Reviewers: dcarney,
Description:
[x86] Allow (v)divsd->(v)mulsd to execute in parallel.
This tweak was already present in CrankShaft for the non-AVX case. As it
turns out, it's also relevant even with AVX. Now the same optimization
is applied in case of TurboFan as well.
[email protected]
Please review this at https://codereview.chromium.org/1081033003/
Base URL: https://chromium.googlesource.com/v8/v8.git@master
Affected files (+30, -6 lines):
M src/compiler/ia32/code-generator-ia32.cc
M src/compiler/x64/code-generator-x64.cc
M src/ia32/lithium-codegen-ia32.cc
M src/x64/lithium-codegen-x64.cc
Index: src/compiler/ia32/code-generator-ia32.cc
diff --git a/src/compiler/ia32/code-generator-ia32.cc
b/src/compiler/ia32/code-generator-ia32.cc
index
aaa34dcfd66f3dcc58fc70edc50f044651074dbf..76a457b8f6645d82b8c97753511fe4777df111e4
100644
--- a/src/compiler/ia32/code-generator-ia32.cc
+++ b/src/compiler/ia32/code-generator-ia32.cc
@@ -468,6 +468,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kSSEFloat32Div:
__ divss(i.InputDoubleRegister(0), i.InputOperand(1));
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulss depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kSSEFloat32Max:
__ maxss(i.InputDoubleRegister(0), i.InputOperand(1));
@@ -506,6 +509,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kSSEFloat64Div:
__ divsd(i.InputDoubleRegister(0), i.InputOperand(1));
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kSSEFloat64Max:
__ maxsd(i.InputDoubleRegister(0), i.InputOperand(1));
@@ -629,6 +635,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
CpuFeatureScope avx_scope(masm(), AVX);
__ vdivss(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputOperand(1));
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulss depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
}
case kAVXFloat32Max: {
@@ -665,6 +674,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
CpuFeatureScope avx_scope(masm(), AVX);
__ vdivsd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputOperand(1));
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
}
case kAVXFloat64Max: {
Index: src/compiler/x64/code-generator-x64.cc
diff --git a/src/compiler/x64/code-generator-x64.cc
b/src/compiler/x64/code-generator-x64.cc
index
1e5d62f7e3c522a336c29bbf9214d5d24afc5d3a..4213c9f135131957dc51e93505a2b7f4444df20f
100644
--- a/src/compiler/x64/code-generator-x64.cc
+++ b/src/compiler/x64/code-generator-x64.cc
@@ -726,6 +726,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kSSEFloat32Div:
ASSEMBLE_SSE_BINOP(divss);
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulss depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kSSEFloat32Abs: {
// TODO(bmeurer): Use RIP relative 128-bit constants.
@@ -767,6 +770,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kSSEFloat64Div:
ASSEMBLE_SSE_BINOP(divsd);
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kSSEFloat64Mod: {
__ subq(rsp, Immediate(kDoubleSize));
@@ -919,6 +925,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kAVXFloat32Div:
ASSEMBLE_AVX_BINOP(vdivss);
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulss depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kAVXFloat32Max:
ASSEMBLE_AVX_BINOP(vmaxss);
@@ -946,6 +955,9 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
break;
case kAVXFloat64Div:
ASSEMBLE_AVX_BINOP(vdivsd);
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result.
+ __ movaps(i.OutputDoubleRegister(), i.OutputDoubleRegister());
break;
case kAVXFloat64Max:
ASSEMBLE_AVX_BINOP(vmaxsd);
Index: src/ia32/lithium-codegen-ia32.cc
diff --git a/src/ia32/lithium-codegen-ia32.cc
b/src/ia32/lithium-codegen-ia32.cc
index
7714e9f6e9bae8a59a76f29e136268c548f0a30c..72bb270e7a6fcfe384da3579b12a989e721b9cd2
100644
--- a/src/ia32/lithium-codegen-ia32.cc
+++ b/src/ia32/lithium-codegen-ia32.cc
@@ -2040,10 +2040,10 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
} else {
DCHECK(result.is(left));
__ divsd(left, right);
- // Don't delete this mov. It may improve performance on some CPUs,
- // when there is a mulsd depending on the result
- __ movaps(left, left);
}
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result
+ __ movaps(result, result);
break;
case Token::MOD: {
// Pass two doubles as arguments on the stack.
Index: src/x64/lithium-codegen-x64.cc
diff --git a/src/x64/lithium-codegen-x64.cc b/src/x64/lithium-codegen-x64.cc
index
8da76fa673ed465cf6228b08ad8ffe1deb68f594..3b72d690fe514c17d17d939b3ad3b13d24748b86
100644
--- a/src/x64/lithium-codegen-x64.cc
+++ b/src/x64/lithium-codegen-x64.cc
@@ -2076,10 +2076,10 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
} else {
DCHECK(result.is(left));
__ divsd(left, right);
- // Don't delete this mov. It may improve performance on some CPUs,
- // when there is a mulsd depending on the result
- __ movaps(left, left);
}
+ // Don't delete this mov. It may improve performance on some CPUs,
+ // when there is a (v)mulsd depending on the result
+ __ movaps(result, result);
break;
case Token::MOD: {
XMMRegister xmm_scratch = double_scratch0();
--
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
---
You received this message because you are subscribed to the Google Groups "v8-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
