AuroraMiddleware/v8
1
0
Fork 0
Code Issues 2 Pull Requests Releases Wiki Activity

[ia32] Fix partial regression in Cvtsi2ss/Cvtsi2sd

Browse Source 
In https://crrev.com/c/3131374 we switched some instructions to use
macro-assembler functions which can handle AVX and SSE. However for
Cvtsi2ss and Cvtsi2sd, the behavior subtly changed. The old behavior
directly called cvtsi2ss/cvtsi2sd in the code-generator. The new
behavior used the macro-assembler functions, which xor the dst operand.
This led to more instructions and larger code size in some benchmarks.
The xor is supposed to help reduce dependence chain length (see comments
on Cvtsi2ss), but doesn't seem to have helped in this benchmark.

So, partially revert the changes, and rename all affected IA32 opcodes
back to SSE.

Bug: chromium:1248509
Change-Id: Ie700e2980fe9ed083c1160bda3a28f64e1e43041
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3154349
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Adam Klein <adamk@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76775}
This commit is contained in:
Ng Zhi An 2021-09-10 14:17:02 -07:00 committed by V8 LUCI CQ
parent d7c9b31a77
commit 5f622d21a5
4 changed files with 12 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
src/compiler/backend/ia32/code-generator-ia32.cc
View File

@ -1439,15 +1439,17 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
__ Cvttsd2ui(i.OutputRegister(), i.InputOperand(0),
i.TempSimd128Register(0));
break;
case kIA32Int32ToFloat32:
__ Cvtsi2ss(i.OutputDoubleRegister(), i.InputOperand(0));
case kSSEInt32ToFloat32:
// Calling Cvtsi2ss (which does a xor) regresses some benchmarks.
__ cvtsi2ss(i.OutputDoubleRegister(), i.InputOperand(0));
break;
case kIA32Uint32ToFloat32:
__ Cvtui2ss(i.OutputDoubleRegister(), i.InputOperand(0),
i.TempRegister(0));
break;
case kIA32Int32ToFloat64:
__ Cvtsi2sd(i.OutputDoubleRegister(), i.InputOperand(0));
case kSSEInt32ToFloat64:
// Calling Cvtsi2sd (which does a xor) regresses some benchmarks.
__ cvtsi2sd(i.OutputDoubleRegister(), i.InputOperand(0));
break;
case kIA32Uint32ToFloat64:
__ Cvtui2sd(i.OutputDoubleRegister(), i.InputOperand(0),

4
src/compiler/backend/ia32/instruction-codes-ia32.h
View File

@ -64,9 +64,9 @@ namespace compiler {
V(IA32Float32ToUint32) \
V(IA32Float64ToInt32) \
V(IA32Float64ToUint32) \
V(IA32Int32ToFloat32) \
V(SSEInt32ToFloat32) \
V(IA32Uint32ToFloat32) \
V(IA32Int32ToFloat64) \
V(SSEInt32ToFloat64) \
V(IA32Uint32ToFloat64) \
V(IA32Float64ExtractLowWord32) \
V(IA32Float64ExtractHighWord32) \

4
src/compiler/backend/ia32/instruction-scheduler-ia32.cc
View File

@ -65,9 +65,9 @@ int InstructionScheduler::GetTargetInstructionFlags(
case kIA32Float32ToUint32:
case kIA32Float64ToInt32:
case kIA32Float64ToUint32:
case kIA32Int32ToFloat32:
case kSSEInt32ToFloat32:
case kIA32Uint32ToFloat32:
case kIA32Int32ToFloat64:
case kSSEInt32ToFloat64:
case kIA32Uint32ToFloat64:
case kIA32Float64ExtractLowWord32:
case kIA32Float64ExtractHighWord32:

4
src/compiler/backend/ia32/instruction-selector-ia32.cc
View File

@ -1128,8 +1128,8 @@ void InstructionSelector::VisitWord32Ror(Node* node) {
V(Word32Ctz, kIA32Tzcnt) \
V(Word32Popcnt, kIA32Popcnt) \
V(ChangeFloat32ToFloat64, kIA32Float32ToFloat64) \
V(RoundInt32ToFloat32, kIA32Int32ToFloat32) \
V(ChangeInt32ToFloat64, kIA32Int32ToFloat64) \
V(RoundInt32ToFloat32, kSSEInt32ToFloat32) \
V(ChangeInt32ToFloat64, kSSEInt32ToFloat64) \
V(TruncateFloat32ToInt32, kIA32Float32ToInt32) \
V(ChangeFloat64ToInt32, kIA32Float64ToInt32) \
V(TruncateFloat64ToFloat32, kIA32Float64ToFloat32) \