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

[wasm][simd] Revert changes to F32x4 Min/Max

Browse Source 
- Restores the old inline code sequence, since the branching version
  doesn't set the NaN high bit.

Bug: v8:10862
Change-Id: Iad8ee47b678cc1c6c04222dd83b2fa588ea9136c
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2387557
Reviewed-by: Zhi An Ng <zhin@chromium.org>
Commit-Queue: Bill Budge <bbudge@chromium.org>
Cr-Commit-Position: refs/heads/master@{#69656}
This commit is contained in:
Bill Budge 2020-09-01 12:05:49 -07:00 committed by Commit Bot
parent d022b74c4e
commit b7704fb184
1 changed files with 40 additions and 122 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

162
src/compiler/backend/x64/code-generator-x64.cc
View File

@ -194,94 +194,6 @@ class OutOfLineLoadFloat64NaN final : public OutOfLineCode {
XMMRegister const result_;
};
class OutOfLineF32x4Min final : public OutOfLineCode {
public:
OutOfLineF32x4Min(CodeGenerator* gen, XMMRegister result, XMMRegister error)
: OutOfLineCode(gen), result_(result), error_(error) {}
void Generate() final {
// |result| is the partial result, |kScratchDoubleReg| is the error.
// propagate -0's and NaNs (possibly non-canonical) from the error.
__ Orps(error_, result_);
// Canonicalize NaNs by quieting and clearing the payload.
__ Cmpps(result_, error_, int8_t{3});
__ Orps(error_, result_);
__ Psrld(result_, byte{10});
__ Andnps(result_, error_);
}
private:
XMMRegister const result_;
XMMRegister const error_;
};
class OutOfLineF64x2Min final : public OutOfLineCode {
public:
OutOfLineF64x2Min(CodeGenerator* gen, XMMRegister result, XMMRegister error)
: OutOfLineCode(gen), result_(result), error_(error) {}
void Generate() final {
// |result| is the partial result, |kScratchDoubleReg| is the error.
// propagate -0's and NaNs (possibly non-canonical) from the error.
__ Orpd(error_, result_);
// Canonicalize NaNs by quieting and clearing the payload.
__ Cmppd(result_, error_, int8_t{3});
__ Orpd(error_, result_);
__ Psrlq(result_, 13);
__ Andnpd(result_, error_);
}
private:
XMMRegister const result_;
XMMRegister const error_;
};
class OutOfLineF32x4Max final : public OutOfLineCode {
public:
OutOfLineF32x4Max(CodeGenerator* gen, XMMRegister result, XMMRegister error)
: OutOfLineCode(gen), result_(result), error_(error) {}
void Generate() final {
// |result| is the partial result, |kScratchDoubleReg| is the error.
// Propagate NaNs (possibly non-canonical).
__ Orps(result_, error_);
// Propagate sign errors and (subtle) quiet NaNs.
__ Subps(result_, error_);
// Canonicalize NaNs by clearing the payload. Sign is non-deterministic.
__ Cmpps(error_, result_, int8_t{3});
__ Psrld(error_, byte{10});
__ Andnps(error_, result_);
__ Movaps(result_, error_);
}
private:
XMMRegister const result_;
XMMRegister const error_;
};
class OutOfLineF64x2Max final : public OutOfLineCode {
public:
OutOfLineF64x2Max(CodeGenerator* gen, XMMRegister result, XMMRegister error)
: OutOfLineCode(gen), result_(result), error_(error) {}
void Generate() final {
// |result| is the partial result, |kScratchDoubleReg| is the error.
// Propagate NaNs (possibly non-canonical).
__ Orpd(result_, error_);
// Propagate sign errors and (subtle) quiet NaNs.
__ Subpd(result_, error_);
// Canonicalize NaNs by clearing the payload. Sign is non-deterministic.
__ Cmppd(error_, result_, int8_t{3});
__ Psrlq(error_, byte{13});
__ Andnpd(error_, result_);
__ Movapd(result_, error_);
}
private:
XMMRegister const result_;
XMMRegister const error_;
};
class OutOfLineTruncateDoubleToI final : public OutOfLineCode {
public:
OutOfLineTruncateDoubleToI(CodeGenerator* gen, Register result,
@ -2419,18 +2331,18 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
XMMRegister src1 = i.InputSimd128Register(1),
dst = i.OutputSimd128Register();
DCHECK_EQ(dst, i.InputSimd128Register(0));
// The minpd instruction doesn't propagate NaNs and -0's in its first
// operand. Perform minpd in both orders and compare results. Handle the
// unlikely case of discrepancies out of line.
// The minpd instruction doesn't propagate NaNs and +0's in its first
// operand. Perform minpd in both orders, merge the resuls, and adjust.
__ Movapd(kScratchDoubleReg, src1);
__ Minpd(kScratchDoubleReg, dst);
__ Minpd(dst, src1);
// Most likely there is no difference and we're done.
__ Xorpd(kScratchDoubleReg, dst);
__ Ptest(kScratchDoubleReg, kScratchDoubleReg);
auto ool = zone()->New<OutOfLineF64x2Min>(this, dst, kScratchDoubleReg);
__ j(not_zero, ool->entry());
__ bind(ool->exit());
// propagate -0's and NaNs, which may be non-canonical.
__ Orpd(kScratchDoubleReg, dst);
// Canonicalize NaNs by quieting and clearing the payload.
__ Cmppd(dst, kScratchDoubleReg, int8_t{3});
__ Orpd(kScratchDoubleReg, dst);
__ Psrlq(dst, 13);
__ Andnpd(dst, kScratchDoubleReg);
break;
}
case kX64F64x2Max: {
@ -2438,17 +2350,20 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
dst = i.OutputSimd128Register();
DCHECK_EQ(dst, i.InputSimd128Register(0));
// The maxpd instruction doesn't propagate NaNs and +0's in its first
// operand. Perform maxpd in both orders and compare results. Handle the
// unlikely case of discrepancies out of line.
// operand. Perform maxpd in both orders, merge the resuls, and adjust.
__ Movapd(kScratchDoubleReg, src1);
__ Maxpd(kScratchDoubleReg, dst);
__ Maxpd(dst, src1);
// Most likely there is no difference and we're done.
__ Xorpd(kScratchDoubleReg, dst);
__ Ptest(kScratchDoubleReg, kScratchDoubleReg);
auto ool = zone()->New<OutOfLineF64x2Max>(this, dst, kScratchDoubleReg);
__ j(not_zero, ool->entry());
__ bind(ool->exit());
// Find discrepancies.
__ Xorpd(dst, kScratchDoubleReg);
// Propagate NaNs, which may be non-canonical.
__ Orpd(kScratchDoubleReg, dst);
// Propagate sign discrepancy and (subtle) quiet NaNs.
__ Subpd(kScratchDoubleReg, dst);
// Canonicalize NaNs by clearing the payload. Sign is non-deterministic.
__ Cmppd(dst, kScratchDoubleReg, int8_t{3});
__ Psrlq(dst, 13);
__ Andnpd(dst, kScratchDoubleReg);
break;
}
case kX64F64x2Eq: {
@ -2612,18 +2527,18 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
XMMRegister src1 = i.InputSimd128Register(1),
dst = i.OutputSimd128Register();
DCHECK_EQ(dst, i.InputSimd128Register(0));
// The minps instruction doesn't propagate NaNs and -0's in its first
// operand. Perform minps in both orders and compare results. Handle the
// unlikely case of discrepancies out of line.
// The minps instruction doesn't propagate NaNs and +0's in its first
// operand. Perform minps in both orders, merge the resuls, and adjust.
__ Movaps(kScratchDoubleReg, src1);
__ Minps(kScratchDoubleReg, dst);
__ Minps(dst, src1);
// Most likely there is no difference and we're done.
__ Xorps(kScratchDoubleReg, dst);
__ Ptest(kScratchDoubleReg, kScratchDoubleReg);
auto ool = zone()->New<OutOfLineF32x4Min>(this, dst, kScratchDoubleReg);
__ j(not_zero, ool->entry());
__ bind(ool->exit());
// propagate -0's and NaNs, which may be non-canonical.
__ Orps(kScratchDoubleReg, dst);
// Canonicalize NaNs by quieting and clearing the payload.
__ Cmpps(dst, kScratchDoubleReg, int8_t{3});
__ Orps(kScratchDoubleReg, dst);
__ Psrld(dst, byte{10});
__ Andnps(dst, kScratchDoubleReg);
break;
}
case kX64F32x4Max: {
@ -2631,17 +2546,20 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
dst = i.OutputSimd128Register();
DCHECK_EQ(dst, i.InputSimd128Register(0));
// The maxps instruction doesn't propagate NaNs and +0's in its first
// operand. Perform maxps in both orders and compare results. Handle the
// unlikely case of discrepancies out of line.
// operand. Perform maxps in both orders, merge the resuls, and adjust.
__ Movaps(kScratchDoubleReg, src1);
__ Maxps(kScratchDoubleReg, dst);
__ Maxps(dst, src1);
// Most likely there is no difference and we're done.
__ Xorps(kScratchDoubleReg, dst);
__ Ptest(kScratchDoubleReg, kScratchDoubleReg);
auto ool = zone()->New<OutOfLineF32x4Max>(this, dst, kScratchDoubleReg);
__ j(not_zero, ool->entry());
__ bind(ool->exit());
// Find discrepancies.
__ Xorps(dst, kScratchDoubleReg);
// Propagate NaNs, which may be non-canonical.
__ Orps(kScratchDoubleReg, dst);
// Propagate sign discrepancy and (subtle) quiet NaNs.
__ Subps(kScratchDoubleReg, dst);
// Canonicalize NaNs by clearing the payload. Sign is non-deterministic.
__ Cmpps(dst, kScratchDoubleReg, int8_t{3});
__ Psrld(dst, byte{10});
__ Andnps(dst, kScratchDoubleReg);
break;
}
case kX64F32x4Eq: {