350,7 +28703,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28366,7 +28719,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28427,7 +28780,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28443,7 +28796,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28459,7 +28812,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28520,7 +28873,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28536,7 +28889,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28552,7 +28905,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28613,7 +28966,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28629,7 +28982,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28645,7 +28998,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28706,7 +29059,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28722,7 +29075,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28738,7 +29091,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -28752,7 +29105,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVDB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28779,7 +29132,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28792,7 +29145,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28833,7 +29186,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVQB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28860,7 +29213,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28873,7 +29226,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28914,7 +29267,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVQW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28941,7 +29294,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSDB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28968,7 +29321,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -28981,7 +29334,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29022,7 +29375,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSQB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29049,7 +29402,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29062,7 +29415,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29103,7 +29456,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSQW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29130,7 +29483,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29143,7 +29496,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29184,7 +29537,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29197,7 +29550,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29210,7 +29563,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29265,7 +29618,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29278,7 +29631,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29291,7 +29644,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29346,7 +29699,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29359,7 +29712,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29372,7 +29725,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29427,7 +29780,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29440,7 +29793,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29453,7 +29806,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29508,7 +29861,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29521,7 +29874,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29534,7 +29887,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29589,7 +29942,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29602,7 +29955,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29615,7 +29968,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVSXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29670,7 +30023,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSDB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29697,7 +30050,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29710,7 +30063,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSDW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29751,7 +30104,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSQB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29778,7 +30131,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29791,7 +30144,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSQD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29832,7 +30185,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSQW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29859,7 +30212,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29872,7 +30225,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVUSWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29913,7 +30266,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29926,7 +30279,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVWB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -29967,7 +30320,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29980,7 +30333,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -29993,7 +30346,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30048,7 +30401,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30061,7 +30414,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30074,7 +30427,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30129,7 +30482,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30142,7 +30495,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30155,7 +30508,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXBW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30210,7 +30563,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30223,7 +30576,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30236,7 +30589,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXDQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30291,7 +30644,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30304,7 +30657,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30317,7 +30670,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30372,7 +30725,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30385,7 +30738,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -30398,7 +30751,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPMOVZXWQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30530,7 +30883,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30546,7 +30899,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30562,7 +30915,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30623,7 +30976,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30639,7 +30992,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30655,7 +31008,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30716,7 +31069,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30732,7 +31085,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30748,7 +31101,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30809,7 +31162,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30825,7 +31178,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30841,7 +31194,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30902,7 +31255,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30918,7 +31271,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30934,7 +31287,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -30978,7 +31331,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -30991,7 +31344,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31004,7 +31357,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTB,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31059,7 +31412,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31072,7 +31425,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31085,7 +31438,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31140,7 +31493,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31153,7 +31506,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31166,7 +31519,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31221,7 +31574,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31234,7 +31587,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31247,7 +31600,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVPOPCNTW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -31349,7 +31702,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31365,7 +31718,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31381,7 +31734,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31412,7 +31765,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31428,7 +31781,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31444,7 +31797,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31501,7 +31854,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31516,7 +31869,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31531,7 +31884,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31588,7 +31941,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31603,7 +31956,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31618,7 +31971,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31675,7 +32028,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31690,7 +32043,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31705,7 +32058,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31762,7 +32115,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31777,7 +32130,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -31792,7 +32145,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32485,7 +32838,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32500,7 +32853,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32515,7 +32868,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32830,7 +33183,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32845,7 +33198,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32860,7 +33213,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32917,7 +33270,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32932,7 +33285,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -32947,7 +33300,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33004,7 +33357,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33019,7 +33372,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33034,7 +33387,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33352,7 +33705,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33367,7 +33720,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33382,7 +33735,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33439,7 +33792,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33454,7 +33807,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33469,7 +33822,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33526,7 +33879,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33541,7 +33894,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33556,7 +33909,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33874,7 +34227,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33889,7 +34242,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33904,7 +34257,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33961,7 +34314,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33976,7 +34329,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -33991,7 +34344,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34048,7 +34401,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34063,7 +34416,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34078,7 +34431,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34222,7 +34575,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34237,7 +34590,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34252,7 +34605,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34309,7 +34662,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34324,7 +34677,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34339,7 +34692,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34396,7 +34749,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34411,7 +34764,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34426,7 +34779,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34483,7 +34836,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34498,7 +34851,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34513,7 +34866,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34570,7 +34923,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34585,7 +34938,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34600,7 +34953,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34657,7 +35010,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34672,7 +35025,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34687,7 +35040,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34744,7 +35097,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34759,7 +35112,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34774,7 +35127,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34831,7 +35184,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34846,7 +35199,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -34861,7 +35214,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35174,7 +35527,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35190,7 +35543,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35206,7 +35559,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35237,7 +35590,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35253,7 +35606,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35269,7 +35622,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35283,7 +35636,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCP14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35296,7 +35649,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCP14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35309,7 +35662,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCP14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35364,7 +35717,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCP14PS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35419,7 +35772,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCPPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35432,7 +35785,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRCPPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35445,7 +35798,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRT14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35458,7 +35811,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRT14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35471,7 +35824,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRT14PD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35526,7 +35879,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRT14PS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35581,7 +35934,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRTPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35594,7 +35947,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVRSQRTPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35650,7 +36003,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35665,7 +36018,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35680,7 +36033,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35737,7 +36090,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35752,7 +36105,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35767,7 +36120,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -35781,7 +36134,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35794,7 +36147,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35807,7 +36160,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35862,7 +36215,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35875,7 +36228,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -35888,7 +36241,7 @@ var opcodeTable = [...]opInfo{
 		asm:    x86.AVSQRTPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -35986,7 +36339,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36001,7 +36354,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36016,7 +36369,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36073,7 +36426,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36088,7 +36441,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36103,7 +36456,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36134,7 +36487,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVAESKEYGENASSIST,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -36346,7 +36699,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVEXTRACTF64X4,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -36360,7 +36713,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVEXTRACTF128,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -36374,7 +36727,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVEXTRACTI64X4,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -36388,7 +36741,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVEXTRACTI128,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -36448,7 +36801,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36464,7 +36817,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36480,7 +36833,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36541,7 +36894,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36557,7 +36910,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -36573,7 +36926,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37322,7 +37675,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37336,7 +37689,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37350,7 +37703,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37409,7 +37762,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37423,7 +37776,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37437,7 +37790,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPROLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37496,7 +37849,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37510,7 +37863,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37524,7 +37877,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37583,7 +37936,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37597,7 +37950,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37611,7 +37964,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPRORQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -37716,7 +38069,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37732,7 +38085,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37748,7 +38101,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37809,7 +38162,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37825,7 +38178,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37841,7 +38194,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37902,7 +38255,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37918,7 +38271,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37934,7 +38287,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -37995,7 +38348,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38011,7 +38364,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38027,7 +38380,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38088,7 +38441,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38104,7 +38457,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38120,7 +38473,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38181,7 +38534,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38197,7 +38550,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38213,7 +38566,7 @@ var opcodeTable = [...]opInfo{
 		reg: regInfo{
 			inputs: []inputInfo{
 				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
-				{0, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
@@ -38228,7 +38581,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38242,7 +38595,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38256,7 +38609,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38315,7 +38668,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFHW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38329,7 +38682,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFHW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38343,7 +38696,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSHUFHW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38402,7 +38755,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38416,7 +38769,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38430,7 +38783,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38489,7 +38842,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38503,7 +38856,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38517,7 +38870,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38576,7 +38929,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38590,7 +38943,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38604,7 +38957,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSLLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38663,7 +39016,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38677,7 +39030,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38691,7 +39044,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38750,7 +39103,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38764,7 +39117,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38778,7 +39131,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38837,7 +39190,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38851,7 +39204,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38865,7 +39218,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRAW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -38924,7 +39277,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38938,7 +39291,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -38952,7 +39305,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39011,7 +39364,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39025,7 +39378,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39039,7 +39392,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLQ,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39098,7 +39451,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39112,7 +39465,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39126,7 +39479,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVPSRLW,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39287,7 +39640,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39301,7 +39654,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39315,7 +39668,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39374,7 +39727,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39388,7 +39741,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39402,7 +39755,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVREDUCEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39461,7 +39814,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39475,7 +39828,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39489,7 +39842,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39548,7 +39901,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39562,7 +39915,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39576,7 +39929,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVRNDSCALEPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 281474976645120}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
 			},
 			outputs: []outputInfo{
 				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
@@ -39635,7 +39988,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVROUNDPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39649,7 +40002,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVROUNDPD,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39663,7 +40016,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVROUNDPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -39677,7 +40030,7 @@ var opcodeTable = [...]opInfo{
 		asm:     x86.AVROUNDPS,
 		reg: regInfo{
 			inputs: []inputInfo{
-				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
 			},
 			outputs: []outputInfo{
 				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
@@ -49020,6 +49373,7421 @@ var opcodeTable = [...]opInfo{
 			},
 		},
 	},
+	{
+		name:         "VADDPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VADDPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VADDPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VADDPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VADDPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VADDPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVADDPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VBROADCASTSDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVBROADCASTSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VBROADCASTSDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVBROADCASTSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VBROADCASTSSMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVBROADCASTSS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VBROADCASTSSMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVBROADCASTSS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VBROADCASTSSMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVBROADCASTSS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTPS2UDQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTPS2UDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTPS2UDQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTPS2UDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTPS2UDQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTPS2UDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTTPS2DQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTTPS2DQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTTPS2DQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTTPS2DQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VCVTTPS2DQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVCVTTPS2DQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VDIVPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVDIVPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VGF2P8MULBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVGF2P8MULB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VGF2P8MULBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVGF2P8MULB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VGF2P8MULBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVGF2P8MULB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMAXPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMAXPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMINPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMINPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VMULPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVMULPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSBMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPABSWMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPABSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKSSDWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKSSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKSSDWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKSSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKSSDWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKSSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKUSDWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKUSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKUSDWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKUSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPACKUSDWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPACKUSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDUSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPADDWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPADDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPANDQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPANDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPAVGWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPAVGW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTBMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPBROADCASTWMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPBROADCASTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPLZCNTQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPLZCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDUBSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDUBSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDUBSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDWDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDWDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMADDWDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMADDWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMAXUWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMAXUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMINUWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMINUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVDBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVDWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVDWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVQBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVQB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVQDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVQDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVQWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVQW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSDBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSDWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSDWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSQBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSQB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSQDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSQDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSQWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSQW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSWBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSWBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXBWMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXDQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXDQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXDQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVSXWQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVSXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSDBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSDB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSDWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSDWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSQBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSQB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSQDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSQDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSQD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSQWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSQW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSWBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVUSWBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVUSWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVWBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVWBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVWB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXBWMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXDQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXDQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXDQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMOVZXWQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPMOVZXWQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHUWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHUWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHUWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHUW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULHWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPMULLWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPMULLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTBMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTBMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTBMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTQMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTQMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTQMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTWMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTWMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPOPCNTWMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPOPCNTW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPORQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLVQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPROLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORVQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPRORVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLVWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSLLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAVWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRAVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLVWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSRLVW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSBMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSBMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSBMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSB,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBUSWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBUSW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBWMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBWMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSUBWMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSUBW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORQMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORQMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPXORQMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPXORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PSMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PSMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRCP14PSMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRCP14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PSMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PSMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRSQRT14PSMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRSQRT14PS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSCALEFPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSCALEFPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPDMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPDMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPDMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPSMasked128Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPSMasked256Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSQRTPSMasked512Merging",
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVSQRTPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPDMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPDMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPDMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPSMasked128Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPSMasked256Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VSUBPSMasked512Merging",
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVSUBPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLQMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLQMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPROLQMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPROLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORQMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORQMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPRORQMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPRORQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDQMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDQMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDQMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDWMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDWMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHLDWMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHLDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDQMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDQMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDQMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDWMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDWMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHRDWMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       4,
+		resultInArg0: true,
+		asm:          x86.AVPSHRDW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{3, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{2, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFHWMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFHWMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSHUFHWMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSHUFHW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLDMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLDMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLDMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLQMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLQMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLQMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLWMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLWMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSLLWMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSLLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRADMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRADMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRADMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAQMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAQMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAQMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAWMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAWMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRAWMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRAW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLDMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLDMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLDMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLQMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLQMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLQMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLQ,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLWMasked128constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLWMasked256constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VPSRLWMasked512constMerging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVPSRLW,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPSMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPSMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VREDUCEPSMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVREDUCEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPDMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPDMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPDMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPSMasked128Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPSMasked256Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
+	{
+		name:         "VRNDSCALEPSMasked512Merging",
+		auxType:      auxUInt8,
+		argLen:       3,
+		resultInArg0: true,
+		asm:          x86.AVRNDSCALEPS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{2, 71494644084506624}, // K1 K2 K3 K4 K5 K6 K7
+				{0, 281472829161472},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+				{1, 281474976645120},   // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+			outputs: []outputInfo{
+				{0, 281472829161472}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X16 X17 X18 X19 X20 X21 X22 X23 X24 X25 X26 X27 X28 X29 X30 X31
+			},
+		},
+	},
 
 	{
 		name:        "ADD",
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 6644615f95..ef0bddc70a 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -889,10 +889,18 @@ func rewriteValueAMD64(v *Value) bool {
 		return rewriteValueAMD64_OpAMD64VPANDQMasked256(v)
 	case OpAMD64VPANDQMasked512:
 		return rewriteValueAMD64_OpAMD64VPANDQMasked512(v)
+	case OpAMD64VPBLENDMBMasked512:
+		return rewriteValueAMD64_OpAMD64VPBLENDMBMasked512(v)
 	case OpAMD64VPBLENDMDMasked512:
 		return rewriteValueAMD64_OpAMD64VPBLENDMDMasked512(v)
 	case OpAMD64VPBLENDMQMasked512:
 		return rewriteValueAMD64_OpAMD64VPBLENDMQMasked512(v)
+	case OpAMD64VPBLENDMWMasked512:
+		return rewriteValueAMD64_OpAMD64VPBLENDMWMasked512(v)
+	case OpAMD64VPBLENDVB128:
+		return rewriteValueAMD64_OpAMD64VPBLENDVB128(v)
+	case OpAMD64VPBLENDVB256:
+		return rewriteValueAMD64_OpAMD64VPBLENDVB256(v)
 	case OpAMD64VPBROADCASTB128:
 		return rewriteValueAMD64_OpAMD64VPBROADCASTB128(v)
 	case OpAMD64VPBROADCASTB256:
@@ -40256,10 +40264,879 @@ func rewriteValueAMD64_OpAMD64VPANDQMasked512(v *Value) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpAMD64VPBLENDMBMasked512(v *Value) bool {
+	v_2 := v.Args[2]
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	// match: (VPBLENDMBMasked512 dst (VGF2P8MULB512 x y) mask)
+	// result: (VGF2P8MULBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VGF2P8MULB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VGF2P8MULBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPOPCNTB512 x) mask)
+	// result: (VPOPCNTBMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTB512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPOPCNTBMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPSUBSB512 x y) mask)
+	// result: (VPSUBSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPSUBB512 x y) mask)
+	// result: (VPSUBBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPMINSB512 x y) mask)
+	// result: (VPMINSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPADDB512 x y) mask)
+	// result: (VPADDBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPMAXUB512 x y) mask)
+	// result: (VPMAXUBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXUBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPADDUSB512 x y) mask)
+	// result: (VPADDUSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDUSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPAVGB512 x y) mask)
+	// result: (VPAVGBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPAVGBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPMINUB512 x y) mask)
+	// result: (VPMINUBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINUBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPMAXSB512 x y) mask)
+	// result: (VPMAXSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPSUBUSB512 x y) mask)
+	// result: (VPSUBUSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBUSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPSHUFB512 x y) mask)
+	// result: (VPSHUFBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHUFBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPABSB512 x) mask)
+	// result: (VPABSBMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSB512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPABSBMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMBMasked512 dst (VPADDSB512 x y) mask)
+	// result: (VPADDSBMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSB512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDSBMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	return false
+}
 func rewriteValueAMD64_OpAMD64VPBLENDMDMasked512(v *Value) bool {
 	v_2 := v.Args[2]
 	v_1 := v.Args[1]
 	v_0 := v.Args[0]
+	// match: (VPBLENDMDMasked512 dst (VPMOVSDW256 x) mask)
+	// result: (VPMOVSDWMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSDW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSDWMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPLZCNTD512 x) mask)
+	// result: (VPLZCNTDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPLZCNTDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMULLD512 x y) mask)
+	// result: (VPMULLDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMULLDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VMAXPS512 x y) mask)
+	// result: (VMAXPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMAXPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMOVUSDB128 x) mask)
+	// result: (VPMOVUSDBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSDB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSDBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VRSQRT14PS512 x) mask)
+	// result: (VRSQRT14PSMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRSQRT14PS512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRSQRT14PSMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMOVDW256 x) mask)
+	// result: (VPMOVDWMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVDW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVDWMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VRCP14PS512 x) mask)
+	// result: (VRCP14PSMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRCP14PS512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRCP14PSMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VREDUCEPS512 [a] x) mask)
+	// result: (VREDUCEPSMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPS512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VREDUCEPSMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VDIVPS512 x y) mask)
+	// result: (VDIVPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VDIVPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSRLVD512 x y) mask)
+	// result: (VPSRLVDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRLVDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSUBD512 x y) mask)
+	// result: (VPSUBDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPROLD512 [a] x) mask)
+	// result: (VPROLDMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPROLDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPORD512 x y) mask)
+	// result: (VPORDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPORD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPORDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSHLDD512 [a] x y) mask)
+	// result: (VPSHLDDMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHLDDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPACKUSDW512 x y) mask)
+	// result: (VPACKUSDWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKUSDW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPACKUSDWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMAXSD512 x y) mask)
+	// result: (VPMAXSDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXSDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VADDPS512 x y) mask)
+	// result: (VADDPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VADDPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMOVUSDW256 x) mask)
+	// result: (VPMOVUSDWMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSDW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSDWMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMOVSDB128 x) mask)
+	// result: (VPMOVSDBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSDB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSDBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VSUBPS512 x y) mask)
+	// result: (VSUBPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSUBPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMAXUD512 x y) mask)
+	// result: (VPMAXUDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXUDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPRORD512 [a] x) mask)
+	// result: (VPRORDMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPRORDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPROLVD512 x y) mask)
+	// result: (VPROLVDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPROLVDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VCVTTPS2DQ512 x) mask)
+	// result: (VCVTTPS2DQMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTTPS2DQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VCVTTPS2DQMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPACKSSDW512 x y) mask)
+	// result: (VPACKSSDWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKSSDW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPACKSSDWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPRORVD512 x y) mask)
+	// result: (VPRORVDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPRORVDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPADDD512 x y) mask)
+	// result: (VPADDDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VRNDSCALEPS512 [a] x) mask)
+	// result: (VRNDSCALEPSMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPS512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRNDSCALEPSMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VCVTPS2UDQ512 x) mask)
+	// result: (VCVTPS2UDQMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTPS2UDQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VCVTPS2UDQMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSHRDD512 [a] x y) mask)
+	// result: (VPSHRDDMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHRDDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPOPCNTD512 x) mask)
+	// result: (VPOPCNTDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPOPCNTDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMOVDB128 x) mask)
+	// result: (VPMOVDBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVDB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVDBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSRAD512const [a] x) mask)
+	// result: (VPSRADMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAD512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRADMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VMINPS512 x y) mask)
+	// result: (VMINPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMINPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPANDD512 x y) mask)
+	// result: (VPANDDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPANDD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPANDDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSHUFD512 [a] x) mask)
+	// result: (VPSHUFDMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHUFDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMINSD512 x y) mask)
+	// result: (VPMINSDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINSDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSRAVD512 x y) mask)
+	// result: (VPSRAVDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRAVDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPXORD512 x y) mask)
+	// result: (VPXORDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPXORD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPXORDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSLLVD512 x y) mask)
+	// result: (VPSLLVDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLVDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPSLLD512const [a] x) mask)
+	// result: (VPSLLDMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLD512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLDMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPMINUD512 x y) mask)
+	// result: (VPMINUDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINUDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VSCALEFPS512 x y) mask)
+	// result: (VSCALEFPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSCALEFPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VSQRTPS512 x) mask)
+	// result: (VSQRTPSMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPS512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSQRTPSMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VPABSD512 x) mask)
+	// result: (VPABSDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPABSDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMDMasked512 dst (VMULPS512 x y) mask)
+	// result: (VMULPSMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPS512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMULPSMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
 	// match: (VPBLENDMDMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask)
 	// cond: canMergeLoad(v, l) && clobber(l)
 	// result: (VPBLENDMDMasked512load {sym} [off] x ptr mask mem)
@@ -40289,6 +41166,631 @@ func rewriteValueAMD64_OpAMD64VPBLENDMQMasked512(v *Value) bool {
 	v_2 := v.Args[2]
 	v_1 := v.Args[1]
 	v_0 := v.Args[0]
+	// match: (VPBLENDMQMasked512 dst (VPSLLQ512const [a] x) mask)
+	// result: (VPSLLQMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLQ512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLQMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSUBQ512 x y) mask)
+	// result: (VPSUBQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPROLQ512 [a] x) mask)
+	// result: (VPROLQMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLQ512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPROLQMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSLLVQ512 x y) mask)
+	// result: (VPSLLVQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLVQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVUSQB128 x) mask)
+	// result: (VPMOVUSQBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSQB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSQBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPADDQ512 x y) mask)
+	// result: (VPADDQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VRNDSCALEPD512 [a] x) mask)
+	// result: (VRNDSCALEPDMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRNDSCALEPDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPABSQ512 x) mask)
+	// result: (VPABSQMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPABSQMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVUSQD256 x) mask)
+	// result: (VPMOVUSQDMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSQD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSQDMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VADDPD512 x y) mask)
+	// result: (VADDPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VADDPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VRCP14PD512 x) mask)
+	// result: (VRCP14PDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRCP14PD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRCP14PDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSRLVQ512 x y) mask)
+	// result: (VPSRLVQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRLVQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPRORVQ512 x y) mask)
+	// result: (VPRORVQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPRORVQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSRAVQ512 x y) mask)
+	// result: (VPSRAVQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRAVQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPANDQ512 x y) mask)
+	// result: (VPANDQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPANDQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPANDQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVQB128 x) mask)
+	// result: (VPMOVQBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVQB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVQBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSHLDQ512 [a] x y) mask)
+	// result: (VPSHLDQMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDQ512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHLDQMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VDIVPD512 x y) mask)
+	// result: (VDIVPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VDIVPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPROLVQ512 x y) mask)
+	// result: (VPROLVQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPROLVQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPRORQ512 [a] x) mask)
+	// result: (VPRORQMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORQ512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPRORQMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMINSQ512 x y) mask)
+	// result: (VPMINSQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINSQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VSQRTPD512 x) mask)
+	// result: (VSQRTPDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSQRTPDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVSQD256 x) mask)
+	// result: (VPMOVSQDMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSQD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSQDMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VMINPD512 x y) mask)
+	// result: (VMINPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMINPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMULLQ512 x y) mask)
+	// result: (VPMULLQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMULLQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VMAXPD512 x y) mask)
+	// result: (VMAXPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMAXPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VMULPD512 x y) mask)
+	// result: (VMULPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VMULPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPORQ512 x y) mask)
+	// result: (VPORQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPORQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPORQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVUSQW128 x) mask)
+	// result: (VPMOVUSQWMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSQW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSQWMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VREDUCEPD512 [a] x) mask)
+	// result: (VREDUCEPDMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPD512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VREDUCEPDMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPOPCNTQ512 x) mask)
+	// result: (VPOPCNTQMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPOPCNTQMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPXORQ512 x y) mask)
+	// result: (VPXORQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPXORQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPXORQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVQD256 x) mask)
+	// result: (VPMOVQDMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVQD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVQDMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMAXUQ512 x y) mask)
+	// result: (VPMAXUQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXUQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VSUBPD512 x y) mask)
+	// result: (VSUBPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSUBPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVQW128 x) mask)
+	// result: (VPMOVQWMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVQW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVQWMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSHRDQ512 [a] x y) mask)
+	// result: (VPSHRDQMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDQ512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHRDQMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPLZCNTQ512 x) mask)
+	// result: (VPLZCNTQMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPLZCNTQMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VSCALEFPD512 x y) mask)
+	// result: (VSCALEFPDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VSCALEFPDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVSQW128 x) mask)
+	// result: (VPMOVSQWMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSQW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSQWMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMINUQ512 x y) mask)
+	// result: (VPMINUQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINUQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMOVSQB128 x) mask)
+	// result: (VPMOVSQBMasked128Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSQB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSQBMasked128Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VRSQRT14PD512 x) mask)
+	// result: (VRSQRT14PDMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRSQRT14PD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VRSQRT14PDMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPSRAQ512const [a] x) mask)
+	// result: (VPSRAQMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAQ512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRAQMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMQMasked512 dst (VPMAXSQ512 x y) mask)
+	// result: (VPMAXSQMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSQ512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXSQMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
 	// match: (VPBLENDMQMasked512 x l:(VMOVDQUload512 {sym} [off] ptr mem) mask)
 	// cond: canMergeLoad(v, l) && clobber(l)
 	// result: (VPBLENDMQMasked512load {sym} [off] x ptr mask mem)
@@ -40314,6 +41816,6036 @@ func rewriteValueAMD64_OpAMD64VPBLENDMQMasked512(v *Value) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpAMD64VPBLENDMWMasked512(v *Value) bool {
+	v_2 := v.Args[2]
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	// match: (VPBLENDMWMasked512 dst (VPMAXSW512 x y) mask)
+	// result: (VPMAXSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMULHW512 x y) mask)
+	// result: (VPMULHWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMULHWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMOVWB256 x) mask)
+	// result: (VPMOVWBMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVWB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVWBMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMADDUBSW512 x y) mask)
+	// result: (VPMADDUBSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDUBSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMADDUBSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSHLDW512 [a] x y) mask)
+	// result: (VPSHLDWMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDW512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHLDWMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMULHUW512 x y) mask)
+	// result: (VPMULHUWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHUW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMULHUWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMOVUSWB256 x) mask)
+	// result: (VPMOVUSWBMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSWB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVUSWBMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMINSW512 x y) mask)
+	// result: (VPMINSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSRAVW512 x y) mask)
+	// result: (VPSRAVWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRAVWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPADDW512 x y) mask)
+	// result: (VPADDWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSHUFHW512 [a] x) mask)
+	// result: (VPSHUFHWMasked512Merging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFHW512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHUFHWMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSHRDW512 [a] x y) mask)
+	// result: (VPSHRDWMasked512Merging dst [a] x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDW512 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSHRDWMasked512Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSUBSW512 x y) mask)
+	// result: (VPSUBSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSUBUSW512 x y) mask)
+	// result: (VPSUBUSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBUSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSUBW512 x y) mask)
+	// result: (VPSUBWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSUBWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMADDWD512 x y) mask)
+	// result: (VPMADDWDMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDWD512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMADDWDMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSLLVW512 x y) mask)
+	// result: (VPSLLVWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLVWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPABSW512 x) mask)
+	// result: (VPABSWMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSW512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPABSWMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSRAW512const [a] x) mask)
+	// result: (VPSRAWMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAW512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRAWMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPADDUSW512 x y) mask)
+	// result: (VPADDUSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDUSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPOPCNTW512 x) mask)
+	// result: (VPOPCNTWMasked512Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTW512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPOPCNTWMasked512Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMINUW512 x y) mask)
+	// result: (VPMINUWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMINUWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPAVGW512 x y) mask)
+	// result: (VPAVGWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPAVGWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMOVSWB256 x) mask)
+	// result: (VPMOVSWBMasked256Merging dst x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSWB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMOVSWBMasked256Merging)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMAXUW512 x y) mask)
+	// result: (VPMAXUWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMAXUWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSRLVW512 x y) mask)
+	// result: (VPSRLVWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSRLVWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPSLLW512const [a] x) mask)
+	// result: (VPSLLWMasked512constMerging dst [a] x mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLW512const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPSLLWMasked512constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v.AddArg3(dst, x, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPADDSW512 x y) mask)
+	// result: (VPADDSWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPADDSWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	// match: (VPBLENDMWMasked512 dst (VPMULLW512 x y) mask)
+	// result: (VPMULLWMasked512Merging dst x y mask)
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLW512 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		v.reset(OpAMD64VPMULLWMasked512Merging)
+		v.AddArg4(dst, x, y, mask)
+		return true
+	}
+	return false
+}
+func rewriteValueAMD64_OpAMD64VPBLENDVB128(v *Value) bool {
+	v_2 := v.Args[2]
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (VPBLENDVB128 dst (VPMINUD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPROLQ128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLQMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLQ128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLQMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMADDUBSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMADDUBSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDUBSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMADDUBSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VBROADCASTSS256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VBROADCASTSSMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VBROADCASTSS256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VBROADCASTSSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBWMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINSQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBUSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBUSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBUSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBQMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXWQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWQMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMULLW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHLDQ128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDQMasked128Merging dst [a] x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDQ128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDQMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBQMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXSQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPOPCNTW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPRORVD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORVDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VBROADCASTSD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VBROADCASTSDMasked256Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VBROADCASTSD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VBROADCASTSDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXDQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXDQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXDQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAQ128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAQMasked128constMerging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAQ128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAQMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPACKUSDW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPACKUSDWMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKUSDW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPACKUSDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPLZCNTD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPLZCNTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPLZCNTDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXUD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPOPCNTB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VBROADCASTSD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VBROADCASTSDMasked512Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VBROADCASTSD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VBROADCASTSDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMINPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMINPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMINPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHRDW128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDWMasked128Merging dst [a] x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDW128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDWMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VADDPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VADDPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VADDPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXWD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWDMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXWQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWQMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VREDUCEPD128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VREDUCEPDMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VREDUCEPDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBDMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPRORQ128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORQMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORQ128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORQMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLVW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBW256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBWMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINSD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VADDPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VADDPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VADDPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBDMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXDQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXDQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXDQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPROLVD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLVDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRLVQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXSD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINUB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMULLQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTDMasked512Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMADDWD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMADDWDMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDWD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMADDWDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPROLD128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAD128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRADMasked128constMerging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAD128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRADMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBUSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBUSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBUSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDUSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDUSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDUSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBWMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXDQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXDQMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXDQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXDQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPROLVQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLVQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDUSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDUSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDUSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VRNDSCALEPS128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRNDSCALEPSMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPS128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRNDSCALEPSMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINUW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMULLD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHUFB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPRORD128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VCVTTPS2DQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VCVTTPS2DQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTTPS2DQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VCVTTPS2DQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMINPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMINPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMINPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSUBPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSUBPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSUBPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTB512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTBMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTB512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTBMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VRCP14PD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRCP14PDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRCP14PD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRCP14PDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXWD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWDMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTW256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTWMasked256Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTDMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VBROADCASTSS128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VBROADCASTSSMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VBROADCASTSS128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VBROADCASTSSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXDQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXDQMasked256Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXDQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXDQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBDMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHLDW128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDWMasked128Merging dst [a] x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDW128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDWMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXUQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHLDD128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDDMasked128Merging dst [a] x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSUBPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSUBPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSUBPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPRORVQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORVQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VBROADCASTSS512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VBROADCASTSSMasked512Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VBROADCASTSS512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VBROADCASTSSMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBDMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBQMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPABSW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSWMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXWQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWQMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VGF2P8MULB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VGF2P8MULBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VGF2P8MULB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VGF2P8MULBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPABSD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTB256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTBMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMAXPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMAXPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMAXPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINUQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMULPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMULPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMULPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMULHUW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULHUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHUW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULHUWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMULPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMULPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMULPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VCVTPS2UDQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VCVTPS2UDQMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTPS2UDQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VCVTPS2UDQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSCALEFPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSCALEFPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSCALEFPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLVQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBQMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXWD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWDMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VRSQRT14PD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRSQRT14PDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRSQRT14PD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRSQRT14PDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAW128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAWMasked128constMerging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAW128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAWMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMULHW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULHWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULHWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHRDD128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDDMasked128Merging dst [a] x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPADDSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMINSB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHUFD128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFDMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTQMasked512Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VREDUCEPS128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VREDUCEPSMasked128Merging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPS128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VREDUCEPSMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXWQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWQMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAVW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSQRTPD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSQRTPDMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSQRTPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPAVGW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPAVGWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPAVGWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VDIVPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VDIVPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VDIVPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VDIVPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VDIVPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VDIVPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPOPCNTD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTQMasked256Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VRNDSCALEPD128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRNDSCALEPDMasked128Merging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPD128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRNDSCALEPDMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXWQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWQMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPOPCNTQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPAVGB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPAVGBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPAVGBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBQMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXSW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBW256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBWMasked256Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBDMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHUFHW128 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFHWMasked128Merging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFHW128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFHWMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLW128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLWMasked128constMerging dst [a] x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLW128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLWMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLVD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRLVD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXWQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWQMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSUBQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLD128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLDMasked128constMerging dst [a] x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLD128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLDMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRLVW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSLLQ128const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLQMasked128constMerging dst [a] x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLQ128const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLQMasked128constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAVD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVDMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVSXBD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBDMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXBQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBQMasked512Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPLZCNTQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPLZCNTQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPLZCNTQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPACKSSDW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPACKSSDWMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKSSDW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPACKSSDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMOVZXWD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWDMasked128Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSRAVQ128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVQMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVQ128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTDMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VMAXPS128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMAXPSMasked128Merging dst x y (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPS128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMAXPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPSHRDQ128 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDQMasked128Merging dst [a] x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDQ128 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDQMasked128Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXUW128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUWMasked128Merging dst x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUW128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPABSB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSBMasked128Merging dst x (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPABSQ128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSQMasked128Merging dst x (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSQ128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSQMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSCALEFPD128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSCALEFPDMasked128Merging dst x y (VPMOVVec64x2ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPD128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSCALEFPDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x2ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VSQRTPS128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSQRTPSMasked128Merging dst x (VPMOVVec32x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPS128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSQRTPSMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPBROADCASTW512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPBROADCASTWMasked512Merging dst x (VPMOVVec16x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPBROADCASTW512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPBROADCASTWMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB128 dst (VPMAXUB128 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUBMasked128Merging dst x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUB128 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	return false
+}
+func rewriteValueAMD64_OpAMD64VPBLENDVB256(v *Value) bool {
+	v_2 := v.Args[2]
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (VPBLENDVB256 dst (VPMOVSXBW512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXBWMasked512Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXBW512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXBWMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDUSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDUSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDUSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMULPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMULPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMULPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPOPCNTB256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTBMasked256Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSUBPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSUBPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSUBPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXUQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPROLD256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAVD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VADDPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VADDPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VADDPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVSXDQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXDQMasked512Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXDQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXDQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVUSWB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVUSWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSWB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVUSWBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAQ256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAQMasked256constMerging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAQ256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAQMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VCVTPS2UDQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VCVTPS2UDQMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTPS2UDQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VCVTPS2UDQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHLDD256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDDMasked256Merging dst [a] x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLVW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRLVQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBUSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBUSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBUSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMINPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMINPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMINPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINSD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VRNDSCALEPS256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRNDSCALEPSMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPS256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRNDSCALEPSMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPROLVQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLVQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMULHW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULHWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULHWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VDIVPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VDIVPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VDIVPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPLZCNTQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPLZCNTQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPLZCNTQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRLVD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVSDW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSDW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPOPCNTD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDUSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDUSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDUSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDUSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSQRTPD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSQRTPDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSQRTPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VREDUCEPS256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VREDUCEPSMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPS256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VREDUCEPSMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVSXWD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSXWDMasked512Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSXWD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSXWDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VGF2P8MULB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VGF2P8MULBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VGF2P8MULB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VGF2P8MULBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLVD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRLVW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRLVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRLVW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRLVWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VREDUCEPD256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VREDUCEPDMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VREDUCEPD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VREDUCEPDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VRNDSCALEPD256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRNDSCALEPDMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRNDSCALEPD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRNDSCALEPDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPRORVD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORVDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHLDW256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDWMasked256Merging dst [a] x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDW256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDWMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VCVTTPS2DQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VCVTTPS2DQMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VCVTTPS2DQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VCVTTPS2DQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSUBPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSUBPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSUBPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSUBPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSQRTPS256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSQRTPSMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSQRTPS256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSQRTPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPACKUSDW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPACKUSDWMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKUSDW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPACKUSDWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMULLD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVWB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVWB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVWBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMADDWD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMADDWDMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDWD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMADDWDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVQD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVQD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVQDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMULHUW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULHUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULHUW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULHUWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMULLQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPROLVD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLVDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLVD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLVDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVUSDW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVUSDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSDW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVUSDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMULLW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMULLWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMULLW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMULLWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPRORD256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAVW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINUD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHUFD256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFDMasked256Merging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLVQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLVQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLVQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVUSQD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVUSQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVUSQD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVUSQDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBUSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBUSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBUSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBUSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VRSQRT14PD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRSQRT14PDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRSQRT14PD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRSQRT14PDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVZXWD512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXWDMasked512Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXWD512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXWDMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPROLQ256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPROLQMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPROLQ256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPROLQMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPAVGB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPAVGBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPAVGBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPRORVQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORVQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORVQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVZXDQ512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXDQMasked512Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXDQ512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXDQMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINUB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLW256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLWMasked256constMerging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLW256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLWMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSCALEFPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSCALEFPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSCALEFPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLQ256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLQMasked256constMerging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLQ256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLQMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPABSQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHUFHW256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFHWMasked256Merging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFHW256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFHWMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMAXPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMAXPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMAXPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXSD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMULPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMULPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMULPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMULPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VDIVPS256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VDIVPSMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VDIVPS256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VDIVPSMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXSQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMINPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMINPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMINPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMINPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHLDQ256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHLDQMasked256Merging dst [a] x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHLDQ256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHLDQMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VSCALEFPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VSCALEFPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VSCALEFPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VSCALEFPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVSWB128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSWBMasked128Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSWB128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSWBMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINSQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINSQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINSQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINSQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPABSD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINUW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHRDW256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDWMasked256Merging dst [a] x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDW256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDWMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVZXBW512 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVZXBWMasked512Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVZXBW512 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVZXBWMasked512Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXUD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUDMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHRDQ256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDQMasked256Merging dst [a] x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDQ256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDQMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMADDUBSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMADDUBSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMADDUBSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMADDUBSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSLLD256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSLLDMasked256constMerging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSLLD256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSLLDMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMINUQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMINUQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMINUQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMINUQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VRCP14PD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VRCP14PDMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VRCP14PD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VRCP14PDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHRDD256 [a] x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHRDDMasked256Merging dst [a] x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHRDD256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHRDDMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPADDQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPADDQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPADDQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPADDQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXUB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPRORQ256 [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPRORQMasked256Merging dst [a] x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPRORQ256 {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPRORQMasked256Merging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VADDPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VADDPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VADDPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VADDPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSHUFB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSHUFBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSHUFB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSHUFBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAD256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRADMasked256constMerging dst [a] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAD256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRADMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAW256const [a] x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAWMasked256constMerging dst [a] x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAW256const {
+			break
+		}
+		a := auxIntToUint8(v_1.AuxInt)
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAWMasked256constMerging)
+		v.AuxInt = uint8ToAuxInt(a)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPABSW256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSWMasked256Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPACKSSDW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPACKSSDWMasked256Merging dst x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPACKSSDW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPACKSSDWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVSQD128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVSQDMasked128Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVSQD128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVSQDMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPLZCNTD256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPLZCNTDMasked256Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPLZCNTD256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPLZCNTDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VMAXPD256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VMAXPDMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VMAXPD256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VMAXPDMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPAVGW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPAVGWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPAVGW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPAVGWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPOPCNTQ256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTQMasked256Merging dst x (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTQ256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBSW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBSWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBSWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMAXUW256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMAXUWMasked256Merging dst x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMAXUW256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMAXUWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPOPCNTW256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPOPCNTWMasked256Merging dst x (VPMOVVec16x16ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPOPCNTW256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPOPCNTWMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec16x16ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSRAVQ256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSRAVQMasked256Merging dst x y (VPMOVVec64x4ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSRAVQ256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSRAVQMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec64x4ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPABSB256 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPABSBMasked256Merging dst x (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPABSB256 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPABSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPMOVDW128 x) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPMOVDWMasked128Merging dst x (VPMOVVec32x8ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPMOVDW128 {
+			break
+		}
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPMOVDWMasked128Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec32x8ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg3(dst, x, v0)
+		return true
+	}
+	// match: (VPBLENDVB256 dst (VPSUBSB256 x y) mask)
+	// cond: v.Block.CPUfeatures.hasFeature(CPUavx512)
+	// result: (VPSUBSBMasked256Merging dst x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+	for {
+		dst := v_0
+		if v_1.Op != OpAMD64VPSUBSB256 {
+			break
+		}
+		y := v_1.Args[1]
+		x := v_1.Args[0]
+		mask := v_2
+		if !(v.Block.CPUfeatures.hasFeature(CPUavx512)) {
+			break
+		}
+		v.reset(OpAMD64VPSUBSBMasked256Merging)
+		v0 := b.NewValue0(v.Pos, OpAMD64VPMOVVec8x32ToM, types.TypeMask)
+		v0.AddArg(mask)
+		v.AddArg4(dst, x, y, v0)
+		return true
+	}
+	return false
+}
 func rewriteValueAMD64_OpAMD64VPBROADCASTB128(v *Value) bool {
 	v_0 := v.Args[0]
 	b := v.Block
author	Junyang Shao <shaojunyang@google.com>	2025-11-05 19:25:00 +0000
committer	Junyang Shao <shaojunyang@google.com>	2025-11-11 13:34:39 -0800
commit	86b4fe31d9b7fe4b249a3a8007290305eaa4f16a (patch)
tree	23f7b2fed426e41c055c39d864835012e602119c /src/cmd/compile
parent	771a1dc216ff02dd23c78ada35a207a363690d11 (diff)
download	go-86b4fe31d9b7fe4b249a3a8007290305eaa4f16a.tar.xz