cmd/compile: merge zero/sign extensions with UBFX/SBFX on arm64

The UBFX and SBFX already zero/sign extend the result. Further
zero/sign extensions are thus unnecessary as long as they leave
the top bits unaltered. This patch absorbs zero/sign extensions
into UBFX/SBFX.

Add the related test cases.

Change-Id: I7c4516c8b52d677f77bf3aaedab87c4a28056ec0
Reviewed-on: https://go-review.googlesource.com/c/go/+/265039
Trust: fannie zhang <Fannie.Zhang@arm.com>
Trust: Keith Randall <khr@golang.org>
Run-TryBot: fannie zhang <Fannie.Zhang@arm.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>

src/cmd/compile/internal/ssa/gen/ARM64.rules

@@ -1834,6 +1834,10 @@
 (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)
+// merge sbfx and sign-extension into sbfx
+(MOVWreg (SBFX [bfc] x)) && bfc.getARM64BFwidth() <= 32 => (SBFX [bfc] x)
+(MOVHreg (SBFX [bfc] x)) && bfc.getARM64BFwidth() <= 16 => (SBFX [bfc] x)
+(MOVBreg (SBFX [bfc] x)) && bfc.getARM64BFwidth() <=  8 => (SBFX [bfc] x)
 
 // sbfiz/sbfx combinations: merge shifts into bitfield ops
 (SRAconst [sc] (SBFIZ [bfc] x)) && sc < bfc.getARM64BFlsb()
@@ -1880,6 +1884,11 @@
 // (x << lc) >> rc
 (SRLconst [rc] (SLLconst [lc] x)) && lc < rc => (UBFX [armBFAuxInt(rc-lc, 64-rc)] x)
 
+// merge ubfx and zerso-extension into ubfx
+(MOVWUreg (UBFX [bfc] x)) && bfc.getARM64BFwidth() <= 32 => (UBFX [bfc] x)
+(MOVHUreg (UBFX [bfc] x)) && bfc.getARM64BFwidth() <= 16 => (UBFX [bfc] x)
+(MOVBUreg (UBFX [bfc] x)) && bfc.getARM64BFwidth() <=  8 => (UBFX [bfc] x)
+
 // ubfiz/ubfx combinations: merge shifts into bitfield ops
 (SRLconst [sc] (UBFX [bfc] x)) && sc < bfc.getARM64BFwidth()
 	=> (UBFX [armBFAuxInt(bfc.getARM64BFlsb()+sc, bfc.getARM64BFwidth()-sc)] x)

src/cmd/compile/internal/ssa/rewriteARM64.go

@@ -7187,6 +7187,23 @@ func rewriteValueARM64_OpARM64MOVBUreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVBUreg (UBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 8
+	// result: (UBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64UBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 8) {
+			break
+		}
+		v.reset(OpARM64UBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVBload(v *Value) bool {
@@ -7401,6 +7418,23 @@ func rewriteValueARM64_OpARM64MOVBreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVBreg (SBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 8
+	// result: (SBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64SBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 8) {
+			break
+		}
+		v.reset(OpARM64SBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVBstore(v *Value) bool {
@@ -10699,6 +10733,23 @@ func rewriteValueARM64_OpARM64MOVHUreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVHUreg (UBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 16
+	// result: (UBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64UBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 16) {
+			break
+		}
+		v.reset(OpARM64UBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVHload(v *Value) bool {
@@ -11096,6 +11147,23 @@ func rewriteValueARM64_OpARM64MOVHreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVHreg (SBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 16
+	// result: (SBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64SBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 16) {
+			break
+		}
+		v.reset(OpARM64SBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVHstore(v *Value) bool {
@@ -12811,6 +12879,23 @@ func rewriteValueARM64_OpARM64MOVWUreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVWUreg (UBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 32
+	// result: (UBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64UBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 32) {
+			break
+		}
+		v.reset(OpARM64UBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVWload(v *Value) bool {
@@ -13266,6 +13351,23 @@ func rewriteValueARM64_OpARM64MOVWreg(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (MOVWreg (SBFX [bfc] x))
+	// cond: bfc.getARM64BFwidth() <= 32
+	// result: (SBFX [bfc] x)
+	for {
+		if v_0.Op != OpARM64SBFX {
+			break
+		}
+		bfc := auxIntToArm64BitField(v_0.AuxInt)
+		x := v_0.Args[0]
+		if !(bfc.getARM64BFwidth() <= 32) {
+			break
+		}
+		v.reset(OpARM64SBFX)
+		v.AuxInt = arm64BitFieldToAuxInt(bfc)
+		v.AddArg(x)
+		return true
+	}
 	return false
 }
 func rewriteValueARM64_OpARM64MOVWstore(v *Value) bool {

test/codegen/bitfield.go

@@ -124,6 +124,12 @@ func sbfx6(x int32) int32 {
 	return (x << 3) >> 4 // arm64:"SBFX\t[$]1, R[0-9]+, [$]28",-"LSL",-"ASR"
 }
 
+// merge sbfx and sign-extension into sbfx.
+func sbfx7(x int32) int64 {
+	c := x + 5
+	return int64(c >> 20) // arm64"SBFX\t[$]20, R[0-9]+, [$]12",-"MOVW\tR[0-9]+, R[0-9]+"
+}
+
 // ubfiz
 func ubfiz1(x uint64) uint64 {
 	// arm64:"UBFIZ\t[$]3, R[0-9]+, [$]12",-"LSL",-"AND"
@@ -237,6 +243,16 @@ func ubfx11(x uint64) uint64 {
 	return ((x & 0xfffff) << 3) >> 4
 }
 
+// merge ubfx and zero-extension into ubfx.
+func ubfx12(x uint64) bool {
+	midr := x + 10
+	part_num := uint16((midr >> 4) & 0xfff)
+	if part_num == 0xd0c { // arm64:"UBFX\t[$]4, R[0-9]+, [$]12",-"MOVHU\tR[0-9]+, R[0-9]+"
+		return true
+	}
+	return false
+}
+
 // Check that we don't emit comparisons for constant shifts.
 //go:nosplit
 func shift_no_cmp(x int) int {