From: Jakub Ciolek <jakub@ciolek.dev>
Date: Tue, 7 Jan 2025 06:00:24 +0000 (+0100)
Subject: cmd/compile: on AMD64, prefer XOR/AND for (x & 1) == 0 check
X-Git-Tag: go1.25rc1~1200
X-Git-Url: http://www.git.cypherpunks.su/?a=commitdiff_plain;h=e57769d5ad251a26d7676cd7a63e5332b26a35f0;p=gostls13.git

cmd/compile: on AMD64, prefer XOR/AND for (x & 1) == 0 check

It's shorter to encode. Additionally, XOR and AND generally
have higher throughput than BT/SET*.

compilecmp:

runtime
runtime.(*sweepClass).split 58 -> 56  (-3.45%)
runtime.sweepClass.split 14 -> 11  (-21.43%)

runtime [cmd/compile]
runtime.(*sweepClass).split 58 -> 56  (-3.45%)
runtime.sweepClass.split 14 -> 11  (-21.43%)

strconv
strconv.ryuFtoaShortest changed

strconv [cmd/compile]
strconv.ryuFtoaShortest changed

math/big
math/big.(*Int).MulRange 255 -> 252  (-1.18%)

testing/quick
testing/quick.sizedValue changed

internal/fuzz
internal/fuzz.(*pcgRand).bool 69 -> 70  (+1.45%)

cmd/internal/obj/x86
cmd/internal/obj/x86.(*AsmBuf).asmevex changed

math/big [cmd/compile]
math/big.(*Int).MulRange 255 -> 252  (-1.18%)

cmd/internal/obj/x86 [cmd/compile]
cmd/internal/obj/x86.(*AsmBuf).asmevex changed

net/http
net/http.(*http2stream).isPushed 11 -> 10  (-9.09%)

cmd/vendor/github.com/google/pprof/internal/binutils
cmd/vendor/github.com/google/pprof/internal/binutils.(*file).computeBase changed

Change-Id: I9cb2987eb263c85ee4e93d6f8455c91a55273173
Reviewed-on: https://go-review.googlesource.com/c/go/+/640975
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Keith Randall <khr@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
---

diff --git a/src/cmd/compile/internal/ssa/_gen/AMD64.rules b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
index ce9a6e9914..716f4f1c32 100644
--- a/src/cmd/compile/internal/ssa/_gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
@@ -629,6 +629,8 @@
 // x & 1 != 0 -> x & 1
 (SETNE (TEST(B|W)const [1] x)) => (AND(L|L)const [1] x)
 (SETB (BT(L|Q)const [0] x)) => (AND(L|Q)const [1] x)
+// x & 1 == 0 -> (x & 1) ^ 1
+(SETAE (BT(L|Q)const [0] x)) => (XORLconst [1] (ANDLconst <typ.Bool> [1] x))
 
 // Recognize bit tests: a&(1<<b) != 0 for b suitably bounded
 // Note that BTx instructions use the carry bit, so we need to convert tests for zero flag
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index f17c4be516..d62f38f0e2 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -16503,6 +16503,8 @@ func rewriteValueAMD64_OpAMD64SETA(v *Value) bool {
 }
 func rewriteValueAMD64_OpAMD64SETAE(v *Value) bool {
 	v_0 := v.Args[0]
+	b := v.Block
+	typ := &b.Func.Config.Types
 	// match: (SETAE (TESTQ x x))
 	// result: (ConstBool [true])
 	for {
@@ -16559,6 +16561,36 @@ func rewriteValueAMD64_OpAMD64SETAE(v *Value) bool {
 		v.AuxInt = boolToAuxInt(true)
 		return true
 	}
+	// match: (SETAE (BTLconst [0] x))
+	// result: (XORLconst [1] (ANDLconst <typ.Bool> [1] x))
+	for {
+		if v_0.Op != OpAMD64BTLconst || auxIntToInt8(v_0.AuxInt) != 0 {
+			break
+		}
+		x := v_0.Args[0]
+		v.reset(OpAMD64XORLconst)
+		v.AuxInt = int32ToAuxInt(1)
+		v0 := b.NewValue0(v.Pos, OpAMD64ANDLconst, typ.Bool)
+		v0.AuxInt = int32ToAuxInt(1)
+		v0.AddArg(x)
+		v.AddArg(v0)
+		return true
+	}
+	// match: (SETAE (BTQconst [0] x))
+	// result: (XORLconst [1] (ANDLconst <typ.Bool> [1] x))
+	for {
+		if v_0.Op != OpAMD64BTQconst || auxIntToInt8(v_0.AuxInt) != 0 {
+			break
+		}
+		x := v_0.Args[0]
+		v.reset(OpAMD64XORLconst)
+		v.AuxInt = int32ToAuxInt(1)
+		v0 := b.NewValue0(v.Pos, OpAMD64ANDLconst, typ.Bool)
+		v0.AuxInt = int32ToAuxInt(1)
+		v0.AddArg(x)
+		v.AddArg(v0)
+		return true
+	}
 	// match: (SETAE (InvertFlags x))
 	// result: (SETBE x)
 	for {
diff --git a/test/codegen/bool.go b/test/codegen/bool.go
index 164ca1b224..2024759a5c 100644
--- a/test/codegen/bool.go
+++ b/test/codegen/bool.go
@@ -47,6 +47,7 @@ func convertNeqBool32(x uint32) bool {
 
 func convertEqBool32(x uint32) bool {
 	// ppc64x:"RLDICL",-"CMPW","XOR",-"ISEL"
+	// amd64:"ANDL","XORL",-"BTL",-"SETCC"
 	return x&1 == 0
 }
 
@@ -57,6 +58,7 @@ func convertNeqBool64(x uint64) bool {
 
 func convertEqBool64(x uint64) bool {
 	// ppc64x:"RLDICL","XOR",-"CMP",-"ISEL"
+	// amd64:"ANDL","XORL",-"BTL",-"SETCC"
 	return x&1 == 0
 }