[dev.simd] cmd/compile: tweak sort order in generator

This CL is created by simdgen CL 684056

Change-Id: Ie4240098bbe701531ab82d5200e92857726f1ba7
Reviewed-on: https://go-review.googlesource.com/c/go/+/684076
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>

1 files changed, 415 insertions, 415 deletions

diff --git a/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules b/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
index bb0476fc20..b21d58b4a4 100644
--- a/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/simdAMD64.rules
@@ -1,807 +1,807 @@
 // Code generated by x/arch/internal/simdgen using 'go run . -xedPath $XED_PATH -o godefs -goroot $GOROOT go.yaml types.yaml categories.yaml'; DO NOT EDIT.
 
+(AbsoluteInt8x16 ...) => (VPABSB128 ...)
+(AbsoluteInt8x32 ...) => (VPABSB256 ...)
+(AbsoluteInt8x64 ...) => (VPABSB512 ...)
+(AbsoluteInt16x8 ...) => (VPABSW128 ...)
 (AbsoluteInt16x16 ...) => (VPABSW256 ...)
 (AbsoluteInt16x32 ...) => (VPABSW512 ...)
-(AbsoluteInt16x8 ...) => (VPABSW128 ...)
-(AbsoluteInt32x16 ...) => (VPABSD512 ...)
 (AbsoluteInt32x4 ...) => (VPABSD128 ...)
 (AbsoluteInt32x8 ...) => (VPABSD256 ...)
+(AbsoluteInt32x16 ...) => (VPABSD512 ...)
 (AbsoluteInt64x2 ...) => (VPABSQ128 ...)
 (AbsoluteInt64x4 ...) => (VPABSQ256 ...)
 (AbsoluteInt64x8 ...) => (VPABSQ512 ...)
-(AbsoluteInt8x16 ...) => (VPABSB128 ...)
-(AbsoluteInt8x32 ...) => (VPABSB256 ...)
-(AbsoluteInt8x64 ...) => (VPABSB512 ...)
-(AddFloat32x16 ...) => (VADDPS512 ...)
 (AddFloat32x4 ...) => (VADDPS128 ...)
 (AddFloat32x8 ...) => (VADDPS256 ...)
+(AddFloat32x16 ...) => (VADDPS512 ...)
 (AddFloat64x2 ...) => (VADDPD128 ...)
 (AddFloat64x4 ...) => (VADDPD256 ...)
 (AddFloat64x8 ...) => (VADDPD512 ...)
+(AddInt8x16 ...) => (VPADDB128 ...)
+(AddInt8x32 ...) => (VPADDB256 ...)
+(AddInt8x64 ...) => (VPADDB512 ...)
+(AddInt16x8 ...) => (VPADDW128 ...)
 (AddInt16x16 ...) => (VPADDW256 ...)
 (AddInt16x32 ...) => (VPADDW512 ...)
-(AddInt16x8 ...) => (VPADDW128 ...)
-(AddInt32x16 ...) => (VPADDD512 ...)
 (AddInt32x4 ...) => (VPADDD128 ...)
 (AddInt32x8 ...) => (VPADDD256 ...)
+(AddInt32x16 ...) => (VPADDD512 ...)
 (AddInt64x2 ...) => (VPADDQ128 ...)
 (AddInt64x4 ...) => (VPADDQ256 ...)
 (AddInt64x8 ...) => (VPADDQ512 ...)
-(AddInt8x16 ...) => (VPADDB128 ...)
-(AddInt8x32 ...) => (VPADDB256 ...)
-(AddInt8x64 ...) => (VPADDB512 ...)
+(AddUint8x16 ...) => (VPADDB128 ...)
+(AddUint8x32 ...) => (VPADDB256 ...)
+(AddUint8x64 ...) => (VPADDB512 ...)
+(AddUint16x8 ...) => (VPADDW128 ...)
 (AddUint16x16 ...) => (VPADDW256 ...)
 (AddUint16x32 ...) => (VPADDW512 ...)
-(AddUint16x8 ...) => (VPADDW128 ...)
-(AddUint32x16 ...) => (VPADDD512 ...)
 (AddUint32x4 ...) => (VPADDD128 ...)
 (AddUint32x8 ...) => (VPADDD256 ...)
+(AddUint32x16 ...) => (VPADDD512 ...)
 (AddUint64x2 ...) => (VPADDQ128 ...)
 (AddUint64x4 ...) => (VPADDQ256 ...)
 (AddUint64x8 ...) => (VPADDQ512 ...)
-(AddUint8x16 ...) => (VPADDB128 ...)
-(AddUint8x32 ...) => (VPADDB256 ...)
-(AddUint8x64 ...) => (VPADDB512 ...)
 (AddSubFloat32x4 ...) => (VADDSUBPS128 ...)
 (AddSubFloat32x8 ...) => (VADDSUBPS256 ...)
 (AddSubFloat64x2 ...) => (VADDSUBPD128 ...)
 (AddSubFloat64x4 ...) => (VADDSUBPD256 ...)
-(AndFloat32x16 ...) => (VANDPS512 ...)
 (AndFloat32x4 ...) => (VANDPS128 ...)
 (AndFloat32x8 ...) => (VANDPS256 ...)
+(AndFloat32x16 ...) => (VANDPS512 ...)
 (AndFloat64x2 ...) => (VANDPD128 ...)
 (AndFloat64x4 ...) => (VANDPD256 ...)
 (AndFloat64x8 ...) => (VANDPD512 ...)
-(AndInt16x16 ...) => (VPAND256 ...)
+(AndInt8x16 ...) => (VPAND128 ...)
+(AndInt8x32 ...) => (VPAND256 ...)
 (AndInt16x8 ...) => (VPAND128 ...)
-(AndInt32x16 ...) => (VPANDD512 ...)
+(AndInt16x16 ...) => (VPAND256 ...)
 (AndInt32x4 ...) => (VPAND128 ...)
 (AndInt32x8 ...) => (VPAND256 ...)
+(AndInt32x16 ...) => (VPANDD512 ...)
 (AndInt64x2 ...) => (VPAND128 ...)
 (AndInt64x4 ...) => (VPAND256 ...)
 (AndInt64x8 ...) => (VPANDQ512 ...)
-(AndInt8x16 ...) => (VPAND128 ...)
-(AndInt8x32 ...) => (VPAND256 ...)
-(AndUint16x16 ...) => (VPAND256 ...)
+(AndUint8x16 ...) => (VPAND128 ...)
+(AndUint8x32 ...) => (VPAND256 ...)
 (AndUint16x8 ...) => (VPAND128 ...)
-(AndUint32x16 ...) => (VPANDD512 ...)
+(AndUint16x16 ...) => (VPAND256 ...)
 (AndUint32x4 ...) => (VPAND128 ...)
 (AndUint32x8 ...) => (VPAND256 ...)
+(AndUint32x16 ...) => (VPANDD512 ...)
 (AndUint64x2 ...) => (VPAND128 ...)
 (AndUint64x4 ...) => (VPAND256 ...)
 (AndUint64x8 ...) => (VPANDQ512 ...)
-(AndUint8x16 ...) => (VPAND128 ...)
-(AndUint8x32 ...) => (VPAND256 ...)
-(AndNotFloat32x16 ...) => (VANDNPS512 ...)
 (AndNotFloat32x4 ...) => (VANDNPS128 ...)
 (AndNotFloat32x8 ...) => (VANDNPS256 ...)
+(AndNotFloat32x16 ...) => (VANDNPS512 ...)
 (AndNotFloat64x2 ...) => (VANDNPD128 ...)
 (AndNotFloat64x4 ...) => (VANDNPD256 ...)
 (AndNotFloat64x8 ...) => (VANDNPD512 ...)
-(AndNotInt16x16 ...) => (VPANDN256 ...)
+(AndNotInt8x16 ...) => (VPANDN128 ...)
+(AndNotInt8x32 ...) => (VPANDN256 ...)
 (AndNotInt16x8 ...) => (VPANDN128 ...)
-(AndNotInt32x16 ...) => (VPANDND512 ...)
+(AndNotInt16x16 ...) => (VPANDN256 ...)
 (AndNotInt32x4 ...) => (VPANDN128 ...)
 (AndNotInt32x8 ...) => (VPANDN256 ...)
+(AndNotInt32x16 ...) => (VPANDND512 ...)
 (AndNotInt64x2 ...) => (VPANDN128 ...)
 (AndNotInt64x4 ...) => (VPANDN256 ...)
 (AndNotInt64x8 ...) => (VPANDNQ512 ...)
-(AndNotInt8x16 ...) => (VPANDN128 ...)
-(AndNotInt8x32 ...) => (VPANDN256 ...)
-(AndNotUint16x16 ...) => (VPANDN256 ...)
+(AndNotUint8x16 ...) => (VPANDN128 ...)
+(AndNotUint8x32 ...) => (VPANDN256 ...)
 (AndNotUint16x8 ...) => (VPANDN128 ...)
-(AndNotUint32x16 ...) => (VPANDND512 ...)
+(AndNotUint16x16 ...) => (VPANDN256 ...)
 (AndNotUint32x4 ...) => (VPANDN128 ...)
 (AndNotUint32x8 ...) => (VPANDN256 ...)
+(AndNotUint32x16 ...) => (VPANDND512 ...)
 (AndNotUint64x2 ...) => (VPANDN128 ...)
 (AndNotUint64x4 ...) => (VPANDN256 ...)
 (AndNotUint64x8 ...) => (VPANDNQ512 ...)
-(AndNotUint8x16 ...) => (VPANDN128 ...)
-(AndNotUint8x32 ...) => (VPANDN256 ...)
-(ApproximateReciprocalFloat32x16 ...) => (VRCP14PS512 ...)
 (ApproximateReciprocalFloat32x4 ...) => (VRCP14PS128 ...)
 (ApproximateReciprocalFloat32x8 ...) => (VRCP14PS256 ...)
+(ApproximateReciprocalFloat32x16 ...) => (VRCP14PS512 ...)
 (ApproximateReciprocalFloat64x2 ...) => (VRCP14PD128 ...)
 (ApproximateReciprocalFloat64x4 ...) => (VRCP14PD256 ...)
 (ApproximateReciprocalFloat64x8 ...) => (VRCP14PD512 ...)
-(ApproximateReciprocalOfSqrtFloat32x16 ...) => (VRSQRT14PS512 ...)
 (ApproximateReciprocalOfSqrtFloat32x4 ...) => (VRSQRTPS128 ...)
 (ApproximateReciprocalOfSqrtFloat32x8 ...) => (VRSQRTPS256 ...)
+(ApproximateReciprocalOfSqrtFloat32x16 ...) => (VRSQRT14PS512 ...)
 (ApproximateReciprocalOfSqrtFloat64x2 ...) => (VRSQRT14PD128 ...)
 (ApproximateReciprocalOfSqrtFloat64x4 ...) => (VRSQRT14PD256 ...)
 (ApproximateReciprocalOfSqrtFloat64x8 ...) => (VRSQRT14PD512 ...)
-(AverageUint16x16 ...) => (VPAVGW256 ...)
-(AverageUint16x32 ...) => (VPAVGW512 ...)
-(AverageUint16x8 ...) => (VPAVGW128 ...)
 (AverageUint8x16 ...) => (VPAVGB128 ...)
 (AverageUint8x32 ...) => (VPAVGB256 ...)
 (AverageUint8x64 ...) => (VPAVGB512 ...)
+(AverageUint16x8 ...) => (VPAVGW128 ...)
+(AverageUint16x16 ...) => (VPAVGW256 ...)
+(AverageUint16x32 ...) => (VPAVGW512 ...)
 (CeilFloat32x4 x) => (VROUNDPS128 [2] x)
 (CeilFloat32x8 x) => (VROUNDPS256 [2] x)
 (CeilFloat64x2 x) => (VROUNDPD128 [2] x)
 (CeilFloat64x4 x) => (VROUNDPD256 [2] x)
-(CeilSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+10] x)
 (CeilSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+10] x)
 (CeilSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+10] x)
+(CeilSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+10] x)
 (CeilSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+10] x)
 (CeilSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+10] x)
 (CeilSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+10] x)
-(CeilWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+2] x)
 (CeilWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+2] x)
 (CeilWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+2] x)
+(CeilWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+2] x)
 (CeilWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+2] x)
 (CeilWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+2] x)
 (CeilWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+2] x)
-(DiffWithCeilSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+10] x)
 (DiffWithCeilSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+10] x)
 (DiffWithCeilSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+10] x)
+(DiffWithCeilSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+10] x)
 (DiffWithCeilSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+10] x)
 (DiffWithCeilSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+10] x)
 (DiffWithCeilSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+10] x)
-(DiffWithCeilWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+2] x)
 (DiffWithCeilWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+2] x)
 (DiffWithCeilWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+2] x)
+(DiffWithCeilWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+2] x)
 (DiffWithCeilWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+2] x)
 (DiffWithCeilWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+2] x)
 (DiffWithCeilWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+2] x)
-(DiffWithFloorSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+9] x)
 (DiffWithFloorSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+9] x)
 (DiffWithFloorSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+9] x)
+(DiffWithFloorSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+9] x)
 (DiffWithFloorSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+9] x)
 (DiffWithFloorSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+9] x)
 (DiffWithFloorSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+9] x)
-(DiffWithFloorWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+1] x)
 (DiffWithFloorWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+1] x)
 (DiffWithFloorWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+1] x)
+(DiffWithFloorWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+1] x)
 (DiffWithFloorWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+1] x)
 (DiffWithFloorWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+1] x)
 (DiffWithFloorWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+1] x)
-(DiffWithRoundSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+8] x)
 (DiffWithRoundSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+8] x)
 (DiffWithRoundSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+8] x)
+(DiffWithRoundSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+8] x)
 (DiffWithRoundSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+8] x)
 (DiffWithRoundSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+8] x)
 (DiffWithRoundSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+8] x)
-(DiffWithRoundWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+0] x)
 (DiffWithRoundWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+0] x)
 (DiffWithRoundWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+0] x)
+(DiffWithRoundWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+0] x)
 (DiffWithRoundWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+0] x)
 (DiffWithRoundWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+0] x)
 (DiffWithRoundWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+0] x)
-(DiffWithTruncSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+11] x)
 (DiffWithTruncSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+11] x)
 (DiffWithTruncSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+11] x)
+(DiffWithTruncSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+11] x)
 (DiffWithTruncSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+11] x)
 (DiffWithTruncSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+11] x)
 (DiffWithTruncSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+11] x)
-(DiffWithTruncWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+3] x)
 (DiffWithTruncWithPrecisionFloat32x4 [a] x) => (VREDUCEPS128 [a+3] x)
 (DiffWithTruncWithPrecisionFloat32x8 [a] x) => (VREDUCEPS256 [a+3] x)
+(DiffWithTruncWithPrecisionFloat32x16 [a] x) => (VREDUCEPS512 [a+3] x)
 (DiffWithTruncWithPrecisionFloat64x2 [a] x) => (VREDUCEPD128 [a+3] x)
 (DiffWithTruncWithPrecisionFloat64x4 [a] x) => (VREDUCEPD256 [a+3] x)
 (DiffWithTruncWithPrecisionFloat64x8 [a] x) => (VREDUCEPD512 [a+3] x)
-(DivFloat32x16 ...) => (VDIVPS512 ...)
 (DivFloat32x4 ...) => (VDIVPS128 ...)
 (DivFloat32x8 ...) => (VDIVPS256 ...)
+(DivFloat32x16 ...) => (VDIVPS512 ...)
 (DivFloat64x2 ...) => (VDIVPD128 ...)
 (DivFloat64x4 ...) => (VDIVPD256 ...)
 (DivFloat64x8 ...) => (VDIVPD512 ...)
 (DotProdBroadcastFloat64x2 x y) => (VDPPD128 [127] x y)
-(EqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [0] x y))
 (EqualFloat32x4 x y) => (VCMPPS128 [0] x y)
 (EqualFloat32x8 x y) => (VCMPPS256 [0] x y)
+(EqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [0] x y))
 (EqualFloat64x2 x y) => (VCMPPD128 [0] x y)
 (EqualFloat64x4 x y) => (VCMPPD256 [0] x y)
 (EqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [0] x y))
+(EqualInt8x16 ...) => (VPCMPEQB128 ...)
+(EqualInt8x32 ...) => (VPCMPEQB256 ...)
+(EqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [0] x y))
+(EqualInt16x8 ...) => (VPCMPEQW128 ...)
 (EqualInt16x16 ...) => (VPCMPEQW256 ...)
 (EqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [0] x y))
-(EqualInt16x8 ...) => (VPCMPEQW128 ...)
-(EqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [0] x y))
 (EqualInt32x4 ...) => (VPCMPEQD128 ...)
 (EqualInt32x8 ...) => (VPCMPEQD256 ...)
+(EqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [0] x y))
 (EqualInt64x2 ...) => (VPCMPEQQ128 ...)
 (EqualInt64x4 ...) => (VPCMPEQQ256 ...)
 (EqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [0] x y))
-(EqualInt8x16 ...) => (VPCMPEQB128 ...)
-(EqualInt8x32 ...) => (VPCMPEQB256 ...)
-(EqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [0] x y))
+(EqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [0] x y))
+(EqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [0] x y))
+(EqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [0] x y))
+(EqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [0] x y))
 (EqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [0] x y))
 (EqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [0] x y))
-(EqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [0] x y))
-(EqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [0] x y))
 (EqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [0] x y))
 (EqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [0] x y))
+(EqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [0] x y))
 (EqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [0] x y))
 (EqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [0] x y))
 (EqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [0] x y))
-(EqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [0] x y))
-(EqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [0] x y))
-(EqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [0] x y))
 (FloorFloat32x4 x) => (VROUNDPS128 [1] x)
 (FloorFloat32x8 x) => (VROUNDPS256 [1] x)
 (FloorFloat64x2 x) => (VROUNDPD128 [1] x)
 (FloorFloat64x4 x) => (VROUNDPD256 [1] x)
-(FloorSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+9] x)
 (FloorSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+9] x)
 (FloorSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+9] x)
+(FloorSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+9] x)
 (FloorSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+9] x)
 (FloorSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+9] x)
 (FloorSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+9] x)
-(FloorWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+1] x)
 (FloorWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+1] x)
 (FloorWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+1] x)
+(FloorWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+1] x)
 (FloorWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+1] x)
 (FloorWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+1] x)
 (FloorWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+1] x)
-(FusedMultiplyAddFloat32x16 ...) => (VFMADD213PS512 ...)
 (FusedMultiplyAddFloat32x4 ...) => (VFMADD213PS128 ...)
 (FusedMultiplyAddFloat32x8 ...) => (VFMADD213PS256 ...)
+(FusedMultiplyAddFloat32x16 ...) => (VFMADD213PS512 ...)
 (FusedMultiplyAddFloat64x2 ...) => (VFMADD213PD128 ...)
 (FusedMultiplyAddFloat64x4 ...) => (VFMADD213PD256 ...)
 (FusedMultiplyAddFloat64x8 ...) => (VFMADD213PD512 ...)
-(FusedMultiplyAddSubFloat32x16 ...) => (VFMADDSUB213PS512 ...)
 (FusedMultiplyAddSubFloat32x4 ...) => (VFMADDSUB213PS128 ...)
 (FusedMultiplyAddSubFloat32x8 ...) => (VFMADDSUB213PS256 ...)
+(FusedMultiplyAddSubFloat32x16 ...) => (VFMADDSUB213PS512 ...)
 (FusedMultiplyAddSubFloat64x2 ...) => (VFMADDSUB213PD128 ...)
 (FusedMultiplyAddSubFloat64x4 ...) => (VFMADDSUB213PD256 ...)
 (FusedMultiplyAddSubFloat64x8 ...) => (VFMADDSUB213PD512 ...)
-(FusedMultiplySubAddFloat32x16 ...) => (VFMSUBADD213PS512 ...)
 (FusedMultiplySubAddFloat32x4 ...) => (VFMSUBADD213PS128 ...)
 (FusedMultiplySubAddFloat32x8 ...) => (VFMSUBADD213PS256 ...)
+(FusedMultiplySubAddFloat32x16 ...) => (VFMSUBADD213PS512 ...)
 (FusedMultiplySubAddFloat64x2 ...) => (VFMSUBADD213PD128 ...)
 (FusedMultiplySubAddFloat64x4 ...) => (VFMSUBADD213PD256 ...)
 (FusedMultiplySubAddFloat64x8 ...) => (VFMSUBADD213PD512 ...)
+(GetElemInt8x16 [a] x) => (VPEXTRB128 [a] x)
 (GetElemInt16x8 [a] x) => (VPEXTRW128 [a] x)
 (GetElemInt32x4 [a] x) => (VPEXTRD128 [a] x)
 (GetElemInt64x2 [a] x) => (VPEXTRQ128 [a] x)
-(GetElemInt8x16 [a] x) => (VPEXTRB128 [a] x)
+(GetElemUint8x16 [a] x) => (VPEXTRB128 [a] x)
 (GetElemUint16x8 [a] x) => (VPEXTRW128 [a] x)
 (GetElemUint32x4 [a] x) => (VPEXTRD128 [a] x)
 (GetElemUint64x2 [a] x) => (VPEXTRQ128 [a] x)
-(GetElemUint8x16 [a] x) => (VPEXTRB128 [a] x)
-(GreaterFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [6] x y))
 (GreaterFloat32x4 x y) => (VCMPPS128 [6] x y)
 (GreaterFloat32x8 x y) => (VCMPPS256 [6] x y)
+(GreaterFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [6] x y))
 (GreaterFloat64x2 x y) => (VCMPPD128 [6] x y)
 (GreaterFloat64x4 x y) => (VCMPPD256 [6] x y)
 (GreaterFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [6] x y))
+(GreaterInt8x16 ...) => (VPCMPGTB128 ...)
+(GreaterInt8x32 ...) => (VPCMPGTB256 ...)
+(GreaterInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [6] x y))
+(GreaterInt16x8 ...) => (VPCMPGTW128 ...)
 (GreaterInt16x16 ...) => (VPCMPGTW256 ...)
 (GreaterInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [6] x y))
-(GreaterInt16x8 ...) => (VPCMPGTW128 ...)
-(GreaterInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [6] x y))
 (GreaterInt32x4 ...) => (VPCMPGTD128 ...)
 (GreaterInt32x8 ...) => (VPCMPGTD256 ...)
+(GreaterInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [6] x y))
 (GreaterInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [6] x y))
 (GreaterInt64x4 ...) => (VPCMPGTQ256 ...)
 (GreaterInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [6] x y))
-(GreaterInt8x16 ...) => (VPCMPGTB128 ...)
-(GreaterInt8x32 ...) => (VPCMPGTB256 ...)
-(GreaterInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [6] x y))
+(GreaterUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [6] x y))
+(GreaterUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [6] x y))
+(GreaterUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [6] x y))
+(GreaterUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [6] x y))
 (GreaterUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [6] x y))
 (GreaterUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [6] x y))
-(GreaterUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [6] x y))
-(GreaterUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [6] x y))
 (GreaterUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [6] x y))
 (GreaterUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [6] x y))
+(GreaterUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [6] x y))
 (GreaterUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [6] x y))
 (GreaterUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [6] x y))
 (GreaterUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [6] x y))
-(GreaterUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [6] x y))
-(GreaterUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [6] x y))
-(GreaterUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [6] x y))
-(GreaterEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [5] x y))
 (GreaterEqualFloat32x4 x y) => (VCMPPS128 [5] x y)
 (GreaterEqualFloat32x8 x y) => (VCMPPS256 [5] x y)
+(GreaterEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [5] x y))
 (GreaterEqualFloat64x2 x y) => (VCMPPD128 [5] x y)
 (GreaterEqualFloat64x4 x y) => (VCMPPD256 [5] x y)
 (GreaterEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [5] x y))
+(GreaterEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [5] x y))
+(GreaterEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [5] x y))
+(GreaterEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [5] x y))
+(GreaterEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [5] x y))
 (GreaterEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [5] x y))
 (GreaterEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [5] x y))
-(GreaterEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [5] x y))
-(GreaterEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [5] x y))
 (GreaterEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [5] x y))
 (GreaterEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [5] x y))
+(GreaterEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [5] x y))
 (GreaterEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [5] x y))
 (GreaterEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [5] x y))
 (GreaterEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [5] x y))
-(GreaterEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [5] x y))
-(GreaterEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [5] x y))
-(GreaterEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [5] x y))
+(GreaterEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [5] x y))
+(GreaterEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [5] x y))
+(GreaterEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [5] x y))
+(GreaterEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [5] x y))
 (GreaterEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [5] x y))
 (GreaterEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [5] x y))
-(GreaterEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [5] x y))
-(GreaterEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [5] x y))
 (GreaterEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [5] x y))
 (GreaterEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [5] x y))
+(GreaterEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [5] x y))
 (GreaterEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [5] x y))
 (GreaterEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [5] x y))
 (GreaterEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [5] x y))
-(GreaterEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [5] x y))
-(GreaterEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [5] x y))
-(GreaterEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [5] x y))
-(IsNanFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [3] x y))
 (IsNanFloat32x4 x y) => (VCMPPS128 [3] x y)
 (IsNanFloat32x8 x y) => (VCMPPS256 [3] x y)
+(IsNanFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [3] x y))
 (IsNanFloat64x2 x y) => (VCMPPD128 [3] x y)
 (IsNanFloat64x4 x y) => (VCMPPD256 [3] x y)
 (IsNanFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [3] x y))
-(LessFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [1] x y))
 (LessFloat32x4 x y) => (VCMPPS128 [1] x y)
 (LessFloat32x8 x y) => (VCMPPS256 [1] x y)
+(LessFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [1] x y))
 (LessFloat64x2 x y) => (VCMPPD128 [1] x y)
 (LessFloat64x4 x y) => (VCMPPD256 [1] x y)
 (LessFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [1] x y))
+(LessInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [1] x y))
+(LessInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [1] x y))
+(LessInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [1] x y))
+(LessInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [1] x y))
 (LessInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [1] x y))
 (LessInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [1] x y))
-(LessInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [1] x y))
-(LessInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [1] x y))
 (LessInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [1] x y))
 (LessInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [1] x y))
+(LessInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [1] x y))
 (LessInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [1] x y))
 (LessInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [1] x y))
 (LessInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [1] x y))
-(LessInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [1] x y))
-(LessInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [1] x y))
-(LessInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [1] x y))
+(LessUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [1] x y))
+(LessUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [1] x y))
+(LessUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [1] x y))
+(LessUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [1] x y))
 (LessUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [1] x y))
 (LessUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [1] x y))
-(LessUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [1] x y))
-(LessUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [1] x y))
 (LessUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [1] x y))
 (LessUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [1] x y))
+(LessUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [1] x y))
 (LessUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [1] x y))
 (LessUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [1] x y))
 (LessUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [1] x y))
-(LessUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [1] x y))
-(LessUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [1] x y))
-(LessUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [1] x y))
-(LessEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [2] x y))
 (LessEqualFloat32x4 x y) => (VCMPPS128 [2] x y)
 (LessEqualFloat32x8 x y) => (VCMPPS256 [2] x y)
+(LessEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [2] x y))
 (LessEqualFloat64x2 x y) => (VCMPPD128 [2] x y)
 (LessEqualFloat64x4 x y) => (VCMPPD256 [2] x y)
 (LessEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [2] x y))
+(LessEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [2] x y))
+(LessEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [2] x y))
+(LessEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [2] x y))
+(LessEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [2] x y))
 (LessEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [2] x y))
 (LessEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [2] x y))
-(LessEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [2] x y))
-(LessEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [2] x y))
 (LessEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [2] x y))
 (LessEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [2] x y))
+(LessEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [2] x y))
 (LessEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [2] x y))
 (LessEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [2] x y))
 (LessEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [2] x y))
-(LessEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [2] x y))
-(LessEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [2] x y))
-(LessEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [2] x y))
+(LessEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [2] x y))
+(LessEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [2] x y))
+(LessEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [2] x y))
+(LessEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [2] x y))
 (LessEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [2] x y))
 (LessEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [2] x y))
-(LessEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [2] x y))
-(LessEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [2] x y))
 (LessEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [2] x y))
 (LessEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [2] x y))
+(LessEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [2] x y))
 (LessEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [2] x y))
 (LessEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [2] x y))
 (LessEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [2] x y))
-(LessEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [2] x y))
-(LessEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [2] x y))
-(LessEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [2] x y))
+(MaskedAbsoluteInt8x16 x mask) => (VPABSBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedAbsoluteInt8x32 x mask) => (VPABSBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedAbsoluteInt8x64 x mask) => (VPABSBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedAbsoluteInt16x8 x mask) => (VPABSWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt16x16 x mask) => (VPABSWMasked256 x (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt16x32 x mask) => (VPABSWMasked512 x (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedAbsoluteInt16x8 x mask) => (VPABSWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedAbsoluteInt32x16 x mask) => (VPABSDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt32x4 x mask) => (VPABSDMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt32x8 x mask) => (VPABSDMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAbsoluteInt32x16 x mask) => (VPABSDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt64x2 x mask) => (VPABSQMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt64x4 x mask) => (VPABSQMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAbsoluteInt64x8 x mask) => (VPABSQMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAbsoluteInt8x16 x mask) => (VPABSBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedAbsoluteInt8x32 x mask) => (VPABSBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedAbsoluteInt8x64 x mask) => (VPABSBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedAddFloat32x16 x y mask) => (VADDPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddFloat32x4 x y mask) => (VADDPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAddFloat32x8 x y mask) => (VADDPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAddFloat32x16 x y mask) => (VADDPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddFloat64x2 x y mask) => (VADDPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAddFloat64x4 x y mask) => (VADDPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAddFloat64x8 x y mask) => (VADDPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedAddInt8x16 x y mask) => (VPADDBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedAddInt8x32 x y mask) => (VPADDBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedAddInt8x64 x y mask) => (VPADDBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedAddInt16x8 x y mask) => (VPADDWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedAddInt16x16 x y mask) => (VPADDWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedAddInt16x32 x y mask) => (VPADDWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedAddInt16x8 x y mask) => (VPADDWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedAddInt32x16 x y mask) => (VPADDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddInt32x4 x y mask) => (VPADDDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAddInt32x8 x y mask) => (VPADDDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAddInt32x16 x y mask) => (VPADDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddInt64x2 x y mask) => (VPADDQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAddInt64x4 x y mask) => (VPADDQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAddInt64x8 x y mask) => (VPADDQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAddInt8x16 x y mask) => (VPADDBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedAddInt8x32 x y mask) => (VPADDBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedAddInt8x64 x y mask) => (VPADDBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedAddUint8x16 x y mask) => (VPADDBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedAddUint8x32 x y mask) => (VPADDBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedAddUint8x64 x y mask) => (VPADDBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedAddUint16x8 x y mask) => (VPADDWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedAddUint16x16 x y mask) => (VPADDWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedAddUint16x32 x y mask) => (VPADDWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedAddUint16x8 x y mask) => (VPADDWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedAddUint32x16 x y mask) => (VPADDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddUint32x4 x y mask) => (VPADDDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAddUint32x8 x y mask) => (VPADDDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAddUint32x16 x y mask) => (VPADDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAddUint64x2 x y mask) => (VPADDQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAddUint64x4 x y mask) => (VPADDQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAddUint64x8 x y mask) => (VPADDQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAddUint8x16 x y mask) => (VPADDBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedAddUint8x32 x y mask) => (VPADDBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedAddUint8x64 x y mask) => (VPADDBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedAndFloat32x16 x y mask) => (VANDPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndFloat32x4 x y mask) => (VANDPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndFloat32x8 x y mask) => (VANDPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndFloat32x16 x y mask) => (VANDPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndFloat64x2 x y mask) => (VANDPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndFloat64x4 x y mask) => (VANDPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndFloat64x8 x y mask) => (VANDPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAndInt32x16 x y mask) => (VPANDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndInt32x4 x y mask) => (VPANDDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndInt32x8 x y mask) => (VPANDDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndInt32x16 x y mask) => (VPANDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndInt64x2 x y mask) => (VPANDQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndInt64x4 x y mask) => (VPANDQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndInt64x8 x y mask) => (VPANDQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAndUint32x16 x y mask) => (VPANDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndUint32x4 x y mask) => (VPANDDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndUint32x8 x y mask) => (VPANDDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndUint32x16 x y mask) => (VPANDDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndUint64x2 x y mask) => (VPANDQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndUint64x4 x y mask) => (VPANDQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndUint64x8 x y mask) => (VPANDQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAndNotFloat32x16 x y mask) => (VANDNPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotFloat32x4 x y mask) => (VANDNPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndNotFloat32x8 x y mask) => (VANDNPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndNotFloat32x16 x y mask) => (VANDNPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotFloat64x2 x y mask) => (VANDNPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndNotFloat64x4 x y mask) => (VANDNPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndNotFloat64x8 x y mask) => (VANDNPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAndNotInt32x16 x y mask) => (VPANDNDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotInt32x4 x y mask) => (VPANDNDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndNotInt32x8 x y mask) => (VPANDNDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndNotInt32x16 x y mask) => (VPANDNDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotInt64x2 x y mask) => (VPANDNQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndNotInt64x4 x y mask) => (VPANDNQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndNotInt64x8 x y mask) => (VPANDNQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAndNotUint32x16 x y mask) => (VPANDNDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotUint32x4 x y mask) => (VPANDNDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedAndNotUint32x8 x y mask) => (VPANDNDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedAndNotUint32x16 x y mask) => (VPANDNDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedAndNotUint64x2 x y mask) => (VPANDNQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedAndNotUint64x4 x y mask) => (VPANDNQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedAndNotUint64x8 x y mask) => (VPANDNQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedApproximateReciprocalFloat32x16 x mask) => (VRCP14PSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalFloat32x4 x mask) => (VRCP14PSMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalFloat32x8 x mask) => (VRCP14PSMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedApproximateReciprocalFloat32x16 x mask) => (VRCP14PSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalFloat64x2 x mask) => (VRCP14PDMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalFloat64x4 x mask) => (VRCP14PDMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalFloat64x8 x mask) => (VRCP14PDMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedApproximateReciprocalOfSqrtFloat32x16 x mask) => (VRSQRT14PSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalOfSqrtFloat32x4 x mask) => (VRSQRT14PSMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalOfSqrtFloat32x8 x mask) => (VRSQRT14PSMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedApproximateReciprocalOfSqrtFloat32x16 x mask) => (VRSQRT14PSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalOfSqrtFloat64x2 x mask) => (VRSQRT14PDMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalOfSqrtFloat64x4 x mask) => (VRSQRT14PDMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedApproximateReciprocalOfSqrtFloat64x8 x mask) => (VRSQRT14PDMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedAverageUint16x16 x y mask) => (VPAVGWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
-(MaskedAverageUint16x32 x y mask) => (VPAVGWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedAverageUint16x8 x y mask) => (VPAVGWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedAverageUint8x16 x y mask) => (VPAVGBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
 (MaskedAverageUint8x32 x y mask) => (VPAVGBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
 (MaskedAverageUint8x64 x y mask) => (VPAVGBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedCeilSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+10] x (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedAverageUint16x8 x y mask) => (VPAVGWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedAverageUint16x16 x y mask) => (VPAVGWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(MaskedAverageUint16x32 x y mask) => (VPAVGWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (MaskedCeilSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+10] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedCeilSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+10] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedCeilSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+10] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedCeilSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+10] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedCeilSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+10] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedCeilSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+10] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedCeilWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+2] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedCeilWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+2] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedCeilWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+2] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedCeilWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+2] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedCeilWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+2] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedCeilWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+2] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedCeilWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+2] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+10] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+10] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+10] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+10] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+10] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+10] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+10] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithCeilWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+2] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+2] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+2] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithCeilWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+2] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+2] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+2] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithCeilWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+2] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+9] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+9] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+9] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+9] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+9] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+9] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+9] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithFloorWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+1] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+1] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+1] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithFloorWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+1] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+1] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+1] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithFloorWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+1] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+8] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+8] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+8] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+8] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+8] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+8] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+8] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithRoundWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+0] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+0] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+0] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithRoundWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+0] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+0] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+0] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithRoundWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+0] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+11] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+11] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+11] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+11] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+11] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+11] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+11] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDiffWithTruncWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+3] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncWithPrecisionFloat32x4 [a] x mask) => (VREDUCEPSMasked128 [a+3] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncWithPrecisionFloat32x8 [a] x mask) => (VREDUCEPSMasked256 [a+3] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDiffWithTruncWithPrecisionFloat32x16 [a] x mask) => (VREDUCEPSMasked512 [a+3] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncWithPrecisionFloat64x2 [a] x mask) => (VREDUCEPDMasked128 [a+3] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncWithPrecisionFloat64x4 [a] x mask) => (VREDUCEPDMasked256 [a+3] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDiffWithTruncWithPrecisionFloat64x8 [a] x mask) => (VREDUCEPDMasked512 [a+3] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedDivFloat32x16 x y mask) => (VDIVPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDivFloat32x4 x y mask) => (VDIVPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedDivFloat32x8 x y mask) => (VDIVPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedDivFloat32x16 x y mask) => (VDIVPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedDivFloat64x2 x y mask) => (VDIVPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedDivFloat64x4 x y mask) => (VDIVPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedDivFloat64x8 x y mask) => (VDIVPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [0] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedEqualFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [0] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [0] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedEqualFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [0] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedEqualFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [0] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [0] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [0] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [0] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [0] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedEqualInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [0] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedEqualInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [0] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [0] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [0] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedEqualInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [0] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [0] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedEqualInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [0] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedEqualInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [0] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [0] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [0] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [0] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [0] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [0] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [0] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [0] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedEqualUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [0] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedEqualUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [0] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [0] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [0] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedEqualUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [0] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [0] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedEqualUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [0] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedEqualUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [0] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedEqualUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [0] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [0] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [0] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [0] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedFloorSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+9] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFloorSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+9] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedFloorSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+9] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedFloorSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+9] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFloorSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+9] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedFloorSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+9] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedFloorSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+9] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedFloorWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+1] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFloorWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+1] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedFloorWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+1] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedFloorWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+1] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFloorWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+1] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedFloorWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+1] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedFloorWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+1] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedFusedMultiplyAddFloat32x16 x y z mask) => (VFMADD213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddFloat32x4 x y z mask) => (VFMADD213PSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddFloat32x8 x y z mask) => (VFMADD213PSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedFusedMultiplyAddFloat32x16 x y z mask) => (VFMADD213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddFloat64x2 x y z mask) => (VFMADD213PDMasked128 x y z (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddFloat64x4 x y z mask) => (VFMADD213PDMasked256 x y z (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddFloat64x8 x y z mask) => (VFMADD213PDMasked512 x y z (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedFusedMultiplyAddSubFloat32x16 x y z mask) => (VFMADDSUB213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddSubFloat32x4 x y z mask) => (VFMADDSUB213PSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddSubFloat32x8 x y z mask) => (VFMADDSUB213PSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedFusedMultiplyAddSubFloat32x16 x y z mask) => (VFMADDSUB213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddSubFloat64x2 x y z mask) => (VFMADDSUB213PDMasked128 x y z (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddSubFloat64x4 x y z mask) => (VFMADDSUB213PDMasked256 x y z (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplyAddSubFloat64x8 x y z mask) => (VFMADDSUB213PDMasked512 x y z (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedFusedMultiplySubAddFloat32x16 x y z mask) => (VFMSUBADD213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplySubAddFloat32x4 x y z mask) => (VFMSUBADD213PSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplySubAddFloat32x8 x y z mask) => (VFMSUBADD213PSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedFusedMultiplySubAddFloat32x16 x y z mask) => (VFMSUBADD213PSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedFusedMultiplySubAddFloat64x2 x y z mask) => (VFMSUBADD213PDMasked128 x y z (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedFusedMultiplySubAddFloat64x4 x y z mask) => (VFMSUBADD213PDMasked256 x y z (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedFusedMultiplySubAddFloat64x8 x y z mask) => (VFMSUBADD213PDMasked512 x y z (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedGreaterFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [6] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [6] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [6] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [6] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [6] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedGreaterInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [6] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedGreaterInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [6] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedGreaterInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [6] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [6] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedGreaterInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [6] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedGreaterInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [6] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedGreaterInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [6] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedGreaterInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [6] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [6] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [6] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [6] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [6] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedGreaterInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [6] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedGreaterInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [6] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedGreaterInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [6] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [6] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedGreaterUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [6] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedGreaterUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [6] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [6] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedGreaterUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [6] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedGreaterUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [6] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedGreaterUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [6] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedGreaterUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [6] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [6] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [6] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [6] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [6] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [6] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedGreaterUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [6] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedGreaterUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [6] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedGreaterUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [6] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [5] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [5] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [5] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [5] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [5] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [5] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [5] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [5] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [5] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [5] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [5] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [5] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [5] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [5] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [5] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [5] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [5] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [5] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [5] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [5] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [5] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [5] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [5] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [5] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [5] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [5] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [5] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [5] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [5] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedGreaterEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [5] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [5] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [5] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedGreaterEqualUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [5] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [5] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [5] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedGreaterEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [5] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedIsNanFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [3] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedIsNanFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [3] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedIsNanFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [3] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedIsNanFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [3] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedIsNanFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [3] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedIsNanFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [3] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedIsNanFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [3] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedLessFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [1] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [1] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [1] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [1] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [1] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedLessInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [1] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedLessInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [1] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedLessInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [1] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [1] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedLessInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [1] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedLessInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [1] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedLessInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [1] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedLessInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [1] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [1] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [1] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [1] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [1] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedLessInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [1] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedLessInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [1] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedLessInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [1] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [1] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedLessUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [1] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedLessUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [1] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [1] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedLessUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [1] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedLessUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [1] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedLessUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [1] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedLessUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [1] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [1] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [1] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [1] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [1] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [1] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedLessUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [1] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedLessUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [1] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedLessUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [1] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedLessEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [2] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [2] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [2] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessEqualFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [2] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [2] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedLessEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [2] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedLessEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [2] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedLessEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [2] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [2] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [2] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [2] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedLessEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [2] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedLessEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [2] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [2] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [2] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [2] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [2] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedLessEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [2] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedLessEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [2] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedLessEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [2] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [2] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedLessEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [2] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedLessEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [2] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedLessEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [2] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [2] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [2] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedLessEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [2] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedLessEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [2] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [2] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedLessEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [2] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [2] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [2] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedLessEqualUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [2] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedLessEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [2] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedLessEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [2] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedLessEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [2] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedMaxFloat32x16 x y mask) => (VMAXPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxFloat32x4 x y mask) => (VMAXPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMaxFloat32x8 x y mask) => (VMAXPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMaxFloat32x16 x y mask) => (VMAXPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxFloat64x2 x y mask) => (VMAXPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMaxFloat64x4 x y mask) => (VMAXPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMaxFloat64x8 x y mask) => (VMAXPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedMaxInt8x16 x y mask) => (VPMAXSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedMaxInt8x32 x y mask) => (VPMAXSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedMaxInt8x64 x y mask) => (VPMAXSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMaxInt16x8 x y mask) => (VPMAXSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMaxInt16x16 x y mask) => (VPMAXSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMaxInt16x32 x y mask) => (VPMAXSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMaxInt16x8 x y mask) => (VPMAXSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedMaxInt32x16 x y mask) => (VPMAXSDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxInt32x4 x y mask) => (VPMAXSDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMaxInt32x8 x y mask) => (VPMAXSDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMaxInt32x16 x y mask) => (VPMAXSDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxInt64x2 x y mask) => (VPMAXSQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMaxInt64x4 x y mask) => (VPMAXSQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMaxInt64x8 x y mask) => (VPMAXSQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedMaxInt8x16 x y mask) => (VPMAXSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedMaxInt8x32 x y mask) => (VPMAXSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedMaxInt8x64 x y mask) => (VPMAXSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMaxUint8x16 x y mask) => (VPMAXUBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedMaxUint8x32 x y mask) => (VPMAXUBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedMaxUint8x64 x y mask) => (VPMAXUBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMaxUint16x8 x y mask) => (VPMAXUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMaxUint16x16 x y mask) => (VPMAXUWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMaxUint16x32 x y mask) => (VPMAXUWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMaxUint16x8 x y mask) => (VPMAXUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedMaxUint32x16 x y mask) => (VPMAXUDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxUint32x4 x y mask) => (VPMAXUDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMaxUint32x8 x y mask) => (VPMAXUDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMaxUint32x16 x y mask) => (VPMAXUDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMaxUint64x2 x y mask) => (VPMAXUQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMaxUint64x4 x y mask) => (VPMAXUQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMaxUint64x8 x y mask) => (VPMAXUQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedMaxUint8x16 x y mask) => (VPMAXUBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedMaxUint8x32 x y mask) => (VPMAXUBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedMaxUint8x64 x y mask) => (VPMAXUBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedMinFloat32x16 x y mask) => (VMINPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinFloat32x4 x y mask) => (VMINPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMinFloat32x8 x y mask) => (VMINPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMinFloat32x16 x y mask) => (VMINPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinFloat64x2 x y mask) => (VMINPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMinFloat64x4 x y mask) => (VMINPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMinFloat64x8 x y mask) => (VMINPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedMinInt8x16 x y mask) => (VPMINSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedMinInt8x32 x y mask) => (VPMINSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedMinInt8x64 x y mask) => (VPMINSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMinInt16x8 x y mask) => (VPMINSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMinInt16x16 x y mask) => (VPMINSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMinInt16x32 x y mask) => (VPMINSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMinInt16x8 x y mask) => (VPMINSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedMinInt32x16 x y mask) => (VPMINSDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinInt32x4 x y mask) => (VPMINSDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMinInt32x8 x y mask) => (VPMINSDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMinInt32x16 x y mask) => (VPMINSDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinInt64x2 x y mask) => (VPMINSQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMinInt64x4 x y mask) => (VPMINSQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMinInt64x8 x y mask) => (VPMINSQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedMinInt8x16 x y mask) => (VPMINSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedMinInt8x32 x y mask) => (VPMINSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedMinInt8x64 x y mask) => (VPMINSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMinUint8x16 x y mask) => (VPMINUBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedMinUint8x32 x y mask) => (VPMINUBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedMinUint8x64 x y mask) => (VPMINUBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedMinUint16x8 x y mask) => (VPMINUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMinUint16x16 x y mask) => (VPMINUWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMinUint16x32 x y mask) => (VPMINUWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMinUint16x8 x y mask) => (VPMINUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedMinUint32x16 x y mask) => (VPMINUDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinUint32x4 x y mask) => (VPMINUDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMinUint32x8 x y mask) => (VPMINUDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMinUint32x16 x y mask) => (VPMINUDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMinUint64x2 x y mask) => (VPMINUQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMinUint64x4 x y mask) => (VPMINUQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMinUint64x8 x y mask) => (VPMINUQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedMinUint8x16 x y mask) => (VPMINUBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedMinUint8x32 x y mask) => (VPMINUBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedMinUint8x64 x y mask) => (VPMINUBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedMulFloat32x16 x y mask) => (VMULPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulFloat32x4 x y mask) => (VMULPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMulFloat32x8 x y mask) => (VMULPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMulFloat32x16 x y mask) => (VMULPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulFloat64x2 x y mask) => (VMULPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMulFloat64x4 x y mask) => (VMULPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMulFloat64x8 x y mask) => (VMULPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedMulByPowOf2Float32x16 x y mask) => (VSCALEFPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulByPowOf2Float32x4 x y mask) => (VSCALEFPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMulByPowOf2Float32x8 x y mask) => (VSCALEFPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMulByPowOf2Float32x16 x y mask) => (VSCALEFPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulByPowOf2Float64x2 x y mask) => (VSCALEFPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMulByPowOf2Float64x4 x y mask) => (VSCALEFPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMulByPowOf2Float64x8 x y mask) => (VSCALEFPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
@@ -811,288 +811,288 @@
 (MaskedMulEvenWidenUint64x2 x y mask) => (VPMULUDQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMulEvenWidenUint64x4 x y mask) => (VPMULUDQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMulEvenWidenUint64x8 x y mask) => (VPMULUDQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedMulHighInt16x8 x y mask) => (VPMULHWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMulHighInt16x16 x y mask) => (VPMULHWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMulHighInt16x32 x y mask) => (VPMULHWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMulHighInt16x8 x y mask) => (VPMULHWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedMulHighUint16x8 x y mask) => (VPMULHUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMulHighUint16x16 x y mask) => (VPMULHUWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMulHighUint16x32 x y mask) => (VPMULHUWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMulHighUint16x8 x y mask) => (VPMULHUWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedMulLowInt16x8 x y mask) => (VPMULLWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedMulLowInt16x16 x y mask) => (VPMULLWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedMulLowInt16x32 x y mask) => (VPMULLWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedMulLowInt16x8 x y mask) => (VPMULLWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedMulLowInt32x16 x y mask) => (VPMULLDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulLowInt32x4 x y mask) => (VPMULLDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedMulLowInt32x8 x y mask) => (VPMULLDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedMulLowInt32x16 x y mask) => (VPMULLDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedMulLowInt64x2 x y mask) => (VPMULLQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedMulLowInt64x4 x y mask) => (VPMULLQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedMulLowInt64x8 x y mask) => (VPMULLQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedNotEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualFloat32x4 x y mask) => (VPMOVMToVec32x4 (VCMPPSMasked128 [4] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualFloat32x8 x y mask) => (VPMOVMToVec32x8 (VCMPPSMasked256 [4] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedNotEqualFloat32x16 x y mask) => (VPMOVMToVec32x16 (VCMPPSMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualFloat64x2 x y mask) => (VPMOVMToVec64x2 (VCMPPDMasked128 [4] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedNotEqualFloat64x4 x y mask) => (VPMOVMToVec64x4 (VCMPPDMasked256 [4] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualFloat64x8 x y mask) => (VPMOVMToVec64x8 (VCMPPDMasked512 [4] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
+(MaskedNotEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [4] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedNotEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [4] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedNotEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [4] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedNotEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [4] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPWMasked256 [4] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPWMasked512 [4] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedNotEqualInt16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPWMasked128 [4] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedNotEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPDMasked128 [4] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPDMasked256 [4] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedNotEqualInt32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPDMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPQMasked128 [4] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPQMasked256 [4] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualInt64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPQMasked512 [4] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedNotEqualInt8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPBMasked128 [4] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedNotEqualInt8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPBMasked256 [4] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedNotEqualInt8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPBMasked512 [4] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedNotEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [4] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
+(MaskedNotEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [4] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
+(MaskedNotEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [4] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
+(MaskedNotEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [4] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint16x16 x y mask) => (VPMOVMToVec16x16 (VPCMPUWMasked256 [4] x y (VPMOVVec16x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint16x32 x y mask) => (VPMOVMToVec16x32 (VPCMPUWMasked512 [4] x y (VPMOVVec16x32ToM <types.TypeMask> mask)))
-(MaskedNotEqualUint16x8 x y mask) => (VPMOVMToVec16x8 (VPCMPUWMasked128 [4] x y (VPMOVVec16x8ToM <types.TypeMask> mask)))
-(MaskedNotEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint32x4 x y mask) => (VPMOVMToVec32x4 (VPCMPUDMasked128 [4] x y (VPMOVVec32x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint32x8 x y mask) => (VPMOVMToVec32x8 (VPCMPUDMasked256 [4] x y (VPMOVVec32x8ToM <types.TypeMask> mask)))
+(MaskedNotEqualUint32x16 x y mask) => (VPMOVMToVec32x16 (VPCMPUDMasked512 [4] x y (VPMOVVec32x16ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint64x2 x y mask) => (VPMOVMToVec64x2 (VPCMPUQMasked128 [4] x y (VPMOVVec64x2ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint64x4 x y mask) => (VPMOVMToVec64x4 (VPCMPUQMasked256 [4] x y (VPMOVVec64x4ToM <types.TypeMask> mask)))
 (MaskedNotEqualUint64x8 x y mask) => (VPMOVMToVec64x8 (VPCMPUQMasked512 [4] x y (VPMOVVec64x8ToM <types.TypeMask> mask)))
-(MaskedNotEqualUint8x16 x y mask) => (VPMOVMToVec8x16 (VPCMPUBMasked128 [4] x y (VPMOVVec8x16ToM <types.TypeMask> mask)))
-(MaskedNotEqualUint8x32 x y mask) => (VPMOVMToVec8x32 (VPCMPUBMasked256 [4] x y (VPMOVVec8x32ToM <types.TypeMask> mask)))
-(MaskedNotEqualUint8x64 x y mask) => (VPMOVMToVec8x64 (VPCMPUBMasked512 [4] x y (VPMOVVec8x64ToM <types.TypeMask> mask)))
-(MaskedOrFloat32x16 x y mask) => (VORPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrFloat32x4 x y mask) => (VORPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedOrFloat32x8 x y mask) => (VORPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedOrFloat32x16 x y mask) => (VORPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrFloat64x2 x y mask) => (VORPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedOrFloat64x4 x y mask) => (VORPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedOrFloat64x8 x y mask) => (VORPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedOrInt32x16 x y mask) => (VPORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrInt32x4 x y mask) => (VPORDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedOrInt32x8 x y mask) => (VPORDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedOrInt32x16 x y mask) => (VPORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrInt64x2 x y mask) => (VPORQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedOrInt64x4 x y mask) => (VPORQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedOrInt64x8 x y mask) => (VPORQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedOrUint32x16 x y mask) => (VPORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrUint32x4 x y mask) => (VPORDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedOrUint32x8 x y mask) => (VPORDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedOrUint32x16 x y mask) => (VPORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedOrUint64x2 x y mask) => (VPORQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedOrUint64x4 x y mask) => (VPORQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedOrUint64x8 x y mask) => (VPORQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedPairDotProdInt16x8 x y mask) => (VPMADDWDMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedPairDotProdInt16x16 x y mask) => (VPMADDWDMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedPairDotProdInt16x32 x y mask) => (VPMADDWDMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedPairDotProdInt16x8 x y mask) => (VPMADDWDMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedPairDotProdAccumulateInt32x16 x y z mask) => (VPDPWSSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedPairDotProdAccumulateInt32x4 x y z mask) => (VPDPWSSDMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedPairDotProdAccumulateInt32x8 x y z mask) => (VPDPWSSDMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedPairDotProdAccumulateInt32x16 x y z mask) => (VPDPWSSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedPopCountInt8x16 x mask) => (VPOPCNTBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedPopCountInt8x32 x mask) => (VPOPCNTBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedPopCountInt8x64 x mask) => (VPOPCNTBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedPopCountInt16x8 x mask) => (VPOPCNTWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedPopCountInt16x16 x mask) => (VPOPCNTWMasked256 x (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedPopCountInt16x32 x mask) => (VPOPCNTWMasked512 x (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedPopCountInt16x8 x mask) => (VPOPCNTWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedPopCountInt32x16 x mask) => (VPOPCNTDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedPopCountInt32x4 x mask) => (VPOPCNTDMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedPopCountInt32x8 x mask) => (VPOPCNTDMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedPopCountInt32x16 x mask) => (VPOPCNTDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedPopCountInt64x2 x mask) => (VPOPCNTQMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedPopCountInt64x4 x mask) => (VPOPCNTQMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedPopCountInt64x8 x mask) => (VPOPCNTQMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedPopCountInt8x16 x mask) => (VPOPCNTBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedPopCountInt8x32 x mask) => (VPOPCNTBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedPopCountInt8x64 x mask) => (VPOPCNTBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedPopCountUint8x16 x mask) => (VPOPCNTBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedPopCountUint8x32 x mask) => (VPOPCNTBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedPopCountUint8x64 x mask) => (VPOPCNTBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedPopCountUint16x8 x mask) => (VPOPCNTWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedPopCountUint16x16 x mask) => (VPOPCNTWMasked256 x (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedPopCountUint16x32 x mask) => (VPOPCNTWMasked512 x (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedPopCountUint16x8 x mask) => (VPOPCNTWMasked128 x (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedPopCountUint32x16 x mask) => (VPOPCNTDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedPopCountUint32x4 x mask) => (VPOPCNTDMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedPopCountUint32x8 x mask) => (VPOPCNTDMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedPopCountUint32x16 x mask) => (VPOPCNTDMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedPopCountUint64x2 x mask) => (VPOPCNTQMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedPopCountUint64x4 x mask) => (VPOPCNTQMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedPopCountUint64x8 x mask) => (VPOPCNTQMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedPopCountUint8x16 x mask) => (VPOPCNTBMasked128 x (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedPopCountUint8x32 x mask) => (VPOPCNTBMasked256 x (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedPopCountUint8x64 x mask) => (VPOPCNTBMasked512 x (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedRoundSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+8] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedRoundSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+8] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedRoundSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+8] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedRoundSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+8] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedRoundSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+8] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedRoundSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+8] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedRoundSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+8] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedRoundWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+0] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedRoundWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+0] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedRoundWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+0] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedRoundWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+0] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedRoundWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+0] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedRoundWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+0] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedRoundWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+0] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedSaturatedAddInt16x16 x y mask) => (VPADDSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
-(MaskedSaturatedAddInt16x32 x y mask) => (VPADDSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSaturatedAddInt16x8 x y mask) => (VPADDSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedSaturatedAddInt8x16 x y mask) => (VPADDSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
 (MaskedSaturatedAddInt8x32 x y mask) => (VPADDSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
 (MaskedSaturatedAddInt8x64 x y mask) => (VPADDSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedSaturatedAddUint16x16 x y mask) => (VPADDSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
-(MaskedSaturatedAddUint16x32 x y mask) => (VPADDSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSaturatedAddUint16x8 x y mask) => (VPADDSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedAddInt16x8 x y mask) => (VPADDSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedAddInt16x16 x y mask) => (VPADDSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(MaskedSaturatedAddInt16x32 x y mask) => (VPADDSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (MaskedSaturatedAddUint8x16 x y mask) => (VPADDSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
 (MaskedSaturatedAddUint8x32 x y mask) => (VPADDSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
 (MaskedSaturatedAddUint8x64 x y mask) => (VPADDSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedSaturatedPairDotProdAccumulateInt32x16 x y z mask) => (VPDPWSSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedSaturatedAddUint16x8 x y mask) => (VPADDSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedAddUint16x16 x y mask) => (VPADDSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(MaskedSaturatedAddUint16x32 x y mask) => (VPADDSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (MaskedSaturatedPairDotProdAccumulateInt32x4 x y z mask) => (VPDPWSSDSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSaturatedPairDotProdAccumulateInt32x8 x y z mask) => (VPDPWSSDSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
-(MaskedSaturatedSubInt16x16 x y mask) => (VPSUBSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
-(MaskedSaturatedSubInt16x32 x y mask) => (VPSUBSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSaturatedSubInt16x8 x y mask) => (VPSUBSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedPairDotProdAccumulateInt32x16 x y z mask) => (VPDPWSSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSaturatedSubInt8x16 x y mask) => (VPSUBSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
 (MaskedSaturatedSubInt8x32 x y mask) => (VPSUBSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
 (MaskedSaturatedSubInt8x64 x y mask) => (VPSUBSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedSaturatedSubUint16x16 x y mask) => (VPSUBSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
-(MaskedSaturatedSubUint16x32 x y mask) => (VPSUBSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSaturatedSubUint16x8 x y mask) => (VPSUBSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedSubInt16x8 x y mask) => (VPSUBSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedSubInt16x16 x y mask) => (VPSUBSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(MaskedSaturatedSubInt16x32 x y mask) => (VPSUBSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (MaskedSaturatedSubUint8x16 x y mask) => (VPSUBSBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
 (MaskedSaturatedSubUint8x32 x y mask) => (VPSUBSBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
 (MaskedSaturatedSubUint8x64 x y mask) => (VPSUBSBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedSaturatedSubUint16x8 x y mask) => (VPSUBSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
+(MaskedSaturatedSubUint16x16 x y mask) => (VPSUBSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
+(MaskedSaturatedSubUint16x32 x y mask) => (VPSUBSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedPairDotProdUint8x16 x y mask) => (VPMADDUBSWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedPairDotProdUint8x32 x y mask) => (VPMADDUBSWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedPairDotProdUint8x64 x y mask) => (VPMADDUBSWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSaturatedUnsignedSignedQuadDotProdAccumulateInt32x16 x y z mask) => (VPDPBUSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedQuadDotProdAccumulateInt32x4 x y z mask) => (VPDPBUSDSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedQuadDotProdAccumulateInt32x8 x y z mask) => (VPDPBUSDSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
-(MaskedSaturatedUnsignedSignedQuadDotProdAccumulateUint32x16 x y z mask) => (VPDPBUSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedSaturatedUnsignedSignedQuadDotProdAccumulateInt32x16 x y z mask) => (VPDPBUSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedQuadDotProdAccumulateUint32x4 x y z mask) => (VPDPBUSDSMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSaturatedUnsignedSignedQuadDotProdAccumulateUint32x8 x y z mask) => (VPDPBUSDSMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
-(MaskedSqrtFloat32x16 x mask) => (VSQRTPSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedSaturatedUnsignedSignedQuadDotProdAccumulateUint32x16 x y z mask) => (VPDPBUSDSMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSqrtFloat32x4 x mask) => (VSQRTPSMasked128 x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSqrtFloat32x8 x mask) => (VSQRTPSMasked256 x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedSqrtFloat32x16 x mask) => (VSQRTPSMasked512 x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSqrtFloat64x2 x mask) => (VSQRTPDMasked128 x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedSqrtFloat64x4 x mask) => (VSQRTPDMasked256 x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedSqrtFloat64x8 x mask) => (VSQRTPDMasked512 x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedSubFloat32x16 x y mask) => (VSUBPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubFloat32x4 x y mask) => (VSUBPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSubFloat32x8 x y mask) => (VSUBPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedSubFloat32x16 x y mask) => (VSUBPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubFloat64x2 x y mask) => (VSUBPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedSubFloat64x4 x y mask) => (VSUBPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedSubFloat64x8 x y mask) => (VSUBPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(MaskedSubInt8x16 x y mask) => (VPSUBBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedSubInt8x32 x y mask) => (VPSUBBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedSubInt8x64 x y mask) => (VPSUBBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedSubInt16x8 x y mask) => (VPSUBWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedSubInt16x16 x y mask) => (VPSUBWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedSubInt16x32 x y mask) => (VPSUBWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSubInt16x8 x y mask) => (VPSUBWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedSubInt32x16 x y mask) => (VPSUBDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubInt32x4 x y mask) => (VPSUBDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSubInt32x8 x y mask) => (VPSUBDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedSubInt32x16 x y mask) => (VPSUBDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubInt64x2 x y mask) => (VPSUBQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedSubInt64x4 x y mask) => (VPSUBQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedSubInt64x8 x y mask) => (VPSUBQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedSubInt8x16 x y mask) => (VPSUBBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedSubInt8x32 x y mask) => (VPSUBBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedSubInt8x64 x y mask) => (VPSUBBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedSubUint8x16 x y mask) => (VPSUBBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
+(MaskedSubUint8x32 x y mask) => (VPSUBBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
+(MaskedSubUint8x64 x y mask) => (VPSUBBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
+(MaskedSubUint16x8 x y mask) => (VPSUBWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
 (MaskedSubUint16x16 x y mask) => (VPSUBWMasked256 x y (VPMOVVec16x16ToM <types.TypeMask> mask))
 (MaskedSubUint16x32 x y mask) => (VPSUBWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
-(MaskedSubUint16x8 x y mask) => (VPSUBWMasked128 x y (VPMOVVec16x8ToM <types.TypeMask> mask))
-(MaskedSubUint32x16 x y mask) => (VPSUBDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubUint32x4 x y mask) => (VPSUBDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedSubUint32x8 x y mask) => (VPSUBDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedSubUint32x16 x y mask) => (VPSUBDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedSubUint64x2 x y mask) => (VPSUBQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedSubUint64x4 x y mask) => (VPSUBQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedSubUint64x8 x y mask) => (VPSUBQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedSubUint8x16 x y mask) => (VPSUBBMasked128 x y (VPMOVVec8x16ToM <types.TypeMask> mask))
-(MaskedSubUint8x32 x y mask) => (VPSUBBMasked256 x y (VPMOVVec8x32ToM <types.TypeMask> mask))
-(MaskedSubUint8x64 x y mask) => (VPSUBBMasked512 x y (VPMOVVec8x64ToM <types.TypeMask> mask))
-(MaskedTruncSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+11] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedTruncSuppressExceptionWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+11] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedTruncSuppressExceptionWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+11] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedTruncSuppressExceptionWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+11] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedTruncSuppressExceptionWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+11] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedTruncSuppressExceptionWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+11] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedTruncSuppressExceptionWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+11] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedTruncWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+3] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedTruncWithPrecisionFloat32x4 [a] x mask) => (VRNDSCALEPSMasked128 [a+3] x (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedTruncWithPrecisionFloat32x8 [a] x mask) => (VRNDSCALEPSMasked256 [a+3] x (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedTruncWithPrecisionFloat32x16 [a] x mask) => (VRNDSCALEPSMasked512 [a+3] x (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedTruncWithPrecisionFloat64x2 [a] x mask) => (VRNDSCALEPDMasked128 [a+3] x (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedTruncWithPrecisionFloat64x4 [a] x mask) => (VRNDSCALEPDMasked256 [a+3] x (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedTruncWithPrecisionFloat64x8 [a] x mask) => (VRNDSCALEPDMasked512 [a+3] x (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedUnsignedSignedQuadDotProdAccumulateInt32x16 x y z mask) => (VPDPBUSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedUnsignedSignedQuadDotProdAccumulateInt32x4 x y z mask) => (VPDPBUSDMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedUnsignedSignedQuadDotProdAccumulateInt32x8 x y z mask) => (VPDPBUSDMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
-(MaskedUnsignedSignedQuadDotProdAccumulateUint32x16 x y z mask) => (VPDPBUSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedUnsignedSignedQuadDotProdAccumulateInt32x16 x y z mask) => (VPDPBUSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedUnsignedSignedQuadDotProdAccumulateUint32x4 x y z mask) => (VPDPBUSDMasked128 x y z (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedUnsignedSignedQuadDotProdAccumulateUint32x8 x y z mask) => (VPDPBUSDMasked256 x y z (VPMOVVec32x8ToM <types.TypeMask> mask))
-(MaskedXorFloat32x16 x y mask) => (VXORPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
+(MaskedUnsignedSignedQuadDotProdAccumulateUint32x16 x y z mask) => (VPDPBUSDMasked512 x y z (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorFloat32x4 x y mask) => (VXORPSMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedXorFloat32x8 x y mask) => (VXORPSMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedXorFloat32x16 x y mask) => (VXORPSMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorFloat64x2 x y mask) => (VXORPDMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedXorFloat64x4 x y mask) => (VXORPDMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedXorFloat64x8 x y mask) => (VXORPDMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedXorInt32x16 x y mask) => (VPXORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorInt32x4 x y mask) => (VPXORDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedXorInt32x8 x y mask) => (VPXORDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedXorInt32x16 x y mask) => (VPXORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorInt64x2 x y mask) => (VPXORQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedXorInt64x4 x y mask) => (VPXORQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedXorInt64x8 x y mask) => (VPXORQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaskedXorUint32x16 x y mask) => (VPXORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorUint32x4 x y mask) => (VPXORDMasked128 x y (VPMOVVec32x4ToM <types.TypeMask> mask))
 (MaskedXorUint32x8 x y mask) => (VPXORDMasked256 x y (VPMOVVec32x8ToM <types.TypeMask> mask))
+(MaskedXorUint32x16 x y mask) => (VPXORDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (MaskedXorUint64x2 x y mask) => (VPXORQMasked128 x y (VPMOVVec64x2ToM <types.TypeMask> mask))
 (MaskedXorUint64x4 x y mask) => (VPXORQMasked256 x y (VPMOVVec64x4ToM <types.TypeMask> mask))
 (MaskedXorUint64x8 x y mask) => (VPXORQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
-(MaxFloat32x16 ...) => (VMAXPS512 ...)
 (MaxFloat32x4 ...) => (VMAXPS128 ...)
 (MaxFloat32x8 ...) => (VMAXPS256 ...)
+(MaxFloat32x16 ...) => (VMAXPS512 ...)
 (MaxFloat64x2 ...) => (VMAXPD128 ...)
 (MaxFloat64x4 ...) => (VMAXPD256 ...)
 (MaxFloat64x8 ...) => (VMAXPD512 ...)
+(MaxInt8x16 ...) => (VPMAXSB128 ...)
+(MaxInt8x32 ...) => (VPMAXSB256 ...)
+(MaxInt8x64 ...) => (VPMAXSB512 ...)
+(MaxInt16x8 ...) => (VPMAXSW128 ...)
 (MaxInt16x16 ...) => (VPMAXSW256 ...)
 (MaxInt16x32 ...) => (VPMAXSW512 ...)
-(MaxInt16x8 ...) => (VPMAXSW128 ...)
-(MaxInt32x16 ...) => (VPMAXSD512 ...)
 (MaxInt32x4 ...) => (VPMAXSD128 ...)
 (MaxInt32x8 ...) => (VPMAXSD256 ...)
+(MaxInt32x16 ...) => (VPMAXSD512 ...)
 (MaxInt64x2 ...) => (VPMAXSQ128 ...)
 (MaxInt64x4 ...) => (VPMAXSQ256 ...)
 (MaxInt64x8 ...) => (VPMAXSQ512 ...)
-(MaxInt8x16 ...) => (VPMAXSB128 ...)
-(MaxInt8x32 ...) => (VPMAXSB256 ...)
-(MaxInt8x64 ...) => (VPMAXSB512 ...)
+(MaxUint8x16 ...) => (VPMAXUB128 ...)
+(MaxUint8x32 ...) => (VPMAXUB256 ...)
+(MaxUint8x64 ...) => (VPMAXUB512 ...)
+(MaxUint16x8 ...) => (VPMAXUW128 ...)
 (MaxUint16x16 ...) => (VPMAXUW256 ...)
 (MaxUint16x32 ...) => (VPMAXUW512 ...)
-(MaxUint16x8 ...) => (VPMAXUW128 ...)
-(MaxUint32x16 ...) => (VPMAXUD512 ...)
 (MaxUint32x4 ...) => (VPMAXUD128 ...)
 (MaxUint32x8 ...) => (VPMAXUD256 ...)
+(MaxUint32x16 ...) => (VPMAXUD512 ...)
 (MaxUint64x2 ...) => (VPMAXUQ128 ...)
 (MaxUint64x4 ...) => (VPMAXUQ256 ...)
 (MaxUint64x8 ...) => (VPMAXUQ512 ...)
-(MaxUint8x16 ...) => (VPMAXUB128 ...)
-(MaxUint8x32 ...) => (VPMAXUB256 ...)
-(MaxUint8x64 ...) => (VPMAXUB512 ...)
-(MinFloat32x16 ...) => (VMINPS512 ...)
 (MinFloat32x4 ...) => (VMINPS128 ...)
 (MinFloat32x8 ...) => (VMINPS256 ...)
+(MinFloat32x16 ...) => (VMINPS512 ...)
 (MinFloat64x2 ...) => (VMINPD128 ...)
 (MinFloat64x4 ...) => (VMINPD256 ...)
 (MinFloat64x8 ...) => (VMINPD512 ...)
+(MinInt8x16 ...) => (VPMINSB128 ...)
+(MinInt8x32 ...) => (VPMINSB256 ...)
+(MinInt8x64 ...) => (VPMINSB512 ...)
+(MinInt16x8 ...) => (VPMINSW128 ...)
 (MinInt16x16 ...) => (VPMINSW256 ...)
 (MinInt16x32 ...) => (VPMINSW512 ...)
-(MinInt16x8 ...) => (VPMINSW128 ...)
-(MinInt32x16 ...) => (VPMINSD512 ...)
 (MinInt32x4 ...) => (VPMINSD128 ...)
 (MinInt32x8 ...) => (VPMINSD256 ...)
+(MinInt32x16 ...) => (VPMINSD512 ...)
 (MinInt64x2 ...) => (VPMINSQ128 ...)
 (MinInt64x4 ...) => (VPMINSQ256 ...)
 (MinInt64x8 ...) => (VPMINSQ512 ...)
-(MinInt8x16 ...) => (VPMINSB128 ...)
-(MinInt8x32 ...) => (VPMINSB256 ...)
-(MinInt8x64 ...) => (VPMINSB512 ...)
+(MinUint8x16 ...) => (VPMINUB128 ...)
+(MinUint8x32 ...) => (VPMINUB256 ...)
+(MinUint8x64 ...) => (VPMINUB512 ...)
+(MinUint16x8 ...) => (VPMINUW128 ...)
 (MinUint16x16 ...) => (VPMINUW256 ...)
 (MinUint16x32 ...) => (VPMINUW512 ...)
-(MinUint16x8 ...) => (VPMINUW128 ...)
-(MinUint32x16 ...) => (VPMINUD512 ...)
 (MinUint32x4 ...) => (VPMINUD128 ...)
 (MinUint32x8 ...) => (VPMINUD256 ...)
+(MinUint32x16 ...) => (VPMINUD512 ...)
 (MinUint64x2 ...) => (VPMINUQ128 ...)
 (MinUint64x4 ...) => (VPMINUQ256 ...)
 (MinUint64x8 ...) => (VPMINUQ512 ...)
-(MinUint8x16 ...) => (VPMINUB128 ...)
-(MinUint8x32 ...) => (VPMINUB256 ...)
-(MinUint8x64 ...) => (VPMINUB512 ...)
-(MulFloat32x16 ...) => (VMULPS512 ...)
 (MulFloat32x4 ...) => (VMULPS128 ...)
 (MulFloat32x8 ...) => (VMULPS256 ...)
+(MulFloat32x16 ...) => (VMULPS512 ...)
 (MulFloat64x2 ...) => (VMULPD128 ...)
 (MulFloat64x4 ...) => (VMULPD256 ...)
 (MulFloat64x8 ...) => (VMULPD512 ...)
-(MulByPowOf2Float32x16 ...) => (VSCALEFPS512 ...)
 (MulByPowOf2Float32x4 ...) => (VSCALEFPS128 ...)
 (MulByPowOf2Float32x8 ...) => (VSCALEFPS256 ...)
+(MulByPowOf2Float32x16 ...) => (VSCALEFPS512 ...)
 (MulByPowOf2Float64x2 ...) => (VSCALEFPD128 ...)
 (MulByPowOf2Float64x4 ...) => (VSCALEFPD256 ...)
 (MulByPowOf2Float64x8 ...) => (VSCALEFPD512 ...)
@@ -1106,282 +1106,282 @@
 (MulEvenWidenUint64x2 ...) => (VPMULUDQ128 ...)
 (MulEvenWidenUint64x4 ...) => (VPMULUDQ256 ...)
 (MulEvenWidenUint64x8 ...) => (VPMULUDQ512 ...)
+(MulHighInt16x8 ...) => (VPMULHW128 ...)
 (MulHighInt16x16 ...) => (VPMULHW256 ...)
 (MulHighInt16x32 ...) => (VPMULHW512 ...)
-(MulHighInt16x8 ...) => (VPMULHW128 ...)
+(MulHighUint16x8 ...) => (VPMULHUW128 ...)
 (MulHighUint16x16 ...) => (VPMULHUW256 ...)
 (MulHighUint16x32 ...) => (VPMULHUW512 ...)
-(MulHighUint16x8 ...) => (VPMULHUW128 ...)
+(MulLowInt16x8 ...) => (VPMULLW128 ...)
 (MulLowInt16x16 ...) => (VPMULLW256 ...)
 (MulLowInt16x32 ...) => (VPMULLW512 ...)
-(MulLowInt16x8 ...) => (VPMULLW128 ...)
-(MulLowInt32x16 ...) => (VPMULLD512 ...)
 (MulLowInt32x4 ...) => (VPMULLD128 ...)
 (MulLowInt32x8 ...) => (VPMULLD256 ...)
+(MulLowInt32x16 ...) => (VPMULLD512 ...)
 (MulLowInt64x2 ...) => (VPMULLQ128 ...)
 (MulLowInt64x4 ...) => (VPMULLQ256 ...)
 (MulLowInt64x8 ...) => (VPMULLQ512 ...)
-(NotEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [4] x y))
 (NotEqualFloat32x4 x y) => (VCMPPS128 [4] x y)
 (NotEqualFloat32x8 x y) => (VCMPPS256 [4] x y)
+(NotEqualFloat32x16 x y) => (VPMOVMToVec32x16 (VCMPPS512 [4] x y))
 (NotEqualFloat64x2 x y) => (VCMPPD128 [4] x y)
 (NotEqualFloat64x4 x y) => (VCMPPD256 [4] x y)
 (NotEqualFloat64x8 x y) => (VPMOVMToVec64x8 (VCMPPD512 [4] x y))
+(NotEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [4] x y))
+(NotEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [4] x y))
+(NotEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [4] x y))
+(NotEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [4] x y))
 (NotEqualInt16x16 x y) => (VPMOVMToVec16x16 (VPCMPW256 [4] x y))
 (NotEqualInt16x32 x y) => (VPMOVMToVec16x32 (VPCMPW512 [4] x y))
-(NotEqualInt16x8 x y) => (VPMOVMToVec16x8 (VPCMPW128 [4] x y))
-(NotEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [4] x y))
 (NotEqualInt32x4 x y) => (VPMOVMToVec32x4 (VPCMPD128 [4] x y))
 (NotEqualInt32x8 x y) => (VPMOVMToVec32x8 (VPCMPD256 [4] x y))
+(NotEqualInt32x16 x y) => (VPMOVMToVec32x16 (VPCMPD512 [4] x y))
 (NotEqualInt64x2 x y) => (VPMOVMToVec64x2 (VPCMPQ128 [4] x y))
 (NotEqualInt64x4 x y) => (VPMOVMToVec64x4 (VPCMPQ256 [4] x y))
 (NotEqualInt64x8 x y) => (VPMOVMToVec64x8 (VPCMPQ512 [4] x y))
-(NotEqualInt8x16 x y) => (VPMOVMToVec8x16 (VPCMPB128 [4] x y))
-(NotEqualInt8x32 x y) => (VPMOVMToVec8x32 (VPCMPB256 [4] x y))
-(NotEqualInt8x64 x y) => (VPMOVMToVec8x64 (VPCMPB512 [4] x y))
+(NotEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [4] x y))
+(NotEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [4] x y))
+(NotEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [4] x y))
+(NotEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [4] x y))
 (NotEqualUint16x16 x y) => (VPMOVMToVec16x16 (VPCMPUW256 [4] x y))
 (NotEqualUint16x32 x y) => (VPMOVMToVec16x32 (VPCMPUW512 [4] x y))
-(NotEqualUint16x8 x y) => (VPMOVMToVec16x8 (VPCMPUW128 [4] x y))
-(NotEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [4] x y))
 (NotEqualUint32x4 x y) => (VPMOVMToVec32x4 (VPCMPUD128 [4] x y))
 (NotEqualUint32x8 x y) => (VPMOVMToVec32x8 (VPCMPUD256 [4] x y))
+(NotEqualUint32x16 x y) => (VPMOVMToVec32x16 (VPCMPUD512 [4] x y))
 (NotEqualUint64x2 x y) => (VPMOVMToVec64x2 (VPCMPUQ128 [4] x y))
 (NotEqualUint64x4 x y) => (VPMOVMToVec64x4 (VPCMPUQ256 [4] x y))
 (NotEqualUint64x8 x y) => (VPMOVMToVec64x8 (VPCMPUQ512 [4] x y))
-(NotEqualUint8x16 x y) => (VPMOVMToVec8x16 (VPCMPUB128 [4] x y))
-(NotEqualUint8x32 x y) => (VPMOVMToVec8x32 (VPCMPUB256 [4] x y))
-(NotEqualUint8x64 x y) => (VPMOVMToVec8x64 (VPCMPUB512 [4] x y))
-(OrFloat32x16 ...) => (VORPS512 ...)
 (OrFloat32x4 ...) => (VORPS128 ...)
 (OrFloat32x8 ...) => (VORPS256 ...)
+(OrFloat32x16 ...) => (VORPS512 ...)
 (OrFloat64x2 ...) => (VORPD128 ...)
 (OrFloat64x4 ...) => (VORPD256 ...)
 (OrFloat64x8 ...) => (VORPD512 ...)
-(OrInt16x16 ...) => (VPOR256 ...)
+(OrInt8x16 ...) => (VPOR128 ...)
+(OrInt8x32 ...) => (VPOR256 ...)
 (OrInt16x8 ...) => (VPOR128 ...)
-(OrInt32x16 ...) => (VPORD512 ...)
+(OrInt16x16 ...) => (VPOR256 ...)
 (OrInt32x4 ...) => (VPOR128 ...)
 (OrInt32x8 ...) => (VPOR256 ...)
+(OrInt32x16 ...) => (VPORD512 ...)
 (OrInt64x2 ...) => (VPOR128 ...)
 (OrInt64x4 ...) => (VPOR256 ...)
 (OrInt64x8 ...) => (VPORQ512 ...)
-(OrInt8x16 ...) => (VPOR128 ...)
-(OrInt8x32 ...) => (VPOR256 ...)
-(OrUint16x16 ...) => (VPOR256 ...)
+(OrUint8x16 ...) => (VPOR128 ...)
+(OrUint8x32 ...) => (VPOR256 ...)
 (OrUint16x8 ...) => (VPOR128 ...)
-(OrUint32x16 ...) => (VPORD512 ...)
+(OrUint16x16 ...) => (VPOR256 ...)
 (OrUint32x4 ...) => (VPOR128 ...)
 (OrUint32x8 ...) => (VPOR256 ...)
+(OrUint32x16 ...) => (VPORD512 ...)
 (OrUint64x2 ...) => (VPOR128 ...)
 (OrUint64x4 ...) => (VPOR256 ...)
 (OrUint64x8 ...) => (VPORQ512 ...)
-(OrUint8x16 ...) => (VPOR128 ...)
-(OrUint8x32 ...) => (VPOR256 ...)
+(PairDotProdInt16x8 ...) => (VPMADDWD128 ...)
 (PairDotProdInt16x16 ...) => (VPMADDWD256 ...)
 (PairDotProdInt16x32 ...) => (VPMADDWD512 ...)
-(PairDotProdInt16x8 ...) => (VPMADDWD128 ...)
-(PairDotProdAccumulateInt32x16 ...) => (VPDPWSSD512 ...)
 (PairDotProdAccumulateInt32x4 ...) => (VPDPWSSD128 ...)
 (PairDotProdAccumulateInt32x8 ...) => (VPDPWSSD256 ...)
+(PairDotProdAccumulateInt32x16 ...) => (VPDPWSSD512 ...)
 (PairwiseAddFloat32x4 ...) => (VHADDPS128 ...)
 (PairwiseAddFloat32x8 ...) => (VHADDPS256 ...)
 (PairwiseAddFloat64x2 ...) => (VHADDPD128 ...)
 (PairwiseAddFloat64x4 ...) => (VHADDPD256 ...)
-(PairwiseAddInt16x16 ...) => (VPHADDW256 ...)
 (PairwiseAddInt16x8 ...) => (VPHADDW128 ...)
+(PairwiseAddInt16x16 ...) => (VPHADDW256 ...)
 (PairwiseAddInt32x4 ...) => (VPHADDD128 ...)
 (PairwiseAddInt32x8 ...) => (VPHADDD256 ...)
-(PairwiseAddUint16x16 ...) => (VPHADDW256 ...)
 (PairwiseAddUint16x8 ...) => (VPHADDW128 ...)
+(PairwiseAddUint16x16 ...) => (VPHADDW256 ...)
 (PairwiseAddUint32x4 ...) => (VPHADDD128 ...)
 (PairwiseAddUint32x8 ...) => (VPHADDD256 ...)
 (PairwiseSubFloat32x4 ...) => (VHSUBPS128 ...)
 (PairwiseSubFloat32x8 ...) => (VHSUBPS256 ...)
 (PairwiseSubFloat64x2 ...) => (VHSUBPD128 ...)
 (PairwiseSubFloat64x4 ...) => (VHSUBPD256 ...)
-(PairwiseSubInt16x16 ...) => (VPHSUBW256 ...)
 (PairwiseSubInt16x8 ...) => (VPHSUBW128 ...)
+(PairwiseSubInt16x16 ...) => (VPHSUBW256 ...)
 (PairwiseSubInt32x4 ...) => (VPHSUBD128 ...)
 (PairwiseSubInt32x8 ...) => (VPHSUBD256 ...)
-(PairwiseSubUint16x16 ...) => (VPHSUBW256 ...)
 (PairwiseSubUint16x8 ...) => (VPHSUBW128 ...)
+(PairwiseSubUint16x16 ...) => (VPHSUBW256 ...)
 (PairwiseSubUint32x4 ...) => (VPHSUBD128 ...)
 (PairwiseSubUint32x8 ...) => (VPHSUBD256 ...)
+(PopCountInt8x16 ...) => (VPOPCNTB128 ...)
+(PopCountInt8x32 ...) => (VPOPCNTB256 ...)
+(PopCountInt8x64 ...) => (VPOPCNTB512 ...)
+(PopCountInt16x8 ...) => (VPOPCNTW128 ...)
 (PopCountInt16x16 ...) => (VPOPCNTW256 ...)
 (PopCountInt16x32 ...) => (VPOPCNTW512 ...)
-(PopCountInt16x8 ...) => (VPOPCNTW128 ...)
-(PopCountInt32x16 ...) => (VPOPCNTD512 ...)
 (PopCountInt32x4 ...) => (VPOPCNTD128 ...)
 (PopCountInt32x8 ...) => (VPOPCNTD256 ...)
+(PopCountInt32x16 ...) => (VPOPCNTD512 ...)
 (PopCountInt64x2 ...) => (VPOPCNTQ128 ...)
 (PopCountInt64x4 ...) => (VPOPCNTQ256 ...)
 (PopCountInt64x8 ...) => (VPOPCNTQ512 ...)
-(PopCountInt8x16 ...) => (VPOPCNTB128 ...)
-(PopCountInt8x32 ...) => (VPOPCNTB256 ...)
-(PopCountInt8x64 ...) => (VPOPCNTB512 ...)
+(PopCountUint8x16 ...) => (VPOPCNTB128 ...)
+(PopCountUint8x32 ...) => (VPOPCNTB256 ...)
+(PopCountUint8x64 ...) => (VPOPCNTB512 ...)
+(PopCountUint16x8 ...) => (VPOPCNTW128 ...)
 (PopCountUint16x16 ...) => (VPOPCNTW256 ...)
 (PopCountUint16x32 ...) => (VPOPCNTW512 ...)
-(PopCountUint16x8 ...) => (VPOPCNTW128 ...)
-(PopCountUint32x16 ...) => (VPOPCNTD512 ...)
 (PopCountUint32x4 ...) => (VPOPCNTD128 ...)
 (PopCountUint32x8 ...) => (VPOPCNTD256 ...)
+(PopCountUint32x16 ...) => (VPOPCNTD512 ...)
 (PopCountUint64x2 ...) => (VPOPCNTQ128 ...)
 (PopCountUint64x4 ...) => (VPOPCNTQ256 ...)
 (PopCountUint64x8 ...) => (VPOPCNTQ512 ...)
-(PopCountUint8x16 ...) => (VPOPCNTB128 ...)
-(PopCountUint8x32 ...) => (VPOPCNTB256 ...)
-(PopCountUint8x64 ...) => (VPOPCNTB512 ...)
 (RoundFloat32x4 x) => (VROUNDPS128 [0] x)
 (RoundFloat32x8 x) => (VROUNDPS256 [0] x)
 (RoundFloat64x2 x) => (VROUNDPD128 [0] x)
 (RoundFloat64x4 x) => (VROUNDPD256 [0] x)
-(RoundSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+8] x)
 (RoundSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+8] x)
 (RoundSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+8] x)
+(RoundSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+8] x)
 (RoundSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+8] x)
 (RoundSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+8] x)
 (RoundSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+8] x)
-(RoundWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+0] x)
 (RoundWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+0] x)
 (RoundWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+0] x)
+(RoundWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+0] x)
 (RoundWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+0] x)
 (RoundWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+0] x)
 (RoundWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+0] x)
-(SaturatedAddInt16x16 ...) => (VPADDSW256 ...)
-(SaturatedAddInt16x32 ...) => (VPADDSW512 ...)
-(SaturatedAddInt16x8 ...) => (VPADDSW128 ...)
 (SaturatedAddInt8x16 ...) => (VPADDSB128 ...)
 (SaturatedAddInt8x32 ...) => (VPADDSB256 ...)
 (SaturatedAddInt8x64 ...) => (VPADDSB512 ...)
-(SaturatedAddUint16x16 ...) => (VPADDSW256 ...)
-(SaturatedAddUint16x32 ...) => (VPADDSW512 ...)
-(SaturatedAddUint16x8 ...) => (VPADDSW128 ...)
+(SaturatedAddInt16x8 ...) => (VPADDSW128 ...)
+(SaturatedAddInt16x16 ...) => (VPADDSW256 ...)
+(SaturatedAddInt16x32 ...) => (VPADDSW512 ...)
 (SaturatedAddUint8x16 ...) => (VPADDSB128 ...)
 (SaturatedAddUint8x32 ...) => (VPADDSB256 ...)
 (SaturatedAddUint8x64 ...) => (VPADDSB512 ...)
-(SaturatedPairDotProdAccumulateInt32x16 ...) => (VPDPWSSDS512 ...)
+(SaturatedAddUint16x8 ...) => (VPADDSW128 ...)
+(SaturatedAddUint16x16 ...) => (VPADDSW256 ...)
+(SaturatedAddUint16x32 ...) => (VPADDSW512 ...)
 (SaturatedPairDotProdAccumulateInt32x4 ...) => (VPDPWSSDS128 ...)
 (SaturatedPairDotProdAccumulateInt32x8 ...) => (VPDPWSSDS256 ...)
-(SaturatedPairwiseAddInt16x16 ...) => (VPHADDSW256 ...)
+(SaturatedPairDotProdAccumulateInt32x16 ...) => (VPDPWSSDS512 ...)
 (SaturatedPairwiseAddInt16x8 ...) => (VPHADDSW128 ...)
-(SaturatedPairwiseSubInt16x16 ...) => (VPHSUBSW256 ...)
+(SaturatedPairwiseAddInt16x16 ...) => (VPHADDSW256 ...)
 (SaturatedPairwiseSubInt16x8 ...) => (VPHSUBSW128 ...)
-(SaturatedSubInt16x16 ...) => (VPSUBSW256 ...)
-(SaturatedSubInt16x32 ...) => (VPSUBSW512 ...)
-(SaturatedSubInt16x8 ...) => (VPSUBSW128 ...)
+(SaturatedPairwiseSubInt16x16 ...) => (VPHSUBSW256 ...)
 (SaturatedSubInt8x16 ...) => (VPSUBSB128 ...)
 (SaturatedSubInt8x32 ...) => (VPSUBSB256 ...)
 (SaturatedSubInt8x64 ...) => (VPSUBSB512 ...)
-(SaturatedSubUint16x16 ...) => (VPSUBSW256 ...)
-(SaturatedSubUint16x32 ...) => (VPSUBSW512 ...)
-(SaturatedSubUint16x8 ...) => (VPSUBSW128 ...)
+(SaturatedSubInt16x8 ...) => (VPSUBSW128 ...)
+(SaturatedSubInt16x16 ...) => (VPSUBSW256 ...)
+(SaturatedSubInt16x32 ...) => (VPSUBSW512 ...)
 (SaturatedSubUint8x16 ...) => (VPSUBSB128 ...)
 (SaturatedSubUint8x32 ...) => (VPSUBSB256 ...)
 (SaturatedSubUint8x64 ...) => (VPSUBSB512 ...)
+(SaturatedSubUint16x8 ...) => (VPSUBSW128 ...)
+(SaturatedSubUint16x16 ...) => (VPSUBSW256 ...)
+(SaturatedSubUint16x32 ...) => (VPSUBSW512 ...)
 (SaturatedUnsignedSignedPairDotProdUint8x16 ...) => (VPMADDUBSW128 ...)
 (SaturatedUnsignedSignedPairDotProdUint8x32 ...) => (VPMADDUBSW256 ...)
 (SaturatedUnsignedSignedPairDotProdUint8x64 ...) => (VPMADDUBSW512 ...)
-(SaturatedUnsignedSignedQuadDotProdAccumulateInt32x16 ...) => (VPDPBUSDS512 ...)
 (SaturatedUnsignedSignedQuadDotProdAccumulateInt32x4 ...) => (VPDPBUSDS128 ...)
 (SaturatedUnsignedSignedQuadDotProdAccumulateInt32x8 ...) => (VPDPBUSDS256 ...)
-(SaturatedUnsignedSignedQuadDotProdAccumulateUint32x16 ...) => (VPDPBUSDS512 ...)
+(SaturatedUnsignedSignedQuadDotProdAccumulateInt32x16 ...) => (VPDPBUSDS512 ...)
 (SaturatedUnsignedSignedQuadDotProdAccumulateUint32x4 ...) => (VPDPBUSDS128 ...)
 (SaturatedUnsignedSignedQuadDotProdAccumulateUint32x8 ...) => (VPDPBUSDS256 ...)
+(SaturatedUnsignedSignedQuadDotProdAccumulateUint32x16 ...) => (VPDPBUSDS512 ...)
+(SetElemInt8x16 [a] x y) => (VPINSRB128 [a] x y)
 (SetElemInt16x8 [a] x y) => (VPINSRW128 [a] x y)
 (SetElemInt32x4 [a] x y) => (VPINSRD128 [a] x y)
 (SetElemInt64x2 [a] x y) => (VPINSRQ128 [a] x y)
-(SetElemInt8x16 [a] x y) => (VPINSRB128 [a] x y)
+(SetElemUint8x16 [a] x y) => (VPINSRB128 [a] x y)
 (SetElemUint16x8 [a] x y) => (VPINSRW128 [a] x y)
 (SetElemUint32x4 [a] x y) => (VPINSRD128 [a] x y)
 (SetElemUint64x2 [a] x y) => (VPINSRQ128 [a] x y)
-(SetElemUint8x16 [a] x y) => (VPINSRB128 [a] x y)
-(SignInt16x16 ...) => (VPSIGNW256 ...)
+(SignInt8x16 ...) => (VPSIGNB128 ...)
+(SignInt8x32 ...) => (VPSIGNB256 ...)
 (SignInt16x8 ...) => (VPSIGNW128 ...)
+(SignInt16x16 ...) => (VPSIGNW256 ...)
 (SignInt32x4 ...) => (VPSIGND128 ...)
 (SignInt32x8 ...) => (VPSIGND256 ...)
-(SignInt8x16 ...) => (VPSIGNB128 ...)
-(SignInt8x32 ...) => (VPSIGNB256 ...)
-(SqrtFloat32x16 ...) => (VSQRTPS512 ...)
 (SqrtFloat32x4 ...) => (VSQRTPS128 ...)
 (SqrtFloat32x8 ...) => (VSQRTPS256 ...)
+(SqrtFloat32x16 ...) => (VSQRTPS512 ...)
 (SqrtFloat64x2 ...) => (VSQRTPD128 ...)
 (SqrtFloat64x4 ...) => (VSQRTPD256 ...)
 (SqrtFloat64x8 ...) => (VSQRTPD512 ...)
-(SubFloat32x16 ...) => (VSUBPS512 ...)
 (SubFloat32x4 ...) => (VSUBPS128 ...)
 (SubFloat32x8 ...) => (VSUBPS256 ...)
+(SubFloat32x16 ...) => (VSUBPS512 ...)
 (SubFloat64x2 ...) => (VSUBPD128 ...)
 (SubFloat64x4 ...) => (VSUBPD256 ...)
 (SubFloat64x8 ...) => (VSUBPD512 ...)
+(SubInt8x16 ...) => (VPSUBB128 ...)
+(SubInt8x32 ...) => (VPSUBB256 ...)
+(SubInt8x64 ...) => (VPSUBB512 ...)
+(SubInt16x8 ...) => (VPSUBW128 ...)
 (SubInt16x16 ...) => (VPSUBW256 ...)
 (SubInt16x32 ...) => (VPSUBW512 ...)
-(SubInt16x8 ...) => (VPSUBW128 ...)
-(SubInt32x16 ...) => (VPSUBD512 ...)
 (SubInt32x4 ...) => (VPSUBD128 ...)
 (SubInt32x8 ...) => (VPSUBD256 ...)
+(SubInt32x16 ...) => (VPSUBD512 ...)
 (SubInt64x2 ...) => (VPSUBQ128 ...)
 (SubInt64x4 ...) => (VPSUBQ256 ...)
 (SubInt64x8 ...) => (VPSUBQ512 ...)
-(SubInt8x16 ...) => (VPSUBB128 ...)
-(SubInt8x32 ...) => (VPSUBB256 ...)
-(SubInt8x64 ...) => (VPSUBB512 ...)
+(SubUint8x16 ...) => (VPSUBB128 ...)
+(SubUint8x32 ...) => (VPSUBB256 ...)
+(SubUint8x64 ...) => (VPSUBB512 ...)
+(SubUint16x8 ...) => (VPSUBW128 ...)
 (SubUint16x16 ...) => (VPSUBW256 ...)
 (SubUint16x32 ...) => (VPSUBW512 ...)
-(SubUint16x8 ...) => (VPSUBW128 ...)
-(SubUint32x16 ...) => (VPSUBD512 ...)
 (SubUint32x4 ...) => (VPSUBD128 ...)
 (SubUint32x8 ...) => (VPSUBD256 ...)
+(SubUint32x16 ...) => (VPSUBD512 ...)
 (SubUint64x2 ...) => (VPSUBQ128 ...)
 (SubUint64x4 ...) => (VPSUBQ256 ...)
 (SubUint64x8 ...) => (VPSUBQ512 ...)
-(SubUint8x16 ...) => (VPSUBB128 ...)
-(SubUint8x32 ...) => (VPSUBB256 ...)
-(SubUint8x64 ...) => (VPSUBB512 ...)
 (TruncFloat32x4 x) => (VROUNDPS128 [3] x)
 (TruncFloat32x8 x) => (VROUNDPS256 [3] x)
 (TruncFloat64x2 x) => (VROUNDPD128 [3] x)
 (TruncFloat64x4 x) => (VROUNDPD256 [3] x)
-(TruncSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+11] x)
 (TruncSuppressExceptionWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+11] x)
 (TruncSuppressExceptionWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+11] x)
+(TruncSuppressExceptionWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+11] x)
 (TruncSuppressExceptionWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+11] x)
 (TruncSuppressExceptionWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+11] x)
 (TruncSuppressExceptionWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+11] x)
-(TruncWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+3] x)
 (TruncWithPrecisionFloat32x4 [a] x) => (VRNDSCALEPS128 [a+3] x)
 (TruncWithPrecisionFloat32x8 [a] x) => (VRNDSCALEPS256 [a+3] x)
+(TruncWithPrecisionFloat32x16 [a] x) => (VRNDSCALEPS512 [a+3] x)
 (TruncWithPrecisionFloat64x2 [a] x) => (VRNDSCALEPD128 [a+3] x)
 (TruncWithPrecisionFloat64x4 [a] x) => (VRNDSCALEPD256 [a+3] x)
 (TruncWithPrecisionFloat64x8 [a] x) => (VRNDSCALEPD512 [a+3] x)
-(UnsignedSignedQuadDotProdAccumulateInt32x16 ...) => (VPDPBUSD512 ...)
 (UnsignedSignedQuadDotProdAccumulateInt32x4 ...) => (VPDPBUSD128 ...)
 (UnsignedSignedQuadDotProdAccumulateInt32x8 ...) => (VPDPBUSD256 ...)
-(UnsignedSignedQuadDotProdAccumulateUint32x16 ...) => (VPDPBUSD512 ...)
+(UnsignedSignedQuadDotProdAccumulateInt32x16 ...) => (VPDPBUSD512 ...)
 (UnsignedSignedQuadDotProdAccumulateUint32x4 ...) => (VPDPBUSD128 ...)
 (UnsignedSignedQuadDotProdAccumulateUint32x8 ...) => (VPDPBUSD256 ...)
-(XorFloat32x16 ...) => (VXORPS512 ...)
+(UnsignedSignedQuadDotProdAccumulateUint32x16 ...) => (VPDPBUSD512 ...)
 (XorFloat32x4 ...) => (VXORPS128 ...)
 (XorFloat32x8 ...) => (VXORPS256 ...)
+(XorFloat32x16 ...) => (VXORPS512 ...)
 (XorFloat64x2 ...) => (VXORPD128 ...)
 (XorFloat64x4 ...) => (VXORPD256 ...)
 (XorFloat64x8 ...) => (VXORPD512 ...)
-(XorInt16x16 ...) => (VPXOR256 ...)
+(XorInt8x16 ...) => (VPXOR128 ...)
+(XorInt8x32 ...) => (VPXOR256 ...)
 (XorInt16x8 ...) => (VPXOR128 ...)
-(XorInt32x16 ...) => (VPXORD512 ...)
+(XorInt16x16 ...) => (VPXOR256 ...)
 (XorInt32x4 ...) => (VPXOR128 ...)
 (XorInt32x8 ...) => (VPXOR256 ...)
+(XorInt32x16 ...) => (VPXORD512 ...)
 (XorInt64x2 ...) => (VPXOR128 ...)
 (XorInt64x4 ...) => (VPXOR256 ...)
 (XorInt64x8 ...) => (VPXORQ512 ...)
-(XorInt8x16 ...) => (VPXOR128 ...)
-(XorInt8x32 ...) => (VPXOR256 ...)
-(XorUint16x16 ...) => (VPXOR256 ...)
+(XorUint8x16 ...) => (VPXOR128 ...)
+(XorUint8x32 ...) => (VPXOR256 ...)
 (XorUint16x8 ...) => (VPXOR128 ...)
-(XorUint32x16 ...) => (VPXORD512 ...)
+(XorUint16x16 ...) => (VPXOR256 ...)
 (XorUint32x4 ...) => (VPXOR128 ...)
 (XorUint32x8 ...) => (VPXOR256 ...)
+(XorUint32x16 ...) => (VPXORD512 ...)
 (XorUint64x2 ...) => (VPXOR128 ...)
 (XorUint64x4 ...) => (VPXOR256 ...)
 (XorUint64x8 ...) => (VPXORQ512 ...)
-(XorUint8x16 ...) => (VPXOR128 ...)
-(XorUint8x32 ...) => (VPXOR256 ...)