Reviewers: ulan, jochen (OOO Wed-Thu),
Message:
This speeds up Kraken by about 2.3-2.7%. Most of that improvement comes from
desaturate and fft.
Description:
ARM64: Use default-NaN mode to canonicalize NaNs.
BUG=
Please review this at https://codereview.chromium.org/255343004/
SVN Base: https://v8.googlecode.com/svn/branches/bleeding_edge
Affected files (+76, -14 lines):
M src/arm64/code-stubs-arm64.cc
M src/arm64/ic-arm64.cc
M src/arm64/lithium-arm64.h
M src/arm64/lithium-codegen-arm64.cc
M src/arm64/macro-assembler-arm64.h
M src/arm64/macro-assembler-arm64.cc
M src/objects.h
Index: src/arm64/code-stubs-arm64.cc
diff --git a/src/arm64/code-stubs-arm64.cc b/src/arm64/code-stubs-arm64.cc
index
884be5800eeddc32b014eae4b21f249457578282..5bffd19c1a26c3c27a3c5adcb44e807e9d4781bb
100644
--- a/src/arm64/code-stubs-arm64.cc
+++ b/src/arm64/code-stubs-arm64.cc
@@ -1577,6 +1577,7 @@ void CEntryStub::Generate(MacroAssembler* masm) {
// jssp[8]: Preserved x22 (used for argc).
// jssp[0]: Preserved x21 (used for argv).
__ Drop(x11);
+ __ AssertFPCRState();
__ Ret();
// The stack pointer is still csp if we aren't returning, and the frame
@@ -1660,6 +1661,11 @@ void JSEntryStub::GenerateBody(MacroAssembler* masm,
bool is_construct) {
__ Mov(jssp, csp);
__ SetStackPointer(jssp);
+ // Configure the FPCR. We don't restore it, so this is technically not
allowed
+ // according to AAPCS64. However, we only set default-NaN mode and this
will
+ // be harmless for most C code. Also, it works for ARM.
+ __ ConfigureFPCR();
+
ProfileEntryHookStub::MaybeCallEntryHook(masm);
// Set up the reserved register for 0.0.
@@ -4519,7 +4525,7 @@ void
StoreArrayLiteralElementStub::Generate(MacroAssembler* masm) {
__ Bind(&double_elements);
__ Ldr(x10, FieldMemOperand(array, JSObject::kElementsOffset));
- __ StoreNumberToDoubleElements(value, index_smi, x10, x11, d0, d1,
+ __ StoreNumberToDoubleElements(value, index_smi, x10, x11, d0,
&slow_elements);
__ Ret();
}
@@ -4621,6 +4627,7 @@ void DirectCEntryStub::Generate(MacroAssembler* masm)
{
__ Blr(x10);
// Return to calling code.
__ Peek(lr, 0);
+ __ AssertFPCRState();
__ Ret();
__ SetStackPointer(old_stack_pointer);
Index: src/arm64/ic-arm64.cc
diff --git a/src/arm64/ic-arm64.cc b/src/arm64/ic-arm64.cc
index
a0a1f03113ba6d60ca82f2069e80a7e3344e53ab..c09b847ba5d469f4cf40177e29c4d605fb0b916a
100644
--- a/src/arm64/ic-arm64.cc
+++ b/src/arm64/ic-arm64.cc
@@ -1021,7 +1021,6 @@ static void KeyedStoreGenerateGenericHelper(
elements,
x10,
d0,
- d1,
&transition_double_elements);
if (increment_length == kIncrementLength) {
// Add 1 to receiver->length.
Index: src/arm64/lithium-arm64.h
diff --git a/src/arm64/lithium-arm64.h b/src/arm64/lithium-arm64.h
index
b036446746ce391c460b3ca08a081422e44492b7..da91bca7bd060291f969494d2dcba4550f3a37fa
100644
--- a/src/arm64/lithium-arm64.h
+++ b/src/arm64/lithium-arm64.h
@@ -2361,6 +2361,10 @@ class LStoreKeyed : public LTemplateInstruction<0,
3, T> {
}
bool NeedsCanonicalization() {
+ if (hydrogen()->value()->IsAdd() || hydrogen()->value()->IsSub() ||
+ hydrogen()->value()->IsMul() || hydrogen()->value()->IsDiv()) {
+ return false;
+ }
return this->hydrogen()->NeedsCanonicalization();
}
uint32_t additional_index() const { return
this->hydrogen()->index_offset(); }
Index: src/arm64/lithium-codegen-arm64.cc
diff --git a/src/arm64/lithium-codegen-arm64.cc
b/src/arm64/lithium-codegen-arm64.cc
index
3ecd2403ba3b6aa0848006d90439a11b746e0bc3..10e0f080a4a6570f6ac897b798c777aab1b23934
100644
--- a/src/arm64/lithium-codegen-arm64.cc
+++ b/src/arm64/lithium-codegen-arm64.cc
@@ -5100,11 +5100,8 @@ void
LCodeGen::DoStoreKeyedFixedDouble(LStoreKeyedFixedDouble* instr) {
}
if (instr->NeedsCanonicalization()) {
- DoubleRegister dbl_scratch = double_scratch();
- __ Fmov(dbl_scratch,
- FixedDoubleArray::canonical_not_the_hole_nan_as_double());
- __ Fmaxnm(dbl_scratch, dbl_scratch, value);
- __ Str(dbl_scratch, FieldMemOperand(store_base, offset));
+ __ CanonicalizeNaN(double_scratch(), value);
+ __ Str(double_scratch(), FieldMemOperand(store_base, offset));
} else {
__ Str(value, FieldMemOperand(store_base, offset));
}
Index: src/arm64/macro-assembler-arm64.cc
diff --git a/src/arm64/macro-assembler-arm64.cc
b/src/arm64/macro-assembler-arm64.cc
index
3f84321a5e7057781a478d76f11e1c8f3e76dfd9..4d277b72c966ad3e95efd279e7825ee4bcb7bd41
100644
--- a/src/arm64/macro-assembler-arm64.cc
+++ b/src/arm64/macro-assembler-arm64.cc
@@ -1222,6 +1222,58 @@ void MacroAssembler::AssertStackConsistency() {
}
+void MacroAssembler::AssertFPCRState(Register fpcr) {
+ if (emit_debug_code()) {
+ Label unexpected_mode, done;
+ UseScratchRegisterScope temps(this);
+ if (fpcr.IsNone()) {
+ fpcr = temps.AcquireX();
+ Mrs(fpcr, FPCR);
+ }
+
+ // Assert that default-NaN mode is set.
+ Tbz(fpcr, DN_offset, &unexpected_mode);
+
+ // Assert that flush-to-zero is not set.
+ Tbnz(fpcr, FZ_offset, &unexpected_mode);
+
+ // Assert that the rounding mode is nearest-with-ties-to-even.
+ STATIC_ASSERT(FPTieEven == 0);
+ Tst(fpcr, RMode_mask);
+ B(eq, &done);
+
+ Bind(&unexpected_mode);
+ Abort(kUnexpectedFPCRMode);
+
+ Bind(&done);
+ }
+}
+
+
+void MacroAssembler::ConfigureFPCR() {
+ UseScratchRegisterScope temps(this);
+ Register fpcr = temps.AcquireX();
+ Mrs(fpcr, FPCR);
+
+ // If necessary, enable default-NaN mode. The default values of the
other FPCR
+ // options should be suitable.
+ Label no_write_required;
+ Tbnz(fpcr, DN_offset, &no_write_required);
+
+ Orr(fpcr, fpcr, DN_mask);
+ Msr(FPCR, fpcr);
+
+ Bind(&no_write_required);
+ AssertFPCRState(fpcr);
+}
+
+
+void MacroAssembler::CanonicalizeNaN(const FPRegister& dst,
+ const FPRegister& src) {
+ Fsub(dst, src, fp_zero);
+}
+
+
void MacroAssembler::LoadRoot(CPURegister destination,
Heap::RootListIndex index) {
// TODO(jbramley): Most root values are constants, and can be synthesized
@@ -3888,7 +3940,6 @@ void
MacroAssembler::StoreNumberToDoubleElements(Register value_reg,
Register elements_reg,
Register scratch1,
FPRegister fpscratch1,
- FPRegister fpscratch2,
Label* fail,
int elements_offset) {
ASSERT(!AreAliased(value_reg, key_reg, elements_reg, scratch1));
@@ -3906,12 +3957,9 @@ void
MacroAssembler::StoreNumberToDoubleElements(Register value_reg,
fail, DONT_DO_SMI_CHECK);
Ldr(fpscratch1, FieldMemOperand(value_reg, HeapNumber::kValueOffset));
- Fmov(fpscratch2,
FixedDoubleArray::canonical_not_the_hole_nan_as_double());
- // Check for NaN by comparing the number to itself: NaN comparison will
- // report unordered, indicated by the overflow flag being set.
- Fcmp(fpscratch1, fpscratch1);
- Fcsel(fpscratch1, fpscratch2, fpscratch1, vs);
+ // Canonicalize NaNs.
+ CanonicalizeNaN(fpscratch1);
// Store the result.
Bind(&store_num);
Index: src/arm64/macro-assembler-arm64.h
diff --git a/src/arm64/macro-assembler-arm64.h
b/src/arm64/macro-assembler-arm64.h
index
965d1dbfd9d5ef125b40d84ad83063079d40fcc3..265a0cef1a5acf3f8b65b1d4ae74104a9521c009
100644
--- a/src/arm64/macro-assembler-arm64.h
+++ b/src/arm64/macro-assembler-arm64.h
@@ -788,6 +788,13 @@ class MacroAssembler : public Assembler {
// Root register.
inline void InitializeRootRegister();
+ void AssertFPCRState(Register fpcr = NoReg);
+ void ConfigureFPCR();
+ void CanonicalizeNaN(const FPRegister& dst, const FPRegister& src);
+ void CanonicalizeNaN(const FPRegister& reg) {
+ CanonicalizeNaN(reg, reg);
+ }
+
// Load an object from the root table.
void LoadRoot(CPURegister destination,
Heap::RootListIndex index);
@@ -1533,7 +1540,6 @@ class MacroAssembler : public Assembler {
Register elements_reg,
Register scratch1,
FPRegister fpscratch1,
- FPRegister fpscratch2,
Label* fail,
int elements_offset = 0);
Index: src/objects.h
diff --git a/src/objects.h b/src/objects.h
index
6402f86d341919e6f16fb3ba33db96d819ae5458..2b51b81c990990e3262608226d7edae960be6b9f
100644
--- a/src/objects.h
+++ b/src/objects.h
@@ -1333,6 +1333,7 @@ class MaybeObject BASE_EMBEDDED {
V(kUnexpectedNegativeValue, "Unexpected negative
value") \
V(kUnexpectedNumberOfPreAllocatedPropertyFields,
\
"Unexpected number of pre-allocated property
fields") \
+ V(kUnexpectedFPCRMode, "Unexpected FPCR
mode.") \
V(kUnexpectedSmi, "Unexpected smi
value") \
V(kUnexpectedStringFunction, "Unexpected String
function") \
V(kUnexpectedStringType, "Unexpected string
type") \
--
--
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.