Reviewers: dcarney,
Description:
[x64] Match -0 - x with sign bit flip.
We can use xorps/xorpd on Intel CPUs to flip the sign bit. Ideally we'd
use a RIP-relative 128-bit constant in the code object, as OCaml/GCC
does, however that requires 128-bit alignment for code objects, which is
not yet implemented. So for now we materialize the mask inline.
[email protected]
Please review this at https://codereview.chromium.org/1046893002/
Base URL: https://chromium.googlesource.com/v8/v8.git@master
Affected files (+82, -12 lines):
M src/compiler/x64/code-generator-x64.cc
M src/compiler/x64/instruction-codes-x64.h
M src/compiler/x64/instruction-selector-x64.cc
M test/cctest/compiler/test-run-machops.cc
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
fee2702332d463a4f724043d00c23b0cb0084bcd..98ec629b4d702ba2d8d033a34e830ec6b3cc2939
100644
--- a/src/compiler/x64/code-generator-x64.cc
+++ b/src/compiler/x64/code-generator-x64.cc
@@ -726,15 +726,23 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
case kSSEFloat32Div:
ASSEMBLE_SSE_BINOP(divss);
break;
+ case kSSEFloat32Neg: {
+ // TODO(bmeurer): Use RIP relative 128-bit constants.
+ // TODO(turbofan): Add AVX version with relaxed register constraints.
+ __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+ __ psllq(kScratchDoubleReg, 31);
+ __ xorps(i.OutputDoubleRegister(), kScratchDoubleReg);
+ break;
+ }
+ case kSSEFloat32Sqrt:
+ ASSEMBLE_SSE_UNOP(sqrtss);
+ break;
case kSSEFloat32Max:
ASSEMBLE_SSE_BINOP(maxss);
break;
case kSSEFloat32Min:
ASSEMBLE_SSE_BINOP(minss);
break;
- case kSSEFloat32Sqrt:
- ASSEMBLE_SSE_UNOP(sqrtss);
- break;
case kSSEFloat32ToFloat64:
ASSEMBLE_SSE_UNOP(cvtss2sd);
break;
@@ -791,6 +799,14 @@ void
CodeGenerator::AssembleArchInstruction(Instruction* instr) {
case kSSEFloat64Min:
ASSEMBLE_SSE_BINOP(minsd);
break;
+ case kSSEFloat64Neg: {
+ // TODO(bmeurer): Use RIP relative 128-bit constants.
+ // TODO(turbofan): Add AVX version with relaxed register constraints.
+ __ pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+ __ psllq(kScratchDoubleReg, 63);
+ __ xorpd(i.OutputDoubleRegister(), kScratchDoubleReg);
+ break;
+ }
case kSSEFloat64Sqrt:
ASSEMBLE_SSE_UNOP(sqrtsd);
break;
Index: src/compiler/x64/instruction-codes-x64.h
diff --git a/src/compiler/x64/instruction-codes-x64.h
b/src/compiler/x64/instruction-codes-x64.h
index
2feafd27e5928c2bec6de475b04410599135fd68..8b0469e1c6d319207f0d5bdbcb2ab3e2d27ed6bd
100644
--- a/src/compiler/x64/instruction-codes-x64.h
+++ b/src/compiler/x64/instruction-codes-x64.h
@@ -52,6 +52,7 @@ namespace compiler {
V(SSEFloat32Sub) \
V(SSEFloat32Mul) \
V(SSEFloat32Div) \
+ V(SSEFloat32Neg) \
V(SSEFloat32Sqrt) \
V(SSEFloat32Max) \
V(SSEFloat32Min) \
@@ -62,6 +63,7 @@ namespace compiler {
V(SSEFloat64Mul) \
V(SSEFloat64Div) \
V(SSEFloat64Mod) \
+ V(SSEFloat64Neg) \
V(SSEFloat64Sqrt) \
V(SSEFloat64Round) \
V(SSEFloat64Max) \
Index: src/compiler/x64/instruction-selector-x64.cc
diff --git a/src/compiler/x64/instruction-selector-x64.cc
b/src/compiler/x64/instruction-selector-x64.cc
index
5f3fc06865d35d4d0d7b816039de5b36d91c2f33..82829a88352e8eb5db3e99af5ae5df4077d074ce
100644
--- a/src/compiler/x64/instruction-selector-x64.cc
+++ b/src/compiler/x64/instruction-selector-x64.cc
@@ -858,6 +858,13 @@ void InstructionSelector::VisitFloat32Add(Node* node) {
void InstructionSelector::VisitFloat32Sub(Node* node) {
+ X64OperandGenerator g(this);
+ Float32BinopMatcher m(node);
+ if (m.left().IsMinusZero()) {
+ Emit(kSSEFloat32Neg, g.DefineSameAsFirst(node),
+ g.UseRegister(m.right().node()));
+ return;
+ }
VisitFloatBinop(this, node, kAVXFloat32Sub, kSSEFloat32Sub);
}
@@ -896,17 +903,22 @@ void InstructionSelector::VisitFloat64Add(Node* node)
{
void InstructionSelector::VisitFloat64Sub(Node* node) {
X64OperandGenerator g(this);
Float64BinopMatcher m(node);
- if (m.left().IsMinusZero() && m.right().IsFloat64RoundDown() &&
- CanCover(m.node(), m.right().node())) {
- if (m.right().InputAt(0)->opcode() == IrOpcode::kFloat64Sub &&
- CanCover(m.right().node(), m.right().InputAt(0))) {
- Float64BinopMatcher mright0(m.right().InputAt(0));
- if (mright0.left().IsMinusZero()) {
- Emit(kSSEFloat64Round | MiscField::encode(kRoundUp),
- g.DefineAsRegister(node),
g.UseRegister(mright0.right().node()));
- return;
+ if (m.left().IsMinusZero()) {
+ if (m.right().IsFloat64RoundDown() &&
+ CanCover(m.node(), m.right().node())) {
+ if (m.right().InputAt(0)->opcode() == IrOpcode::kFloat64Sub &&
+ CanCover(m.right().node(), m.right().InputAt(0))) {
+ Float64BinopMatcher mright0(m.right().InputAt(0));
+ if (mright0.left().IsMinusZero()) {
+ Emit(kSSEFloat64Round | MiscField::encode(kRoundUp),
+ g.DefineAsRegister(node),
g.UseRegister(mright0.right().node()));
+ return;
+ }
}
}
+ Emit(kSSEFloat64Neg, g.DefineSameAsFirst(node),
+ g.UseRegister(m.right().node()));
+ return;
}
VisitFloatBinop(this, node, kAVXFloat64Sub, kSSEFloat64Sub);
}
Index: test/cctest/compiler/test-run-machops.cc
diff --git a/test/cctest/compiler/test-run-machops.cc
b/test/cctest/compiler/test-run-machops.cc
index
fdc1a48eafb6ea3afee1b0ad06c93ac389e22832..1686d377883e9c68898e301461b46d36a384384a
100644
--- a/test/cctest/compiler/test-run-machops.cc
+++ b/test/cctest/compiler/test-run-machops.cc
@@ -3212,6 +3212,46 @@ TEST(RunFloat32SubP) {
}
+TEST(RunFloat32SubImm1) {
+ float input = 0.0f;
+ float output = 0.0f;
+
+ FOR_FLOAT32_INPUTS(i) {
+ RawMachineAssemblerTester<int32_t> m;
+ Node* t0 = m.LoadFromPointer(&input, kMachFloat32);
+ Node* t1 = m.Float32Sub(m.Float32Constant(*i), t0);
+ m.StoreToPointer(&output, kMachFloat32, t1);
+ m.Return(m.Int32Constant(0));
+ FOR_FLOAT32_INPUTS(j) {
+ input = *j;
+ float expected = *i - input;
+ CHECK_EQ(0, m.Call());
+ CheckFloatEq(expected, output);
+ }
+ }
+}
+
+
+TEST(RunFloat32SubImm2) {
+ float input = 0.0f;
+ float output = 0.0f;
+
+ FOR_FLOAT32_INPUTS(i) {
+ RawMachineAssemblerTester<int32_t> m;
+ Node* t0 = m.LoadFromPointer(&input, kMachFloat32);
+ Node* t1 = m.Float32Sub(t0, m.Float32Constant(*i));
+ m.StoreToPointer(&output, kMachFloat32, t1);
+ m.Return(m.Int32Constant(0));
+ FOR_FLOAT32_INPUTS(j) {
+ input = *j;
+ float expected = input - *i;
+ CHECK_EQ(0, m.Call());
+ CheckFloatEq(expected, output);
+ }
+ }
+}
+
+
TEST(RunFloat64SubP) {
RawMachineAssemblerTester<int32_t> m;
Float64BinopTester bt(&m);
--
--
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.
