cmd/compile: add an optimization rule for math/bits.ReverseBytes16 on arm

This CL adds two rules to turn patterns like ((x<<8) | (x>>8)) (the type of
x is uint16, "|" can also be "+" or "^") to a REV16 instruction on arm v6+.
This optimization rule can be used for math/bits.ReverseBytes16.

Benchmarks on arm v6:
name               old time/op  new time/op  delta
ReverseBytes-32    2.86ns ± 0%  2.86ns ± 0%   ~     (all equal)
ReverseBytes16-32  2.86ns ± 0%  2.86ns ± 0%   ~     (all equal)
ReverseBytes32-32  1.29ns ± 0%  1.29ns ± 0%   ~     (all equal)
ReverseBytes64-32  1.43ns ± 0%  1.43ns ± 0%   ~     (all equal)

Change-Id: I819e633c9a9d308f8e476fb0c82d73fb73dd019f
Reviewed-on: https://go-review.googlesource.com/c/go/+/159019
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 files changed, 445 insertions, 210 deletions

diff --git a/src/cmd/compile/internal/arm/ssa.go b/src/cmd/compile/internal/arm/ssa.go
index 9a8fabf622..320fe98707 100644
--- a/src/cmd/compile/internal/arm/ssa.go
+++ b/src/cmd/compile/internal/arm/ssa.go
@@ -659,6 +659,7 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 	case ssa.OpARMMVN,
 		ssa.OpARMCLZ,
 		ssa.OpARMREV,
+		ssa.OpARMREV16,
 		ssa.OpARMRBIT,
 		ssa.OpARMSQRTD,
 		ssa.OpARMNEGF,
diff --git a/src/cmd/compile/internal/ssa/gen/ARM.rules b/src/cmd/compile/internal/ssa/gen/ARM.rules
index 8b0e82f154..db418b76a6 100644
--- a/src/cmd/compile/internal/ssa/gen/ARM.rules
+++ b/src/cmd/compile/internal/ssa/gen/ARM.rules
@@ -1216,6 +1216,12 @@
 ( ORshiftRL [c] (SLLconst x [32-c]) x) -> (SRRconst [   c] x)
 (XORshiftRL [c] (SLLconst x [32-c]) x) -> (SRRconst [   c] x)
 
+// ((x>>8) | (x<<8)) -> (REV16 x), the type of x is uint16, "|" can also be "^" or "+".
+// UBFX instruction is supported by ARMv6T2, ARMv7 and above versions, REV16 is supported by
+// ARMv6 and above versions. So for ARMv6, we need to match SLLconst, SRLconst and ORshiftLL.
+((ADDshiftLL|ORshiftLL|XORshiftLL) <typ.UInt16> [8] (BFXU <typ.UInt16> [armBFAuxInt(8, 8)] x) x) -> (REV16 x)
+((ADDshiftLL|ORshiftLL|XORshiftLL) <typ.UInt16> [8] (SRLconst <typ.UInt16> [24] (SLLconst [16] x)) x) && objabi.GOARM>=6 -> (REV16 x)
+
 // use indexed loads and stores
 (MOVWload [0] {sym} (ADD ptr idx) mem) && sym == nil && !config.nacl -> (MOVWloadidx ptr idx mem)
 (MOVWstore [0] {sym} (ADD ptr idx) val mem) && sym == nil && !config.nacl -> (MOVWstoreidx ptr idx val mem)
diff --git a/src/cmd/compile/internal/ssa/gen/ARM64.rules b/src/cmd/compile/internal/ssa/gen/ARM64.rules
index 133a893610..ca123d7375 100644
--- a/src/cmd/compile/internal/ssa/gen/ARM64.rules
+++ b/src/cmd/compile/internal/ssa/gen/ARM64.rules
@@ -1751,11 +1751,11 @@
 ( ORshiftRL [c] (SLLconst x [64-c]) x) -> (RORconst [   c] x)
 (XORshiftRL [c] (SLLconst x [64-c]) x) -> (RORconst [   c] x)
 
-(ADDshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+(ADDshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (RORWconst [32-c] x)
-( ORshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+( ORshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (RORWconst [32-c] x)
-(XORshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+(XORshiftLL <t> [c] (UBFX [bfc] x) x) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (RORWconst [32-c] x)
 (ADDshiftRL <t> [c] (SLLconst x [32-c]) (MOVWUreg x)) && c < 32 && t.Size() == 4 -> (RORWconst [c] x)
 ( ORshiftRL <t> [c] (SLLconst x [32-c]) (MOVWUreg x)) && c < 32 && t.Size() == 4 -> (RORWconst [c] x)
@@ -1794,18 +1794,18 @@
 	-> (RORW x y)
 
 // ((x>>8) | (x<<8)) -> (REV16W x), the type of x is uint16, "|" can also be "^" or "+".
-((ADDshiftLL|ORshiftLL|XORshiftLL) <typ.UInt16> [8] (UBFX <typ.UInt16> [arm64BFAuxInt(8, 8)] x) x) -> (REV16W x)
+((ADDshiftLL|ORshiftLL|XORshiftLL) <typ.UInt16> [8] (UBFX <typ.UInt16> [armBFAuxInt(8, 8)] x) x) -> (REV16W x)
 
 // Extract from reg pair
 (ADDshiftLL [c] (SRLconst x [64-c]) x2) -> (EXTRconst [64-c] x2 x)
 ( ORshiftLL [c] (SRLconst x [64-c]) x2) -> (EXTRconst [64-c] x2 x)
 (XORshiftLL [c] (SRLconst x [64-c]) x2) -> (EXTRconst [64-c] x2 x)
 
-(ADDshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+(ADDshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (EXTRWconst [32-c] x2 x)
-( ORshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+( ORshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (EXTRWconst [32-c] x2 x)
-(XORshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+(XORshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (EXTRWconst [32-c] x2 x)
 
 // Generic rules rewrite certain AND to a pair of shifts.
@@ -1821,88 +1821,88 @@
 
 // sbfiz
 // (x << lc) >> rc
-(SRAconst [rc] (SLLconst [lc] x)) && lc > rc -> (SBFIZ [arm64BFAuxInt(lc-rc, 64-lc)] x)
-(MOVWreg (SLLconst [lc] x)) && lc < 32 -> (SBFIZ [arm64BFAuxInt(lc, 32-lc)] x)
-(MOVHreg (SLLconst [lc] x)) && lc < 16 -> (SBFIZ [arm64BFAuxInt(lc, 16-lc)] x)
-(MOVBreg (SLLconst [lc] x)) && lc < 8 -> (SBFIZ [arm64BFAuxInt(lc, 8-lc)] x)
+(SRAconst [rc] (SLLconst [lc] x)) && lc > rc -> (SBFIZ [armBFAuxInt(lc-rc, 64-lc)] x)
+(MOVWreg (SLLconst [lc] x)) && lc < 32 -> (SBFIZ [armBFAuxInt(lc, 32-lc)] x)
+(MOVHreg (SLLconst [lc] x)) && lc < 16 -> (SBFIZ [armBFAuxInt(lc, 16-lc)] x)
+(MOVBreg (SLLconst [lc] x)) && lc < 8 -> (SBFIZ [armBFAuxInt(lc, 8-lc)] x)
 
 // sbfx
 // (x << lc) >> rc
-(SRAconst [rc] (SLLconst [lc] x)) && lc <= rc -> (SBFX [arm64BFAuxInt(rc-lc, 64-rc)] x)
-(SRAconst [rc] (MOVWreg x)) && rc < 32 -> (SBFX [arm64BFAuxInt(rc, 32-rc)] x)
-(SRAconst [rc] (MOVHreg x)) && rc < 16 -> (SBFX [arm64BFAuxInt(rc, 16-rc)] x)
-(SRAconst [rc] (MOVBreg x)) && rc < 8 -> (SBFX [arm64BFAuxInt(rc, 8-rc)] x)
+(SRAconst [rc] (SLLconst [lc] x)) && lc <= rc -> (SBFX [armBFAuxInt(rc-lc, 64-rc)] x)
+(SRAconst [rc] (MOVWreg x)) && rc < 32 -> (SBFX [armBFAuxInt(rc, 32-rc)] x)
+(SRAconst [rc] (MOVHreg x)) && rc < 16 -> (SBFX [armBFAuxInt(rc, 16-rc)] x)
+(SRAconst [rc] (MOVBreg x)) && rc < 8 -> (SBFX [armBFAuxInt(rc, 8-rc)] x)
 
 // sbfiz/sbfx combinations: merge shifts into bitfield ops
 (SRAconst [sc] (SBFIZ [bfc] x)) && sc < getARM64BFlsb(bfc)
-	-> (SBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	-> (SBFIZ [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 (SRAconst [sc] (SBFIZ [bfc] x)) && sc >= getARM64BFlsb(bfc)
 	&& sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	-> (SBFX [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	-> (SBFX [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 
 // ubfiz
 // (x & ac) << sc
 (SLLconst [sc] (ANDconst [ac] x)) && isARM64BFMask(sc, ac, 0)
-	-> (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(ac, 0))] x)
-(SLLconst [sc] (MOVWUreg x)) && isARM64BFMask(sc, 1<<32-1, 0) -> (UBFIZ [arm64BFAuxInt(sc, 32)] x)
-(SLLconst [sc] (MOVHUreg x)) && isARM64BFMask(sc, 1<<16-1, 0) -> (UBFIZ [arm64BFAuxInt(sc, 16)] x)
-(SLLconst [sc] (MOVBUreg x)) && isARM64BFMask(sc, 1<<8-1, 0) -> (UBFIZ [arm64BFAuxInt(sc, 8)] x)
+	-> (UBFIZ [armBFAuxInt(sc, arm64BFWidth(ac, 0))] x)
+(SLLconst [sc] (MOVWUreg x)) && isARM64BFMask(sc, 1<<32-1, 0) -> (UBFIZ [armBFAuxInt(sc, 32)] x)
+(SLLconst [sc] (MOVHUreg x)) && isARM64BFMask(sc, 1<<16-1, 0) -> (UBFIZ [armBFAuxInt(sc, 16)] x)
+(SLLconst [sc] (MOVBUreg x)) && isARM64BFMask(sc, 1<<8-1, 0) -> (UBFIZ [armBFAuxInt(sc, 8)] x)
 // (x << sc) & ac
 (ANDconst [ac] (SLLconst [sc] x)) && isARM64BFMask(sc, ac, sc)
-	-> (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(ac, sc))] x)
+	-> (UBFIZ [armBFAuxInt(sc, arm64BFWidth(ac, sc))] x)
 (MOVWUreg (SLLconst [sc] x)) && isARM64BFMask(sc, 1<<32-1, sc)
-	-> (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
+	-> (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
 (MOVHUreg (SLLconst [sc] x)) && isARM64BFMask(sc, 1<<16-1, sc)
-	-> (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
+	-> (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
 (MOVBUreg (SLLconst [sc] x)) && isARM64BFMask(sc, 1<<8-1, sc)
-	-> (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
+	-> (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
 // (x << lc) >> rc
-(SRLconst [rc] (SLLconst [lc] x)) && lc > rc -> (UBFIZ [arm64BFAuxInt(lc-rc, 64-lc)] x)
+(SRLconst [rc] (SLLconst [lc] x)) && lc > rc -> (UBFIZ [armBFAuxInt(lc-rc, 64-lc)] x)
 
 // ubfx
 // (x >> sc) & ac
 (ANDconst [ac] (SRLconst [sc] x)) && isARM64BFMask(sc, ac, 0)
-	-> (UBFX [arm64BFAuxInt(sc, arm64BFWidth(ac, 0))] x)
-(MOVWUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<32-1, 0) -> (UBFX [arm64BFAuxInt(sc, 32)] x)
-(MOVHUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<16-1, 0) -> (UBFX [arm64BFAuxInt(sc, 16)] x)
-(MOVBUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<8-1, 0) -> (UBFX [arm64BFAuxInt(sc, 8)] x)
+	-> (UBFX [armBFAuxInt(sc, arm64BFWidth(ac, 0))] x)
+(MOVWUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<32-1, 0) -> (UBFX [armBFAuxInt(sc, 32)] x)
+(MOVHUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<16-1, 0) -> (UBFX [armBFAuxInt(sc, 16)] x)
+(MOVBUreg (SRLconst [sc] x)) && isARM64BFMask(sc, 1<<8-1, 0) -> (UBFX [armBFAuxInt(sc, 8)] x)
 // (x & ac) >> sc
 (SRLconst [sc] (ANDconst [ac] x)) && isARM64BFMask(sc, ac, sc)
-	-> (UBFX [arm64BFAuxInt(sc, arm64BFWidth(ac, sc))] x)
+	-> (UBFX [armBFAuxInt(sc, arm64BFWidth(ac, sc))] x)
 (SRLconst [sc] (MOVWUreg x)) && isARM64BFMask(sc, 1<<32-1, sc)
-	-> (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
+	-> (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
 (SRLconst [sc] (MOVHUreg x)) && isARM64BFMask(sc, 1<<16-1, sc)
-	-> (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
+	-> (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
 (SRLconst [sc] (MOVBUreg x)) && isARM64BFMask(sc, 1<<8-1, sc)
-	-> (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
+	-> (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
 // (x << lc) >> rc
-(SRLconst [rc] (SLLconst [lc] x)) && lc < rc -> (UBFX [arm64BFAuxInt(rc-lc, 64-rc)] x)
+(SRLconst [rc] (SLLconst [lc] x)) && lc < rc -> (UBFX [armBFAuxInt(rc-lc, 64-rc)] x)
 
 // ubfiz/ubfx combinations: merge shifts into bitfield ops
 (SRLconst [sc] (UBFX [bfc] x)) && sc < getARM64BFwidth(bfc)
-	-> (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
+	-> (UBFX [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
 (UBFX [bfc] (SRLconst [sc] x)) && sc+getARM64BFwidth(bfc)+getARM64BFlsb(bfc) < 64
-	-> (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
+	-> (UBFX [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
 (SLLconst [sc] (UBFIZ [bfc] x)) && sc+getARM64BFwidth(bfc)+getARM64BFlsb(bfc) < 64
-	-> (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
+	-> (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
 (UBFIZ [bfc] (SLLconst [sc] x)) && sc < getARM64BFwidth(bfc)
-	-> (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
+	-> (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
 // ((x << c1) >> c2) >> c3
 (SRLconst [sc] (UBFIZ [bfc] x)) && sc == getARM64BFlsb(bfc)
 	-> (ANDconst [1<<uint(getARM64BFwidth(bfc))-1] x)
 (SRLconst [sc] (UBFIZ [bfc] x)) && sc < getARM64BFlsb(bfc)
-	-> (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	-> (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 (SRLconst [sc] (UBFIZ [bfc] x)) && sc > getARM64BFlsb(bfc)
 	&& sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	-> (UBFX [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	-> (UBFX [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 // ((x << c1) << c2) >> c3
 (UBFX [bfc] (SLLconst [sc] x)) && sc == getARM64BFlsb(bfc)
 	-> (ANDconst [1<<uint(getARM64BFwidth(bfc))-1] x)
 (UBFX [bfc] (SLLconst [sc] x)) && sc < getARM64BFlsb(bfc)
-	-> (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	-> (UBFX [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 (UBFX [bfc] (SLLconst [sc] x)) && sc > getARM64BFlsb(bfc)
 	&& sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	-> (UBFIZ [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	-> (UBFIZ [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 
 // bfi
 (OR (UBFIZ [bfc] x) (ANDconst [ac] y))
@@ -1910,7 +1910,7 @@
 	-> (BFI [bfc] y x)
 (ORshiftRL [rc] (ANDconst [ac] x) (SLLconst [lc] y))
 	&& lc > rc && ac == ^((1<<uint(64-lc)-1) << uint64(lc-rc))
-	-> (BFI [arm64BFAuxInt(lc-rc, 64-lc)] x y)
+	-> (BFI [armBFAuxInt(lc-rc, 64-lc)] x y)
 // bfxil
 (OR (UBFX [bfc] x) (ANDconst [ac] y)) && ac == ^(1<<uint(getARM64BFwidth(bfc))-1)
 	-> (BFXIL [bfc] y x)
@@ -2560,23 +2560,23 @@
 	&& x.Uses == 1
 	&& clobber(x)
 	-> (MOVHstoreidx ptr idx w mem)
-(MOVBstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(8, 8)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
+(MOVBstore [i] {s} ptr0 (UBFX [armBFAuxInt(8, 8)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
 	&& x.Uses == 1
 	&& isSamePtr(ptr0, ptr1)
 	&& clobber(x)
 	-> (MOVHstore [i-1] {s} ptr0 w mem)
-(MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(8, 8)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
+(MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(8, 8)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
 	&& clobber(x)
 	-> (MOVHstoreidx ptr1 idx1 w mem)
-(MOVBstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(8, 24)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
+(MOVBstore [i] {s} ptr0 (UBFX [armBFAuxInt(8, 24)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
 	&& x.Uses == 1
 	&& isSamePtr(ptr0, ptr1)
 	&& clobber(x)
 	-> (MOVHstore [i-1] {s} ptr0 w mem)
-(MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(8, 24)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
+(MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(8, 24)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
@@ -2653,18 +2653,18 @@
 	&& isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1)
 	&& clobber(x)
 	-> (MOVWstoreidx ptr1 (SLLconst <idx1.Type> [1] idx1) w mem)
-(MOVHstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstore [i-2] {s} ptr1 w mem))
+(MOVHstore [i] {s} ptr0 (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstore [i-2] {s} ptr1 w mem))
 	&& x.Uses == 1
 	&& isSamePtr(ptr0, ptr1)
 	&& clobber(x)
 	-> (MOVWstore [i-2] {s} ptr0 w mem)
-(MOVHstore [2] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstoreidx ptr1 idx1 w mem))
+(MOVHstore [2] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstoreidx ptr1 idx1 w mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
 	&& clobber(x)
 	-> (MOVWstoreidx ptr1 idx1 w mem)
-(MOVHstore [2] {s} (ADDshiftLL [1] ptr0 idx0) (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstoreidx2 ptr1 idx1 w mem))
+(MOVHstore [2] {s} (ADDshiftLL [1] ptr0 idx0) (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstoreidx2 ptr1 idx1 w mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1)
@@ -2792,9 +2792,9 @@
 	&& clobber(x6)
 	-> (MOVDstoreidx ptr0 idx0 (REV <w.Type> w) mem)
 (MOVBstore [i] {s} ptr w
-	x0:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 24)] w)
-	x1:(MOVBstore [i-2] {s} ptr (UBFX [arm64BFAuxInt(16, 16)] w)
-	x2:(MOVBstore [i-3] {s} ptr (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	x0:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 24)] w)
+	x1:(MOVBstore [i-2] {s} ptr (UBFX [armBFAuxInt(16, 16)] w)
+	x2:(MOVBstore [i-3] {s} ptr (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	&& x0.Uses == 1
 	&& x1.Uses == 1
 	&& x2.Uses == 1
@@ -2803,9 +2803,9 @@
 	&& clobber(x2)
 	-> (MOVWstore [i-3] {s} ptr (REVW <w.Type> w) mem)
 (MOVBstore [3] {s} p w
-	x0:(MOVBstore [2] {s} p (UBFX [arm64BFAuxInt(8, 24)] w)
-	x1:(MOVBstore [1] {s} p1:(ADD ptr1 idx1) (UBFX [arm64BFAuxInt(16, 16)] w)
-	x2:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	x0:(MOVBstore [2] {s} p (UBFX [armBFAuxInt(8, 24)] w)
+	x1:(MOVBstore [1] {s} p1:(ADD ptr1 idx1) (UBFX [armBFAuxInt(16, 16)] w)
+	x2:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	&& x0.Uses == 1
 	&& x1.Uses == 1
 	&& x2.Uses == 1
@@ -2817,9 +2817,9 @@
 	&& clobber(x2)
 	-> (MOVWstoreidx ptr0 idx0 (REVW <w.Type> w) mem)
 (MOVBstoreidx ptr (ADDconst [3] idx) w
-	x0:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [arm64BFAuxInt(8, 24)] w)
-	x1:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(16, 16)] w)
-	x2:(MOVBstoreidx ptr idx (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	x0:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [armBFAuxInt(8, 24)] w)
+	x1:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(16, 16)] w)
+	x2:(MOVBstoreidx ptr idx (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	&& x0.Uses == 1
 	&& x1.Uses == 1
 	&& x2.Uses == 1
@@ -2828,9 +2828,9 @@
 	&& clobber(x2)
 	-> (MOVWstoreidx ptr idx (REVW <w.Type> w) mem)
 (MOVBstoreidx ptr idx w
-	x0:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(8, 24)] w)
-	x1:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [arm64BFAuxInt(16, 16)] w)
-	x2:(MOVBstoreidx ptr (ADDconst [3] idx) (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	x0:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(8, 24)] w)
+	x1:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [armBFAuxInt(16, 16)] w)
+	x2:(MOVBstoreidx ptr (ADDconst [3] idx) (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	&& x0.Uses == 1
 	&& x1.Uses == 1
 	&& x2.Uses == 1
@@ -2898,21 +2898,21 @@
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
 	&& clobber(x)
 	-> (MOVHstoreidx ptr0 idx0 (REV16W <w.Type> w) mem)
-(MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+(MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 8)] w) mem))
 	&& x.Uses == 1
 	&& clobber(x)
 	-> (MOVHstore [i-1] {s} ptr (REV16W <w.Type> w) mem)
-(MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+(MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(8, 8)] w) mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
 	&& clobber(x)
 	-> (MOVHstoreidx ptr0 idx0 (REV16W <w.Type> w) mem)
-(MOVBstoreidx ptr (ADDconst [1] idx) w x:(MOVBstoreidx ptr idx (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+(MOVBstoreidx ptr (ADDconst [1] idx) w x:(MOVBstoreidx ptr idx (UBFX [armBFAuxInt(8, 8)] w) mem))
 	&& x.Uses == 1
 	&& clobber(x)
 	-> (MOVHstoreidx ptr idx (REV16W <w.Type> w) mem)
-(MOVBstoreidx ptr idx w x:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+(MOVBstoreidx ptr idx w x:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(8, 8)] w) mem))
 	&& x.Uses == 1
 	&& clobber(x)
 	-> (MOVHstoreidx ptr idx w mem)
@@ -2926,11 +2926,11 @@
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
 	&& clobber(x)
 	-> (MOVHstoreidx ptr0 idx0 (REV16W <w.Type> w) mem)
-(MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 24)] w) mem))
+(MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 24)] w) mem))
 	&& x.Uses == 1
 	&& clobber(x)
 	-> (MOVHstore [i-1] {s} ptr (REV16W <w.Type> w) mem)
-(MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(8, 24)] w) mem))
+(MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(8, 24)] w) mem))
 	&& x.Uses == 1
 	&& s == nil
 	&& (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1))
diff --git a/src/cmd/compile/internal/ssa/gen/ARMOps.go b/src/cmd/compile/internal/ssa/gen/ARMOps.go
index 4e2b0c5a5d..86d7e5f8ec 100644
--- a/src/cmd/compile/internal/ssa/gen/ARMOps.go
+++ b/src/cmd/compile/internal/ssa/gen/ARMOps.go
@@ -207,9 +207,10 @@ func init() {
 		{name: "NEGD", argLength: 1, reg: fp11, asm: "NEGD"},   // -arg0, float64
 		{name: "SQRTD", argLength: 1, reg: fp11, asm: "SQRTD"}, // sqrt(arg0), float64
 
-		{name: "CLZ", argLength: 1, reg: gp11, asm: "CLZ"},   // count leading zero
-		{name: "REV", argLength: 1, reg: gp11, asm: "REV"},   // reverse byte order
-		{name: "RBIT", argLength: 1, reg: gp11, asm: "RBIT"}, // reverse bit order
+		{name: "CLZ", argLength: 1, reg: gp11, asm: "CLZ"},     // count leading zero
+		{name: "REV", argLength: 1, reg: gp11, asm: "REV"},     // reverse byte order
+		{name: "REV16", argLength: 1, reg: gp11, asm: "REV16"}, // reverse byte order in 16-bit halfwords
+		{name: "RBIT", argLength: 1, reg: gp11, asm: "RBIT"},   // reverse bit order
 
 		// shifts
 		{name: "SLL", argLength: 2, reg: gp21, asm: "SLL"},                    // arg0 << arg1, shift amount is mod 256
diff --git a/src/cmd/compile/internal/ssa/opGen.go b/src/cmd/compile/internal/ssa/opGen.go
index 948bbdc32a..5fcc64f460 100644
--- a/src/cmd/compile/internal/ssa/opGen.go
+++ b/src/cmd/compile/internal/ssa/opGen.go
@@ -907,6 +907,7 @@ const (
 	OpARMSQRTD
 	OpARMCLZ
 	OpARMREV
+	OpARMREV16
 	OpARMRBIT
 	OpARMSLL
 	OpARMSLLconst
@@ -12037,6 +12038,19 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
+		name:   "REV16",
+		argLen: 1,
+		asm:    arm.AREV16,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 22527}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 g R12 R14
+			},
+			outputs: []outputInfo{
+				{0, 21503}, // R0 R1 R2 R3 R4 R5 R6 R7 R8 R9 R12 R14
+			},
+		},
+	},
+	{
 		name:   "RBIT",
 		argLen: 1,
 		asm:    arm.ARBIT,
diff --git a/src/cmd/compile/internal/ssa/rewrite.go b/src/cmd/compile/internal/ssa/rewrite.go
index 9c9de750b2..dbbb33c171 100644
--- a/src/cmd/compile/internal/ssa/rewrite.go
+++ b/src/cmd/compile/internal/ssa/rewrite.go
@@ -1037,13 +1037,13 @@ func isInlinableMemmove(dst, src *Value, sz int64, c *Config) bool {
 	return false
 }
 
-// encodes the lsb and width for arm64 bitfield ops into the expected auxInt format.
-func arm64BFAuxInt(lsb, width int64) int64 {
+// encodes the lsb and width for arm(64) bitfield ops into the expected auxInt format.
+func armBFAuxInt(lsb, width int64) int64 {
 	if lsb < 0 || lsb > 63 {
-		panic("ARM64 bit field lsb constant out of range")
+		panic("ARM(64) bit field lsb constant out of range")
 	}
 	if width < 1 || width > 64 {
-		panic("ARM64 bit field width constant out of range")
+		panic("ARM(64) bit field width constant out of range")
 	}
 	return width | lsb<<8
 }
diff --git a/src/cmd/compile/internal/ssa/rewriteARM.go b/src/cmd/compile/internal/ssa/rewriteARM.go
index 4fc7fdfbe1..c190ef779c 100644
--- a/src/cmd/compile/internal/ssa/rewriteARM.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM.go
@@ -2933,6 +2933,8 @@ func rewriteValueARM_OpARMADDconst_0(v *Value) bool {
 func rewriteValueARM_OpARMADDshiftLL_0(v *Value) bool {
 	b := v.Block
 	_ = b
+	typ := &b.Func.Config.Types
+	_ = typ
 	// match: (ADDshiftLL (MOVWconst [c]) x [d])
 	// cond:
 	// result: (ADDconst [c] (SLLconst <x.Type> x [d]))
@@ -2992,6 +2994,74 @@ func rewriteValueARM_OpARMADDshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (ADDshiftLL <typ.UInt16> [8] (BFXU <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
+	// cond:
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMBFXU {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
+			break
+		}
+		x := v_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
+	// match: (ADDshiftLL <typ.UInt16> [8] (SRLconst <typ.UInt16> [24] (SLLconst [16] x)) x)
+	// cond: objabi.GOARM>=6
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMSRLconst {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != 24 {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpARMSLLconst {
+			break
+		}
+		if v_0_0.AuxInt != 16 {
+			break
+		}
+		x := v_0_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		if !(objabi.GOARM >= 6) {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM_OpARMADDshiftLLreg_0(v *Value) bool {
@@ -11952,6 +12022,8 @@ func rewriteValueARM_OpARMORconst_0(v *Value) bool {
 func rewriteValueARM_OpARMORshiftLL_0(v *Value) bool {
 	b := v.Block
 	_ = b
+	typ := &b.Func.Config.Types
+	_ = typ
 	// match: (ORshiftLL (MOVWconst [c]) x [d])
 	// cond:
 	// result: (ORconst [c] (SLLconst <x.Type> x [d]))
@@ -12011,6 +12083,74 @@ func rewriteValueARM_OpARMORshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (ORshiftLL <typ.UInt16> [8] (BFXU <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
+	// cond:
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMBFXU {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
+			break
+		}
+		x := v_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
+	// match: (ORshiftLL <typ.UInt16> [8] (SRLconst <typ.UInt16> [24] (SLLconst [16] x)) x)
+	// cond: objabi.GOARM>=6
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMSRLconst {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != 24 {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpARMSLLconst {
+			break
+		}
+		if v_0_0.AuxInt != 16 {
+			break
+		}
+		x := v_0_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		if !(objabi.GOARM >= 6) {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
 	// match: (ORshiftLL x y:(SLLconst x [c]) [d])
 	// cond: c==d
 	// result: y
@@ -17230,6 +17370,8 @@ func rewriteValueARM_OpARMXORconst_0(v *Value) bool {
 func rewriteValueARM_OpARMXORshiftLL_0(v *Value) bool {
 	b := v.Block
 	_ = b
+	typ := &b.Func.Config.Types
+	_ = typ
 	// match: (XORshiftLL (MOVWconst [c]) x [d])
 	// cond:
 	// result: (XORconst [c] (SLLconst <x.Type> x [d]))
@@ -17289,6 +17431,74 @@ func rewriteValueARM_OpARMXORshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (XORshiftLL <typ.UInt16> [8] (BFXU <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
+	// cond:
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMBFXU {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
+			break
+		}
+		x := v_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
+	// match: (XORshiftLL <typ.UInt16> [8] (SRLconst <typ.UInt16> [24] (SLLconst [16] x)) x)
+	// cond: objabi.GOARM>=6
+	// result: (REV16 x)
+	for {
+		if v.Type != typ.UInt16 {
+			break
+		}
+		if v.AuxInt != 8 {
+			break
+		}
+		_ = v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpARMSRLconst {
+			break
+		}
+		if v_0.Type != typ.UInt16 {
+			break
+		}
+		if v_0.AuxInt != 24 {
+			break
+		}
+		v_0_0 := v_0.Args[0]
+		if v_0_0.Op != OpARMSLLconst {
+			break
+		}
+		if v_0_0.AuxInt != 16 {
+			break
+		}
+		x := v_0_0.Args[0]
+		if x != v.Args[1] {
+			break
+		}
+		if !(objabi.GOARM >= 6) {
+			break
+		}
+		v.reset(OpARMREV16)
+		v.AddArg(x)
+		return true
+	}
 	// match: (XORshiftLL x (SLLconst x [c]) [d])
 	// cond: c==d
 	// result: (MOVWconst [0])
diff --git a/src/cmd/compile/internal/ssa/rewriteARM64.go b/src/cmd/compile/internal/ssa/rewriteARM64.go
index 45801a4003..25246ce5e5 100644
--- a/src/cmd/compile/internal/ssa/rewriteARM64.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM64.go
@@ -2366,7 +2366,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (ADDshiftLL <t> [c] (UBFX [bfc] x) x)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (RORWconst [32-c] x)
 	for {
 		t := v.Type
@@ -2381,7 +2381,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		if x != v.Args[1] {
 			break
 		}
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64RORWconst)
@@ -2389,7 +2389,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
-	// match: (ADDshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [arm64BFAuxInt(8, 8)] x) x)
+	// match: (ADDshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
 	// cond:
 	// result: (REV16W x)
 	for {
@@ -2407,7 +2407,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		if v_0.Type != typ.UInt16 {
 			break
 		}
-		if v_0.AuxInt != arm64BFAuxInt(8, 8) {
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		x := v_0.Args[0]
@@ -2440,7 +2440,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (ADDshiftLL <t> [c] (UBFX [bfc] x) x2)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (EXTRWconst [32-c] x2 x)
 	for {
 		t := v.Type
@@ -2453,7 +2453,7 @@ func rewriteValueARM64_OpARM64ADDshiftLL_0(v *Value) bool {
 		bfc := v_0.AuxInt
 		x := v_0.Args[0]
 		x2 := v.Args[1]
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64EXTRWconst)
@@ -2912,7 +2912,7 @@ func rewriteValueARM64_OpARM64ANDconst_0(v *Value) bool {
 	}
 	// match: (ANDconst [ac] (SLLconst [sc] x))
 	// cond: isARM64BFMask(sc, ac, sc)
-	// result: (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(ac, sc))] x)
+	// result: (UBFIZ [armBFAuxInt(sc, arm64BFWidth(ac, sc))] x)
 	for {
 		ac := v.AuxInt
 		v_0 := v.Args[0]
@@ -2925,13 +2925,13 @@ func rewriteValueARM64_OpARM64ANDconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(ac, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(ac, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (ANDconst [ac] (SRLconst [sc] x))
 	// cond: isARM64BFMask(sc, ac, 0)
-	// result: (UBFX [arm64BFAuxInt(sc, arm64BFWidth(ac, 0))] x)
+	// result: (UBFX [armBFAuxInt(sc, arm64BFWidth(ac, 0))] x)
 	for {
 		ac := v.AuxInt
 		v_0 := v.Args[0]
@@ -2944,7 +2944,7 @@ func rewriteValueARM64_OpARM64ANDconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(ac, 0))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(ac, 0))
 		v.AddArg(x)
 		return true
 	}
@@ -9130,7 +9130,7 @@ func rewriteValueARM64_OpARM64MOVBUreg_0(v *Value) bool {
 	}
 	// match: (MOVBUreg (SLLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<8-1, sc)
-	// result: (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
+	// result: (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -9142,13 +9142,13 @@ func rewriteValueARM64_OpARM64MOVBUreg_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (MOVBUreg (SRLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<8-1, 0)
-	// result: (UBFX [arm64BFAuxInt(sc, 8)] x)
+	// result: (UBFX [armBFAuxInt(sc, 8)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SRLconst {
@@ -9160,7 +9160,7 @@ func rewriteValueARM64_OpARM64MOVBUreg_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, 8)
+		v.AuxInt = armBFAuxInt(sc, 8)
 		v.AddArg(x)
 		return true
 	}
@@ -9383,7 +9383,7 @@ func rewriteValueARM64_OpARM64MOVBreg_0(v *Value) bool {
 	}
 	// match: (MOVBreg (SLLconst [lc] x))
 	// cond: lc < 8
-	// result: (SBFIZ [arm64BFAuxInt(lc, 8-lc)] x)
+	// result: (SBFIZ [armBFAuxInt(lc, 8-lc)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -9395,7 +9395,7 @@ func rewriteValueARM64_OpARM64MOVBreg_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFIZ)
-		v.AuxInt = arm64BFAuxInt(lc, 8-lc)
+		v.AuxInt = armBFAuxInt(lc, 8-lc)
 		v.AddArg(x)
 		return true
 	}
@@ -9731,7 +9731,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(8, 8)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
+	// match: (MOVBstore [i] {s} ptr0 (UBFX [armBFAuxInt(8, 8)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
 	// cond: x.Uses == 1 && isSamePtr(ptr0, ptr1) && clobber(x)
 	// result: (MOVHstore [i-1] {s} ptr0 w mem)
 	for {
@@ -9743,7 +9743,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(8, 8) {
+		if v_1.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		w := v_1.Args[0]
@@ -9774,7 +9774,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(8, 8)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
+	// match: (MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(8, 8)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
 	// cond: x.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && clobber(x)
 	// result: (MOVHstoreidx ptr1 idx1 w mem)
 	for {
@@ -9794,7 +9794,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(8, 8) {
+		if v_1.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		w := v_1.Args[0]
@@ -9819,7 +9819,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(8, 24)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
+	// match: (MOVBstore [i] {s} ptr0 (UBFX [armBFAuxInt(8, 24)] w) x:(MOVBstore [i-1] {s} ptr1 w mem))
 	// cond: x.Uses == 1 && isSamePtr(ptr0, ptr1) && clobber(x)
 	// result: (MOVHstore [i-1] {s} ptr0 w mem)
 	for {
@@ -9831,7 +9831,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(8, 24) {
+		if v_1.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		w := v_1.Args[0]
@@ -9862,7 +9862,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(8, 24)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
+	// match: (MOVBstore [1] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(8, 24)] w) x:(MOVBstoreidx ptr1 idx1 w mem))
 	// cond: x.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && clobber(x)
 	// result: (MOVHstoreidx ptr1 idx1 w mem)
 	for {
@@ -9882,7 +9882,7 @@ func rewriteValueARM64_OpARM64MOVBstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(8, 24) {
+		if v_1.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		w := v_1.Args[0]
@@ -10694,7 +10694,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [i] {s} ptr w x0:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 24)] w) x1:(MOVBstore [i-2] {s} ptr (UBFX [arm64BFAuxInt(16, 16)] w) x2:(MOVBstore [i-3] {s} ptr (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	// match: (MOVBstore [i] {s} ptr w x0:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 24)] w) x1:(MOVBstore [i-2] {s} ptr (UBFX [armBFAuxInt(16, 16)] w) x2:(MOVBstore [i-3] {s} ptr (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	// cond: x0.Uses == 1 && x1.Uses == 1 && x2.Uses == 1 && clobber(x0) && clobber(x1) && clobber(x2)
 	// result: (MOVWstore [i-3] {s} ptr (REVW <w.Type> w) mem)
 	for {
@@ -10721,7 +10721,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x0_1.Op != OpARM64UBFX {
 			break
 		}
-		if x0_1.AuxInt != arm64BFAuxInt(8, 24) {
+		if x0_1.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x0_1.Args[0] {
@@ -10745,7 +10745,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x1_1.Op != OpARM64UBFX {
 			break
 		}
-		if x1_1.AuxInt != arm64BFAuxInt(16, 16) {
+		if x1_1.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		if w != x1_1.Args[0] {
@@ -10769,7 +10769,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x2_1.Op != OpARM64UBFX {
 			break
 		}
-		if x2_1.AuxInt != arm64BFAuxInt(24, 8) {
+		if x2_1.AuxInt != armBFAuxInt(24, 8) {
 			break
 		}
 		if w != x2_1.Args[0] {
@@ -10789,7 +10789,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [3] {s} p w x0:(MOVBstore [2] {s} p (UBFX [arm64BFAuxInt(8, 24)] w) x1:(MOVBstore [1] {s} p1:(ADD ptr1 idx1) (UBFX [arm64BFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	// match: (MOVBstore [3] {s} p w x0:(MOVBstore [2] {s} p (UBFX [armBFAuxInt(8, 24)] w) x1:(MOVBstore [1] {s} p1:(ADD ptr1 idx1) (UBFX [armBFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	// cond: x0.Uses == 1 && x1.Uses == 1 && x2.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && isSamePtr(p1, p) && clobber(x0) && clobber(x1) && clobber(x2)
 	// result: (MOVWstoreidx ptr0 idx0 (REVW <w.Type> w) mem)
 	for {
@@ -10818,7 +10818,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x0_1.Op != OpARM64UBFX {
 			break
 		}
-		if x0_1.AuxInt != arm64BFAuxInt(8, 24) {
+		if x0_1.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x0_1.Args[0] {
@@ -10846,7 +10846,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x1_1.Op != OpARM64UBFX {
 			break
 		}
-		if x1_1.AuxInt != arm64BFAuxInt(16, 16) {
+		if x1_1.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		if w != x1_1.Args[0] {
@@ -10863,7 +10863,7 @@ func rewriteValueARM64_OpARM64MOVBstore_20(v *Value) bool {
 		if x2_2.Op != OpARM64UBFX {
 			break
 		}
-		if x2_2.AuxInt != arm64BFAuxInt(24, 8) {
+		if x2_2.AuxInt != armBFAuxInt(24, 8) {
 			break
 		}
 		if w != x2_2.Args[0] {
@@ -11381,7 +11381,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+	// match: (MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 8)] w) mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: (MOVHstore [i-1] {s} ptr (REV16W <w.Type> w) mem)
 	for {
@@ -11408,7 +11408,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		if x_1.Op != OpARM64UBFX {
 			break
 		}
-		if x_1.AuxInt != arm64BFAuxInt(8, 8) {
+		if x_1.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		if w != x_1.Args[0] {
@@ -11428,7 +11428,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+	// match: (MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(8, 8)] w) mem))
 	// cond: x.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && clobber(x)
 	// result: (MOVHstoreidx ptr0 idx0 (REV16W <w.Type> w) mem)
 	for {
@@ -11456,7 +11456,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		if x_2.Op != OpARM64UBFX {
 			break
 		}
-		if x_2.AuxInt != arm64BFAuxInt(8, 8) {
+		if x_2.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		if w != x_2.Args[0] {
@@ -11577,7 +11577,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [arm64BFAuxInt(8, 24)] w) mem))
+	// match: (MOVBstore [i] {s} ptr w x:(MOVBstore [i-1] {s} ptr (UBFX [armBFAuxInt(8, 24)] w) mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: (MOVHstore [i-1] {s} ptr (REV16W <w.Type> w) mem)
 	for {
@@ -11604,7 +11604,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		if x_1.Op != OpARM64UBFX {
 			break
 		}
-		if x_1.AuxInt != arm64BFAuxInt(8, 24) {
+		if x_1.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x_1.Args[0] {
@@ -11624,7 +11624,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [arm64BFAuxInt(8, 24)] w) mem))
+	// match: (MOVBstore [1] {s} (ADD ptr1 idx1) w x:(MOVBstoreidx ptr0 idx0 (UBFX [armBFAuxInt(8, 24)] w) mem))
 	// cond: x.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && clobber(x)
 	// result: (MOVHstoreidx ptr0 idx0 (REV16W <w.Type> w) mem)
 	for {
@@ -11652,7 +11652,7 @@ func rewriteValueARM64_OpARM64MOVBstore_30(v *Value) bool {
 		if x_2.Op != OpARM64UBFX {
 			break
 		}
-		if x_2.AuxInt != arm64BFAuxInt(8, 24) {
+		if x_2.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x_2.Args[0] {
@@ -12014,7 +12014,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_0(v *Value) bool {
 func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 	b := v.Block
 	_ = b
-	// match: (MOVBstoreidx ptr (ADDconst [3] idx) w x0:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [arm64BFAuxInt(8, 24)] w) x1:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr idx (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	// match: (MOVBstoreidx ptr (ADDconst [3] idx) w x0:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [armBFAuxInt(8, 24)] w) x1:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr idx (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	// cond: x0.Uses == 1 && x1.Uses == 1 && x2.Uses == 1 && clobber(x0) && clobber(x1) && clobber(x2)
 	// result: (MOVWstoreidx ptr idx (REVW <w.Type> w) mem)
 	for {
@@ -12051,7 +12051,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x0_2.Op != OpARM64UBFX {
 			break
 		}
-		if x0_2.AuxInt != arm64BFAuxInt(8, 24) {
+		if x0_2.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x0_2.Args[0] {
@@ -12079,7 +12079,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x1_2.Op != OpARM64UBFX {
 			break
 		}
-		if x1_2.AuxInt != arm64BFAuxInt(16, 16) {
+		if x1_2.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		if w != x1_2.Args[0] {
@@ -12100,7 +12100,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x2_2.Op != OpARM64UBFX {
 			break
 		}
-		if x2_2.AuxInt != arm64BFAuxInt(24, 8) {
+		if x2_2.AuxInt != armBFAuxInt(24, 8) {
 			break
 		}
 		if w != x2_2.Args[0] {
@@ -12119,7 +12119,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstoreidx ptr idx w x0:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(8, 24)] w) x1:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [arm64BFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr (ADDconst [3] idx) (UBFX [arm64BFAuxInt(24, 8)] w) mem))))
+	// match: (MOVBstoreidx ptr idx w x0:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(8, 24)] w) x1:(MOVBstoreidx ptr (ADDconst [2] idx) (UBFX [armBFAuxInt(16, 16)] w) x2:(MOVBstoreidx ptr (ADDconst [3] idx) (UBFX [armBFAuxInt(24, 8)] w) mem))))
 	// cond: x0.Uses == 1 && x1.Uses == 1 && x2.Uses == 1 && clobber(x0) && clobber(x1) && clobber(x2)
 	// result: (MOVWstoreidx ptr idx w mem)
 	for {
@@ -12149,7 +12149,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x0_2.Op != OpARM64UBFX {
 			break
 		}
-		if x0_2.AuxInt != arm64BFAuxInt(8, 24) {
+		if x0_2.AuxInt != armBFAuxInt(8, 24) {
 			break
 		}
 		if w != x0_2.Args[0] {
@@ -12177,7 +12177,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x1_2.Op != OpARM64UBFX {
 			break
 		}
-		if x1_2.AuxInt != arm64BFAuxInt(16, 16) {
+		if x1_2.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		if w != x1_2.Args[0] {
@@ -12205,7 +12205,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x2_2.Op != OpARM64UBFX {
 			break
 		}
-		if x2_2.AuxInt != arm64BFAuxInt(24, 8) {
+		if x2_2.AuxInt != armBFAuxInt(24, 8) {
 			break
 		}
 		if w != x2_2.Args[0] {
@@ -12222,7 +12222,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstoreidx ptr (ADDconst [1] idx) w x:(MOVBstoreidx ptr idx (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+	// match: (MOVBstoreidx ptr (ADDconst [1] idx) w x:(MOVBstoreidx ptr idx (UBFX [armBFAuxInt(8, 8)] w) mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: (MOVHstoreidx ptr idx (REV16W <w.Type> w) mem)
 	for {
@@ -12252,7 +12252,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x_2.Op != OpARM64UBFX {
 			break
 		}
-		if x_2.AuxInt != arm64BFAuxInt(8, 8) {
+		if x_2.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		if w != x_2.Args[0] {
@@ -12271,7 +12271,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVBstoreidx ptr idx w x:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [arm64BFAuxInt(8, 8)] w) mem))
+	// match: (MOVBstoreidx ptr idx w x:(MOVBstoreidx ptr (ADDconst [1] idx) (UBFX [armBFAuxInt(8, 8)] w) mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: (MOVHstoreidx ptr idx w mem)
 	for {
@@ -12301,7 +12301,7 @@ func rewriteValueARM64_OpARM64MOVBstoreidx_10(v *Value) bool {
 		if x_2.Op != OpARM64UBFX {
 			break
 		}
-		if x_2.AuxInt != arm64BFAuxInt(8, 8) {
+		if x_2.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		if w != x_2.Args[0] {
@@ -13941,7 +13941,7 @@ func rewriteValueARM64_OpARM64MOVHUreg_0(v *Value) bool {
 	}
 	// match: (MOVHUreg (SLLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<16-1, sc)
-	// result: (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
+	// result: (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -13953,7 +13953,7 @@ func rewriteValueARM64_OpARM64MOVHUreg_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))
 		v.AddArg(x)
 		return true
 	}
@@ -13962,7 +13962,7 @@ func rewriteValueARM64_OpARM64MOVHUreg_0(v *Value) bool {
 func rewriteValueARM64_OpARM64MOVHUreg_10(v *Value) bool {
 	// match: (MOVHUreg (SRLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<16-1, 0)
-	// result: (UBFX [arm64BFAuxInt(sc, 16)] x)
+	// result: (UBFX [armBFAuxInt(sc, 16)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SRLconst {
@@ -13974,7 +13974,7 @@ func rewriteValueARM64_OpARM64MOVHUreg_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, 16)
+		v.AuxInt = armBFAuxInt(sc, 16)
 		v.AddArg(x)
 		return true
 	}
@@ -14423,7 +14423,7 @@ func rewriteValueARM64_OpARM64MOVHreg_10(v *Value) bool {
 	}
 	// match: (MOVHreg (SLLconst [lc] x))
 	// cond: lc < 16
-	// result: (SBFIZ [arm64BFAuxInt(lc, 16-lc)] x)
+	// result: (SBFIZ [armBFAuxInt(lc, 16-lc)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -14435,7 +14435,7 @@ func rewriteValueARM64_OpARM64MOVHreg_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFIZ)
-		v.AuxInt = arm64BFAuxInt(lc, 16-lc)
+		v.AuxInt = armBFAuxInt(lc, 16-lc)
 		v.AddArg(x)
 		return true
 	}
@@ -14809,7 +14809,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVHstore [i] {s} ptr0 (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstore [i-2] {s} ptr1 w mem))
+	// match: (MOVHstore [i] {s} ptr0 (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstore [i-2] {s} ptr1 w mem))
 	// cond: x.Uses == 1 && isSamePtr(ptr0, ptr1) && clobber(x)
 	// result: (MOVWstore [i-2] {s} ptr0 w mem)
 	for {
@@ -14821,7 +14821,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(16, 16) {
+		if v_1.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		w := v_1.Args[0]
@@ -14852,7 +14852,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVHstore [2] {s} (ADD ptr0 idx0) (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstoreidx ptr1 idx1 w mem))
+	// match: (MOVHstore [2] {s} (ADD ptr0 idx0) (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstoreidx ptr1 idx1 w mem))
 	// cond: x.Uses == 1 && s == nil && (isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) || isSamePtr(ptr0, idx1) && isSamePtr(idx0, ptr1)) && clobber(x)
 	// result: (MOVWstoreidx ptr1 idx1 w mem)
 	for {
@@ -14872,7 +14872,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(16, 16) {
+		if v_1.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		w := v_1.Args[0]
@@ -14897,7 +14897,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		v.AddArg(mem)
 		return true
 	}
-	// match: (MOVHstore [2] {s} (ADDshiftLL [1] ptr0 idx0) (UBFX [arm64BFAuxInt(16, 16)] w) x:(MOVHstoreidx2 ptr1 idx1 w mem))
+	// match: (MOVHstore [2] {s} (ADDshiftLL [1] ptr0 idx0) (UBFX [armBFAuxInt(16, 16)] w) x:(MOVHstoreidx2 ptr1 idx1 w mem))
 	// cond: x.Uses == 1 && s == nil && isSamePtr(ptr0, ptr1) && isSamePtr(idx0, idx1) && clobber(x)
 	// result: (MOVWstoreidx ptr1 (SLLconst <idx1.Type> [1] idx1) w mem)
 	for {
@@ -14920,7 +14920,7 @@ func rewriteValueARM64_OpARM64MOVHstore_10(v *Value) bool {
 		if v_1.Op != OpARM64UBFX {
 			break
 		}
-		if v_1.AuxInt != arm64BFAuxInt(16, 16) {
+		if v_1.AuxInt != armBFAuxInt(16, 16) {
 			break
 		}
 		w := v_1.Args[0]
@@ -16610,7 +16610,7 @@ func rewriteValueARM64_OpARM64MOVWUreg_10(v *Value) bool {
 	}
 	// match: (MOVWUreg (SLLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<32-1, sc)
-	// result: (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
+	// result: (UBFIZ [armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -16622,13 +16622,13 @@ func rewriteValueARM64_OpARM64MOVWUreg_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (MOVWUreg (SRLconst [sc] x))
 	// cond: isARM64BFMask(sc, 1<<32-1, 0)
-	// result: (UBFX [arm64BFAuxInt(sc, 32)] x)
+	// result: (UBFX [armBFAuxInt(sc, 32)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SRLconst {
@@ -16640,7 +16640,7 @@ func rewriteValueARM64_OpARM64MOVWUreg_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, 32)
+		v.AuxInt = armBFAuxInt(sc, 32)
 		v.AddArg(x)
 		return true
 	}
@@ -17168,7 +17168,7 @@ func rewriteValueARM64_OpARM64MOVWreg_10(v *Value) bool {
 	}
 	// match: (MOVWreg (SLLconst [lc] x))
 	// cond: lc < 32
-	// result: (SBFIZ [arm64BFAuxInt(lc, 32-lc)] x)
+	// result: (SBFIZ [armBFAuxInt(lc, 32-lc)] x)
 	for {
 		v_0 := v.Args[0]
 		if v_0.Op != OpARM64SLLconst {
@@ -17180,7 +17180,7 @@ func rewriteValueARM64_OpARM64MOVWreg_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFIZ)
-		v.AuxInt = arm64BFAuxInt(lc, 32-lc)
+		v.AuxInt = armBFAuxInt(lc, 32-lc)
 		v.AddArg(x)
 		return true
 	}
@@ -26620,7 +26620,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (ORshiftLL <t> [c] (UBFX [bfc] x) x)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (RORWconst [32-c] x)
 	for {
 		t := v.Type
@@ -26635,7 +26635,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		if x != v.Args[1] {
 			break
 		}
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64RORWconst)
@@ -26643,7 +26643,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
-	// match: (ORshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [arm64BFAuxInt(8, 8)] x) x)
+	// match: (ORshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
 	// cond:
 	// result: (REV16W x)
 	for {
@@ -26661,7 +26661,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		if v_0.Type != typ.UInt16 {
 			break
 		}
-		if v_0.AuxInt != arm64BFAuxInt(8, 8) {
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		x := v_0.Args[0]
@@ -26694,7 +26694,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (ORshiftLL <t> [c] (UBFX [bfc] x) x2)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (EXTRWconst [32-c] x2 x)
 	for {
 		t := v.Type
@@ -26707,7 +26707,7 @@ func rewriteValueARM64_OpARM64ORshiftLL_0(v *Value) bool {
 		bfc := v_0.AuxInt
 		x := v_0.Args[0]
 		x2 := v.Args[1]
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64EXTRWconst)
@@ -28883,7 +28883,7 @@ func rewriteValueARM64_OpARM64ORshiftRL_0(v *Value) bool {
 	}
 	// match: (ORshiftRL [rc] (ANDconst [ac] x) (SLLconst [lc] y))
 	// cond: lc > rc && ac == ^((1<<uint(64-lc)-1) << uint64(lc-rc))
-	// result: (BFI [arm64BFAuxInt(lc-rc, 64-lc)] x y)
+	// result: (BFI [armBFAuxInt(lc-rc, 64-lc)] x y)
 	for {
 		rc := v.AuxInt
 		_ = v.Args[1]
@@ -28903,7 +28903,7 @@ func rewriteValueARM64_OpARM64ORshiftRL_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64BFI)
-		v.AuxInt = arm64BFAuxInt(lc-rc, 64-lc)
+		v.AuxInt = armBFAuxInt(lc-rc, 64-lc)
 		v.AddArg(x)
 		v.AddArg(y)
 		return true
@@ -29005,7 +29005,7 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 	}
 	// match: (SLLconst [sc] (ANDconst [ac] x))
 	// cond: isARM64BFMask(sc, ac, 0)
-	// result: (UBFIZ [arm64BFAuxInt(sc, arm64BFWidth(ac, 0))] x)
+	// result: (UBFIZ [armBFAuxInt(sc, arm64BFWidth(ac, 0))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29018,13 +29018,13 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(ac, 0))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(ac, 0))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SLLconst [sc] (MOVWUreg x))
 	// cond: isARM64BFMask(sc, 1<<32-1, 0)
-	// result: (UBFIZ [arm64BFAuxInt(sc, 32)] x)
+	// result: (UBFIZ [armBFAuxInt(sc, 32)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29036,13 +29036,13 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, 32)
+		v.AuxInt = armBFAuxInt(sc, 32)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SLLconst [sc] (MOVHUreg x))
 	// cond: isARM64BFMask(sc, 1<<16-1, 0)
-	// result: (UBFIZ [arm64BFAuxInt(sc, 16)] x)
+	// result: (UBFIZ [armBFAuxInt(sc, 16)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29054,13 +29054,13 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, 16)
+		v.AuxInt = armBFAuxInt(sc, 16)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SLLconst [sc] (MOVBUreg x))
 	// cond: isARM64BFMask(sc, 1<<8-1, 0)
-	// result: (UBFIZ [arm64BFAuxInt(sc, 8)] x)
+	// result: (UBFIZ [armBFAuxInt(sc, 8)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29072,13 +29072,13 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc, 8)
+		v.AuxInt = armBFAuxInt(sc, 8)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SLLconst [sc] (UBFIZ [bfc] x))
 	// cond: sc+getARM64BFwidth(bfc)+getARM64BFlsb(bfc) < 64
-	// result: (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
+	// result: (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29091,7 +29091,7 @@ func rewriteValueARM64_OpARM64SLLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))
 		v.AddArg(x)
 		return true
 	}
@@ -29133,7 +29133,7 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 	}
 	// match: (SRAconst [rc] (SLLconst [lc] x))
 	// cond: lc > rc
-	// result: (SBFIZ [arm64BFAuxInt(lc-rc, 64-lc)] x)
+	// result: (SBFIZ [armBFAuxInt(lc-rc, 64-lc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29146,13 +29146,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFIZ)
-		v.AuxInt = arm64BFAuxInt(lc-rc, 64-lc)
+		v.AuxInt = armBFAuxInt(lc-rc, 64-lc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [rc] (SLLconst [lc] x))
 	// cond: lc <= rc
-	// result: (SBFX [arm64BFAuxInt(rc-lc, 64-rc)] x)
+	// result: (SBFX [armBFAuxInt(rc-lc, 64-rc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29165,13 +29165,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFX)
-		v.AuxInt = arm64BFAuxInt(rc-lc, 64-rc)
+		v.AuxInt = armBFAuxInt(rc-lc, 64-rc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [rc] (MOVWreg x))
 	// cond: rc < 32
-	// result: (SBFX [arm64BFAuxInt(rc, 32-rc)] x)
+	// result: (SBFX [armBFAuxInt(rc, 32-rc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29183,13 +29183,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFX)
-		v.AuxInt = arm64BFAuxInt(rc, 32-rc)
+		v.AuxInt = armBFAuxInt(rc, 32-rc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [rc] (MOVHreg x))
 	// cond: rc < 16
-	// result: (SBFX [arm64BFAuxInt(rc, 16-rc)] x)
+	// result: (SBFX [armBFAuxInt(rc, 16-rc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29201,13 +29201,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFX)
-		v.AuxInt = arm64BFAuxInt(rc, 16-rc)
+		v.AuxInt = armBFAuxInt(rc, 16-rc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [rc] (MOVBreg x))
 	// cond: rc < 8
-	// result: (SBFX [arm64BFAuxInt(rc, 8-rc)] x)
+	// result: (SBFX [armBFAuxInt(rc, 8-rc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29219,13 +29219,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFX)
-		v.AuxInt = arm64BFAuxInt(rc, 8-rc)
+		v.AuxInt = armBFAuxInt(rc, 8-rc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [sc] (SBFIZ [bfc] x))
 	// cond: sc < getARM64BFlsb(bfc)
-	// result: (SBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	// result: (SBFIZ [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29238,13 +29238,13 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFIZ)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRAconst [sc] (SBFIZ [bfc] x))
 	// cond: sc >= getARM64BFlsb(bfc) && sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	// result: (SBFX [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	// result: (SBFX [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29257,7 +29257,7 @@ func rewriteValueARM64_OpARM64SRAconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64SBFX)
-		v.AuxInt = arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
+		v.AuxInt = armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
 		v.AddArg(x)
 		return true
 	}
@@ -29320,7 +29320,7 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 	}
 	// match: (SRLconst [rc] (SLLconst [lc] x))
 	// cond: lc > rc
-	// result: (UBFIZ [arm64BFAuxInt(lc-rc, 64-lc)] x)
+	// result: (UBFIZ [armBFAuxInt(lc-rc, 64-lc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29333,13 +29333,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(lc-rc, 64-lc)
+		v.AuxInt = armBFAuxInt(lc-rc, 64-lc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (ANDconst [ac] x))
 	// cond: isARM64BFMask(sc, ac, sc)
-	// result: (UBFX [arm64BFAuxInt(sc, arm64BFWidth(ac, sc))] x)
+	// result: (UBFX [armBFAuxInt(sc, arm64BFWidth(ac, sc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29352,13 +29352,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(ac, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(ac, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (MOVWUreg x))
 	// cond: isARM64BFMask(sc, 1<<32-1, sc)
-	// result: (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
+	// result: (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29370,13 +29370,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<32-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<32-1, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (MOVHUreg x))
 	// cond: isARM64BFMask(sc, 1<<16-1, sc)
-	// result: (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
+	// result: (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29388,13 +29388,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<16-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<16-1, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (MOVBUreg x))
 	// cond: isARM64BFMask(sc, 1<<8-1, sc)
-	// result: (UBFX [arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
+	// result: (UBFX [armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29406,13 +29406,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc, arm64BFWidth(1<<8-1, sc))
+		v.AuxInt = armBFAuxInt(sc, arm64BFWidth(1<<8-1, sc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [rc] (SLLconst [lc] x))
 	// cond: lc < rc
-	// result: (UBFX [arm64BFAuxInt(rc-lc, 64-rc)] x)
+	// result: (UBFX [armBFAuxInt(rc-lc, 64-rc)] x)
 	for {
 		rc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29425,13 +29425,13 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(rc-lc, 64-rc)
+		v.AuxInt = armBFAuxInt(rc-lc, 64-rc)
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (UBFX [bfc] x))
 	// cond: sc < getARM64BFwidth(bfc)
-	// result: (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
+	// result: (UBFX [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29444,7 +29444,7 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)
 		v.AddArg(x)
 		return true
 	}
@@ -29472,7 +29472,7 @@ func rewriteValueARM64_OpARM64SRLconst_0(v *Value) bool {
 func rewriteValueARM64_OpARM64SRLconst_10(v *Value) bool {
 	// match: (SRLconst [sc] (UBFIZ [bfc] x))
 	// cond: sc < getARM64BFlsb(bfc)
-	// result: (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	// result: (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29485,13 +29485,13 @@ func rewriteValueARM64_OpARM64SRLconst_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (SRLconst [sc] (UBFIZ [bfc] x))
 	// cond: sc > getARM64BFlsb(bfc) && sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	// result: (UBFX [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	// result: (UBFX [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 	for {
 		sc := v.AuxInt
 		v_0 := v.Args[0]
@@ -29504,7 +29504,7 @@ func rewriteValueARM64_OpARM64SRLconst_10(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
+		v.AuxInt = armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
 		v.AddArg(x)
 		return true
 	}
@@ -30449,7 +30449,7 @@ func rewriteValueARM64_OpARM64TSTshiftRL_0(v *Value) bool {
 func rewriteValueARM64_OpARM64UBFIZ_0(v *Value) bool {
 	// match: (UBFIZ [bfc] (SLLconst [sc] x))
 	// cond: sc < getARM64BFwidth(bfc)
-	// result: (UBFIZ [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
+	// result: (UBFIZ [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)] x)
 	for {
 		bfc := v.AuxInt
 		v_0 := v.Args[0]
@@ -30462,7 +30462,7 @@ func rewriteValueARM64_OpARM64UBFIZ_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc)-sc)
 		v.AddArg(x)
 		return true
 	}
@@ -30471,7 +30471,7 @@ func rewriteValueARM64_OpARM64UBFIZ_0(v *Value) bool {
 func rewriteValueARM64_OpARM64UBFX_0(v *Value) bool {
 	// match: (UBFX [bfc] (SRLconst [sc] x))
 	// cond: sc+getARM64BFwidth(bfc)+getARM64BFlsb(bfc) < 64
-	// result: (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
+	// result: (UBFX [armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))] x)
 	for {
 		bfc := v.AuxInt
 		v_0 := v.Args[0]
@@ -30484,7 +30484,7 @@ func rewriteValueARM64_OpARM64UBFX_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)+sc, getARM64BFwidth(bfc))
 		v.AddArg(x)
 		return true
 	}
@@ -30509,7 +30509,7 @@ func rewriteValueARM64_OpARM64UBFX_0(v *Value) bool {
 	}
 	// match: (UBFX [bfc] (SLLconst [sc] x))
 	// cond: sc < getARM64BFlsb(bfc)
-	// result: (UBFX [arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
+	// result: (UBFX [armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))] x)
 	for {
 		bfc := v.AuxInt
 		v_0 := v.Args[0]
@@ -30522,13 +30522,13 @@ func rewriteValueARM64_OpARM64UBFX_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFX)
-		v.AuxInt = arm64BFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
+		v.AuxInt = armBFAuxInt(getARM64BFlsb(bfc)-sc, getARM64BFwidth(bfc))
 		v.AddArg(x)
 		return true
 	}
 	// match: (UBFX [bfc] (SLLconst [sc] x))
 	// cond: sc > getARM64BFlsb(bfc) && sc < getARM64BFlsb(bfc)+getARM64BFwidth(bfc)
-	// result: (UBFIZ [arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
+	// result: (UBFIZ [armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)] x)
 	for {
 		bfc := v.AuxInt
 		v_0 := v.Args[0]
@@ -30541,7 +30541,7 @@ func rewriteValueARM64_OpARM64UBFX_0(v *Value) bool {
 			break
 		}
 		v.reset(OpARM64UBFIZ)
-		v.AuxInt = arm64BFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
+		v.AuxInt = armBFAuxInt(sc-getARM64BFlsb(bfc), getARM64BFlsb(bfc)+getARM64BFwidth(bfc)-sc)
 		v.AddArg(x)
 		return true
 	}
@@ -32099,7 +32099,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (XORshiftLL <t> [c] (UBFX [bfc] x) x)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (RORWconst [32-c] x)
 	for {
 		t := v.Type
@@ -32114,7 +32114,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		if x != v.Args[1] {
 			break
 		}
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64RORWconst)
@@ -32122,7 +32122,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
-	// match: (XORshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [arm64BFAuxInt(8, 8)] x) x)
+	// match: (XORshiftLL <typ.UInt16> [8] (UBFX <typ.UInt16> [armBFAuxInt(8, 8)] x) x)
 	// cond:
 	// result: (REV16W x)
 	for {
@@ -32140,7 +32140,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		if v_0.Type != typ.UInt16 {
 			break
 		}
-		if v_0.AuxInt != arm64BFAuxInt(8, 8) {
+		if v_0.AuxInt != armBFAuxInt(8, 8) {
 			break
 		}
 		x := v_0.Args[0]
@@ -32173,7 +32173,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		return true
 	}
 	// match: (XORshiftLL <t> [c] (UBFX [bfc] x) x2)
-	// cond: c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)
+	// cond: c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	// result: (EXTRWconst [32-c] x2 x)
 	for {
 		t := v.Type
@@ -32186,7 +32186,7 @@ func rewriteValueARM64_OpARM64XORshiftLL_0(v *Value) bool {
 		bfc := v_0.AuxInt
 		x := v_0.Args[0]
 		x2 := v.Args[1]
-		if !(c < 32 && t.Size() == 4 && bfc == arm64BFAuxInt(32-c, c)) {
+		if !(c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)) {
 			break
 		}
 		v.reset(OpARM64EXTRWconst)
diff --git a/test/codegen/mathbits.go b/test/codegen/mathbits.go
index cc3c91eb0d..09939bb6be 100644
--- a/test/codegen/mathbits.go
+++ b/test/codegen/mathbits.go
@@ -171,6 +171,9 @@ func ReverseBytes32(n uint32) uint32 {
 func ReverseBytes16(n uint16) uint16 {
 	// amd64:"ROLW"
 	// arm64:"REV16W",-"UBFX",-"ORR"
+	// arm/5:"SLL","SRL","ORR"
+	// arm/6:"REV16"
+	// arm/7:"REV16"
 	return bits.ReverseBytes16(n)
 }