summaryrefslogtreecommitdiff
path: root/compiler/optimizing/intrinsics_x86.cc
diff options
context:
space:
mode:
Diffstat (limited to 'compiler/optimizing/intrinsics_x86.cc')
-rw-r--r--compiler/optimizing/intrinsics_x86.cc116
1 files changed, 81 insertions, 35 deletions
diff --git a/compiler/optimizing/intrinsics_x86.cc b/compiler/optimizing/intrinsics_x86.cc
index acc40bc998..529f678761 100644
--- a/compiler/optimizing/intrinsics_x86.cc
+++ b/compiler/optimizing/intrinsics_x86.cc
@@ -37,10 +37,12 @@ namespace x86 {
static constexpr int kDoubleNaNHigh = 0x7FF80000;
static constexpr int kDoubleNaNLow = 0x00000000;
-static constexpr int kFloatNaN = 0x7FC00000;
+static constexpr int64_t kDoubleNaN = INT64_C(0x7FF8000000000000);
+static constexpr int32_t kFloatNaN = INT32_C(0x7FC00000);
IntrinsicLocationsBuilderX86::IntrinsicLocationsBuilderX86(CodeGeneratorX86* codegen)
- : arena_(codegen->GetGraph()->GetArena()), codegen_(codegen) {
+ : arena_(codegen->GetGraph()->GetArena()),
+ codegen_(codegen) {
}
@@ -256,15 +258,37 @@ static void CreateFloatToFloat(ArenaAllocator* arena, HInvoke* invoke) {
LocationSummary::kNoCall,
kIntrinsified);
locations->SetInAt(0, Location::RequiresFpuRegister());
- // TODO: Allow x86 to work with memory. This requires assembler support, see below.
- // locations->SetInAt(0, Location::Any()); // X86 can work on memory directly.
locations->SetOut(Location::SameAsFirstInput());
+ HInvokeStaticOrDirect* static_or_direct = invoke->AsInvokeStaticOrDirect();
+ DCHECK(static_or_direct != nullptr);
+ if (invoke->InputAt(static_or_direct->GetSpecialInputIndex())->IsX86ComputeBaseMethodAddress()) {
+ // We need addressibility for the constant area.
+ locations->SetInAt(1, Location::RequiresRegister());
+ // We need a temporary to hold the constant.
+ locations->AddTemp(Location::RequiresFpuRegister());
+ }
}
-static void MathAbsFP(LocationSummary* locations, bool is64bit, X86Assembler* assembler) {
+static void MathAbsFP(LocationSummary* locations,
+ bool is64bit,
+ X86Assembler* assembler,
+ CodeGeneratorX86* codegen) {
Location output = locations->Out();
- if (output.IsFpuRegister()) {
+ DCHECK(output.IsFpuRegister());
+ if (locations->InAt(1).IsValid()) {
+ DCHECK(locations->InAt(1).IsRegister());
+ // We also have a constant area pointer.
+ Register constant_area = locations->InAt(1).AsRegister<Register>();
+ XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
+ if (is64bit) {
+ __ movsd(temp, codegen->LiteralInt64Address(INT64_C(0x7FFFFFFFFFFFFFFF), constant_area));
+ __ andpd(output.AsFpuRegister<XmmRegister>(), temp);
+ } else {
+ __ movss(temp, codegen->LiteralInt32Address(INT32_C(0x7FFFFFFF), constant_area));
+ __ andps(output.AsFpuRegister<XmmRegister>(), temp);
+ }
+ } else {
// Create the right constant on an aligned stack.
if (is64bit) {
__ subl(ESP, Immediate(8));
@@ -277,19 +301,6 @@ static void MathAbsFP(LocationSummary* locations, bool is64bit, X86Assembler* as
__ andps(output.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
}
__ addl(ESP, Immediate(16));
- } else {
- // TODO: update when assember support is available.
- UNIMPLEMENTED(FATAL) << "Needs assembler support.";
-// Once assembler support is available, in-memory operations look like this:
-// if (is64bit) {
-// DCHECK(output.IsDoubleStackSlot());
-// __ andl(Address(Register(RSP), output.GetHighStackIndex(kX86WordSize)),
-// Immediate(0x7FFFFFFF));
-// } else {
-// DCHECK(output.IsStackSlot());
-// // Can use and with a literal directly.
-// __ andl(Address(Register(RSP), output.GetStackIndex()), Immediate(0x7FFFFFFF));
-// }
}
}
@@ -298,7 +309,7 @@ void IntrinsicLocationsBuilderX86::VisitMathAbsDouble(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathAbsDouble(HInvoke* invoke) {
- MathAbsFP(invoke->GetLocations(), /* is64bit */ true, GetAssembler());
+ MathAbsFP(invoke->GetLocations(), /* is64bit */ true, GetAssembler(), codegen_);
}
void IntrinsicLocationsBuilderX86::VisitMathAbsFloat(HInvoke* invoke) {
@@ -306,7 +317,7 @@ void IntrinsicLocationsBuilderX86::VisitMathAbsFloat(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathAbsFloat(HInvoke* invoke) {
- MathAbsFP(invoke->GetLocations(), /* is64bit */ false, GetAssembler());
+ MathAbsFP(invoke->GetLocations(), /* is64bit */ false, GetAssembler(), codegen_);
}
static void CreateAbsIntLocation(ArenaAllocator* arena, HInvoke* invoke) {
@@ -388,8 +399,11 @@ void IntrinsicCodeGeneratorX86::VisitMathAbsLong(HInvoke* invoke) {
GenAbsLong(invoke->GetLocations(), GetAssembler());
}
-static void GenMinMaxFP(LocationSummary* locations, bool is_min, bool is_double,
- X86Assembler* assembler) {
+static void GenMinMaxFP(LocationSummary* locations,
+ bool is_min,
+ bool is_double,
+ X86Assembler* assembler,
+ CodeGeneratorX86* codegen) {
Location op1_loc = locations->InAt(0);
Location op2_loc = locations->InAt(1);
Location out_loc = locations->Out();
@@ -450,15 +464,26 @@ static void GenMinMaxFP(LocationSummary* locations, bool is_min, bool is_double,
// NaN handling.
__ Bind(&nan);
- if (is_double) {
- __ pushl(Immediate(kDoubleNaNHigh));
- __ pushl(Immediate(kDoubleNaNLow));
- __ movsd(out, Address(ESP, 0));
- __ addl(ESP, Immediate(8));
+ // Do we have a constant area pointer?
+ if (locations->InAt(2).IsValid()) {
+ DCHECK(locations->InAt(2).IsRegister());
+ Register constant_area = locations->InAt(2).AsRegister<Register>();
+ if (is_double) {
+ __ movsd(out, codegen->LiteralInt64Address(kDoubleNaN, constant_area));
+ } else {
+ __ movss(out, codegen->LiteralInt32Address(kFloatNaN, constant_area));
+ }
} else {
- __ pushl(Immediate(kFloatNaN));
- __ movss(out, Address(ESP, 0));
- __ addl(ESP, Immediate(4));
+ if (is_double) {
+ __ pushl(Immediate(kDoubleNaNHigh));
+ __ pushl(Immediate(kDoubleNaNLow));
+ __ movsd(out, Address(ESP, 0));
+ __ addl(ESP, Immediate(8));
+ } else {
+ __ pushl(Immediate(kFloatNaN));
+ __ movss(out, Address(ESP, 0));
+ __ addl(ESP, Immediate(4));
+ }
}
__ jmp(&done);
@@ -483,6 +508,11 @@ static void CreateFPFPToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
// The following is sub-optimal, but all we can do for now. It would be fine to also accept
// the second input to be the output (we can simply swap inputs).
locations->SetOut(Location::SameAsFirstInput());
+ HInvokeStaticOrDirect* static_or_direct = invoke->AsInvokeStaticOrDirect();
+ DCHECK(static_or_direct != nullptr);
+ if (invoke->InputAt(static_or_direct->GetSpecialInputIndex())->IsX86ComputeBaseMethodAddress()) {
+ locations->SetInAt(2, Location::RequiresRegister());
+ }
}
void IntrinsicLocationsBuilderX86::VisitMathMinDoubleDouble(HInvoke* invoke) {
@@ -490,7 +520,11 @@ void IntrinsicLocationsBuilderX86::VisitMathMinDoubleDouble(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathMinDoubleDouble(HInvoke* invoke) {
- GenMinMaxFP(invoke->GetLocations(), /* is_min */ true, /* is_double */ true, GetAssembler());
+ GenMinMaxFP(invoke->GetLocations(),
+ /* is_min */ true,
+ /* is_double */ true,
+ GetAssembler(),
+ codegen_);
}
void IntrinsicLocationsBuilderX86::VisitMathMinFloatFloat(HInvoke* invoke) {
@@ -498,7 +532,11 @@ void IntrinsicLocationsBuilderX86::VisitMathMinFloatFloat(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathMinFloatFloat(HInvoke* invoke) {
- GenMinMaxFP(invoke->GetLocations(), /* is_min */ true, /* is_double */ false, GetAssembler());
+ GenMinMaxFP(invoke->GetLocations(),
+ /* is_min */ true,
+ /* is_double */ false,
+ GetAssembler(),
+ codegen_);
}
void IntrinsicLocationsBuilderX86::VisitMathMaxDoubleDouble(HInvoke* invoke) {
@@ -506,7 +544,11 @@ void IntrinsicLocationsBuilderX86::VisitMathMaxDoubleDouble(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathMaxDoubleDouble(HInvoke* invoke) {
- GenMinMaxFP(invoke->GetLocations(), /* is_min */ false, /* is_double */ true, GetAssembler());
+ GenMinMaxFP(invoke->GetLocations(),
+ /* is_min */ false,
+ /* is_double */ true,
+ GetAssembler(),
+ codegen_);
}
void IntrinsicLocationsBuilderX86::VisitMathMaxFloatFloat(HInvoke* invoke) {
@@ -514,7 +556,11 @@ void IntrinsicLocationsBuilderX86::VisitMathMaxFloatFloat(HInvoke* invoke) {
}
void IntrinsicCodeGeneratorX86::VisitMathMaxFloatFloat(HInvoke* invoke) {
- GenMinMaxFP(invoke->GetLocations(), /* is_min */ false, /* is_double */ false, GetAssembler());
+ GenMinMaxFP(invoke->GetLocations(),
+ /* is_min */ false,
+ /* is_double */ false,
+ GetAssembler(),
+ codegen_);
}
static void GenMinMax(LocationSummary* locations, bool is_min, bool is_long,
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2025-10-21 01:02:36 +0000