blob: 0212da106b2223682f5965787e448cc4f443b955 [file] [log] [blame]
/*
* Copyright (C) 2014 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "code_generator_x86.h"
#include "code_generator_utils.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "entrypoints/quick/quick_entrypoints_enum.h"
#include "gc/accounting/card_table.h"
#include "intrinsics.h"
#include "intrinsics_x86.h"
#include "mirror/array-inl.h"
#include "mirror/art_method.h"
#include "mirror/class.h"
#include "thread.h"
#include "utils/assembler.h"
#include "utils/stack_checks.h"
#include "utils/x86/assembler_x86.h"
#include "utils/x86/managed_register_x86.h"
namespace art {
namespace x86 {
static constexpr int kCurrentMethodStackOffset = 0;
static constexpr Register kCoreCalleeSaves[] = { EBP, ESI, EDI };
static constexpr int kC2ConditionMask = 0x400;
static constexpr int kFakeReturnRegister = Register(8);
#define __ reinterpret_cast<X86Assembler*>(codegen->GetAssembler())->
class NullCheckSlowPathX86 : public SlowPathCodeX86 {
public:
explicit NullCheckSlowPathX86(HNullCheck* instruction) : instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
__ Bind(GetEntryLabel());
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowNullPointer)));
RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
}
private:
HNullCheck* const instruction_;
DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathX86);
};
class DivZeroCheckSlowPathX86 : public SlowPathCodeX86 {
public:
explicit DivZeroCheckSlowPathX86(HDivZeroCheck* instruction) : instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
__ Bind(GetEntryLabel());
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowDivZero)));
RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
}
private:
HDivZeroCheck* const instruction_;
DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathX86);
};
class DivRemMinusOneSlowPathX86 : public SlowPathCodeX86 {
public:
explicit DivRemMinusOneSlowPathX86(Register reg, bool is_div) : reg_(reg), is_div_(is_div) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
__ Bind(GetEntryLabel());
if (is_div_) {
__ negl(reg_);
} else {
__ movl(reg_, Immediate(0));
}
__ jmp(GetExitLabel());
}
private:
Register reg_;
bool is_div_;
DISALLOW_COPY_AND_ASSIGN(DivRemMinusOneSlowPathX86);
};
class BoundsCheckSlowPathX86 : public SlowPathCodeX86 {
public:
BoundsCheckSlowPathX86(HBoundsCheck* instruction,
Location index_location,
Location length_location)
: instruction_(instruction),
index_location_(index_location),
length_location_(length_location) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
__ Bind(GetEntryLabel());
// We're moving two locations to locations that could overlap, so we need a parallel
// move resolver.
InvokeRuntimeCallingConvention calling_convention;
x86_codegen->EmitParallelMoves(
index_location_,
Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
Primitive::kPrimInt,
length_location_,
Location::RegisterLocation(calling_convention.GetRegisterAt(1)),
Primitive::kPrimInt);
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowArrayBounds)));
RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
}
private:
HBoundsCheck* const instruction_;
const Location index_location_;
const Location length_location_;
DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathX86);
};
class SuspendCheckSlowPathX86 : public SlowPathCodeX86 {
public:
SuspendCheckSlowPathX86(HSuspendCheck* instruction, HBasicBlock* successor)
: instruction_(instruction), successor_(successor) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, instruction_->GetLocations());
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pTestSuspend)));
RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
RestoreLiveRegisters(codegen, instruction_->GetLocations());
if (successor_ == nullptr) {
__ jmp(GetReturnLabel());
} else {
__ jmp(x86_codegen->GetLabelOf(successor_));
}
}
Label* GetReturnLabel() {
DCHECK(successor_ == nullptr);
return &return_label_;
}
HBasicBlock* GetSuccessor() const {
return successor_;
}
private:
HSuspendCheck* const instruction_;
HBasicBlock* const successor_;
Label return_label_;
DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathX86);
};
class LoadStringSlowPathX86 : public SlowPathCodeX86 {
public:
explicit LoadStringSlowPathX86(HLoadString* instruction) : instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
__ movl(calling_convention.GetRegisterAt(0), Immediate(instruction_->GetStringIndex()));
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pResolveString)));
RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX));
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
private:
HLoadString* const instruction_;
DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathX86);
};
class LoadClassSlowPathX86 : public SlowPathCodeX86 {
public:
LoadClassSlowPathX86(HLoadClass* cls,
HInstruction* at,
uint32_t dex_pc,
bool do_clinit)
: cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) {
DCHECK(at->IsLoadClass() || at->IsClinitCheck());
}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = at_->GetLocations();
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
__ movl(calling_convention.GetRegisterAt(0), Immediate(cls_->GetTypeIndex()));
__ fs()->call(Address::Absolute(do_clinit_
? QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pInitializeStaticStorage)
: QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pInitializeType)));
RecordPcInfo(codegen, at_, dex_pc_);
// Move the class to the desired location.
Location out = locations->Out();
if (out.IsValid()) {
DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
x86_codegen->Move32(out, Location::RegisterLocation(EAX));
}
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
private:
// The class this slow path will load.
HLoadClass* const cls_;
// The instruction where this slow path is happening.
// (Might be the load class or an initialization check).
HInstruction* const at_;
// The dex PC of `at_`.
const uint32_t dex_pc_;
// Whether to initialize the class.
const bool do_clinit_;
DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathX86);
};
class TypeCheckSlowPathX86 : public SlowPathCodeX86 {
public:
TypeCheckSlowPathX86(HInstruction* instruction,
Location class_to_check,
Location object_class,
uint32_t dex_pc)
: instruction_(instruction),
class_to_check_(class_to_check),
object_class_(object_class),
dex_pc_(dex_pc) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
DCHECK(instruction_->IsCheckCast()
|| !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
// We're moving two locations to locations that could overlap, so we need a parallel
// move resolver.
InvokeRuntimeCallingConvention calling_convention;
x86_codegen->EmitParallelMoves(
class_to_check_,
Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
Primitive::kPrimNot,
object_class_,
Location::RegisterLocation(calling_convention.GetRegisterAt(1)),
Primitive::kPrimNot);
if (instruction_->IsInstanceOf()) {
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize,
pInstanceofNonTrivial)));
} else {
DCHECK(instruction_->IsCheckCast());
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pCheckCast)));
}
RecordPcInfo(codegen, instruction_, dex_pc_);
if (instruction_->IsInstanceOf()) {
x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX));
}
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
private:
HInstruction* const instruction_;
const Location class_to_check_;
const Location object_class_;
const uint32_t dex_pc_;
DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathX86);
};
class DeoptimizationSlowPathX86 : public SlowPathCodeX86 {
public:
explicit DeoptimizationSlowPathX86(HInstruction* instruction)
: instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, instruction_->GetLocations());
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pDeoptimize)));
// No need to restore live registers.
DCHECK(instruction_->IsDeoptimize());
HDeoptimize* deoptimize = instruction_->AsDeoptimize();
uint32_t dex_pc = deoptimize->GetDexPc();
codegen->RecordPcInfo(instruction_, dex_pc, this);
}
private:
HInstruction* const instruction_;
DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathX86);
};
#undef __
#define __ reinterpret_cast<X86Assembler*>(GetAssembler())->
inline Condition X86Condition(IfCondition cond) {
switch (cond) {
case kCondEQ: return kEqual;
case kCondNE: return kNotEqual;
case kCondLT: return kLess;
case kCondLE: return kLessEqual;
case kCondGT: return kGreater;
case kCondGE: return kGreaterEqual;
default:
LOG(FATAL) << "Unknown if condition";
}
return kEqual;
}
void CodeGeneratorX86::DumpCoreRegister(std::ostream& stream, int reg) const {
stream << Register(reg);
}
void CodeGeneratorX86::DumpFloatingPointRegister(std::ostream& stream, int reg) const {
stream << XmmRegister(reg);
}
size_t CodeGeneratorX86::SaveCoreRegister(size_t stack_index, uint32_t reg_id) {
__ movl(Address(ESP, stack_index), static_cast<Register>(reg_id));
return kX86WordSize;
}
size_t CodeGeneratorX86::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) {
__ movl(static_cast<Register>(reg_id), Address(ESP, stack_index));
return kX86WordSize;
}
size_t CodeGeneratorX86::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
__ movsd(Address(ESP, stack_index), XmmRegister(reg_id));
return GetFloatingPointSpillSlotSize();
}
size_t CodeGeneratorX86::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
__ movsd(XmmRegister(reg_id), Address(ESP, stack_index));
return GetFloatingPointSpillSlotSize();
}
CodeGeneratorX86::CodeGeneratorX86(HGraph* graph,
const X86InstructionSetFeatures& isa_features,
const CompilerOptions& compiler_options)
: CodeGenerator(graph,
kNumberOfCpuRegisters,
kNumberOfXmmRegisters,
kNumberOfRegisterPairs,
ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves),
arraysize(kCoreCalleeSaves))
| (1 << kFakeReturnRegister),
0,
compiler_options),
block_labels_(graph->GetArena(), 0),
location_builder_(graph, this),
instruction_visitor_(graph, this),
move_resolver_(graph->GetArena(), this),
isa_features_(isa_features) {
// Use a fake return address register to mimic Quick.
AddAllocatedRegister(Location::RegisterLocation(kFakeReturnRegister));
}
Location CodeGeneratorX86::AllocateFreeRegister(Primitive::Type type) const {
switch (type) {
case Primitive::kPrimLong: {
size_t reg = FindFreeEntry(blocked_register_pairs_, kNumberOfRegisterPairs);
X86ManagedRegister pair =
X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(reg));
DCHECK(!blocked_core_registers_[pair.AsRegisterPairLow()]);
DCHECK(!blocked_core_registers_[pair.AsRegisterPairHigh()]);
blocked_core_registers_[pair.AsRegisterPairLow()] = true;
blocked_core_registers_[pair.AsRegisterPairHigh()] = true;
UpdateBlockedPairRegisters();
return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh());
}
case Primitive::kPrimByte:
case Primitive::kPrimBoolean:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot: {
Register reg = static_cast<Register>(
FindFreeEntry(blocked_core_registers_, kNumberOfCpuRegisters));
// Block all register pairs that contain `reg`.
for (int i = 0; i < kNumberOfRegisterPairs; i++) {
X86ManagedRegister current =
X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i));
if (current.AsRegisterPairLow() == reg || current.AsRegisterPairHigh() == reg) {
blocked_register_pairs_[i] = true;
}
}
return Location::RegisterLocation(reg);
}
case Primitive::kPrimFloat:
case Primitive::kPrimDouble: {
return Location::FpuRegisterLocation(
FindFreeEntry(blocked_fpu_registers_, kNumberOfXmmRegisters));
}
case Primitive::kPrimVoid:
LOG(FATAL) << "Unreachable type " << type;
}
return Location();
}
void CodeGeneratorX86::SetupBlockedRegisters(bool is_baseline) const {
// Don't allocate the dalvik style register pair passing.
blocked_register_pairs_[ECX_EDX] = true;
// Stack register is always reserved.
blocked_core_registers_[ESP] = true;
if (is_baseline) {
blocked_core_registers_[EBP] = true;
blocked_core_registers_[ESI] = true;
blocked_core_registers_[EDI] = true;
}
UpdateBlockedPairRegisters();
}
void CodeGeneratorX86::UpdateBlockedPairRegisters() const {
for (int i = 0; i < kNumberOfRegisterPairs; i++) {
X86ManagedRegister current =
X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i));
if (blocked_core_registers_[current.AsRegisterPairLow()]
|| blocked_core_registers_[current.AsRegisterPairHigh()]) {
blocked_register_pairs_[i] = true;
}
}
}
InstructionCodeGeneratorX86::InstructionCodeGeneratorX86(HGraph* graph, CodeGeneratorX86* codegen)
: HGraphVisitor(graph),
assembler_(codegen->GetAssembler()),
codegen_(codegen) {}
static dwarf::Reg DWARFReg(Register reg) {
return dwarf::Reg::X86Core(static_cast<int>(reg));
}
void CodeGeneratorX86::GenerateFrameEntry() {
__ cfi().SetCurrentCFAOffset(kX86WordSize); // return address
__ Bind(&frame_entry_label_);
bool skip_overflow_check =
IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kX86);
DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks());
if (!skip_overflow_check) {
__ testl(EAX, Address(ESP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kX86))));
RecordPcInfo(nullptr, 0);
}
if (HasEmptyFrame()) {
return;
}
for (int i = arraysize(kCoreCalleeSaves) - 1; i >= 0; --i) {
Register reg = kCoreCalleeSaves[i];
if (allocated_registers_.ContainsCoreRegister(reg)) {
__ pushl(reg);
__ cfi().AdjustCFAOffset(kX86WordSize);
__ cfi().RelOffset(DWARFReg(reg), 0);
}
}
int adjust = GetFrameSize() - FrameEntrySpillSize();
__ subl(ESP, Immediate(adjust));
__ cfi().AdjustCFAOffset(adjust);
__ movl(Address(ESP, kCurrentMethodStackOffset), EAX);
}
void CodeGeneratorX86::GenerateFrameExit() {
__ cfi().RememberState();
if (!HasEmptyFrame()) {
int adjust = GetFrameSize() - FrameEntrySpillSize();
__ addl(ESP, Immediate(adjust));
__ cfi().AdjustCFAOffset(-adjust);
for (size_t i = 0; i < arraysize(kCoreCalleeSaves); ++i) {
Register reg = kCoreCalleeSaves[i];
if (allocated_registers_.ContainsCoreRegister(reg)) {
__ popl(reg);
__ cfi().AdjustCFAOffset(-static_cast<int>(kX86WordSize));
__ cfi().Restore(DWARFReg(reg));
}
}
}
__ ret();
__ cfi().RestoreState();
__ cfi().DefCFAOffset(GetFrameSize());
}
void CodeGeneratorX86::Bind(HBasicBlock* block) {
__ Bind(GetLabelOf(block));
}
void CodeGeneratorX86::LoadCurrentMethod(Register reg) {
DCHECK(RequiresCurrentMethod());
__ movl(reg, Address(ESP, kCurrentMethodStackOffset));
}
Location CodeGeneratorX86::GetStackLocation(HLoadLocal* load) const {
switch (load->GetType()) {
case Primitive::kPrimLong:
case Primitive::kPrimDouble:
return Location::DoubleStackSlot(GetStackSlot(load->GetLocal()));
case Primitive::kPrimInt:
case Primitive::kPrimNot:
case Primitive::kPrimFloat:
return Location::StackSlot(GetStackSlot(load->GetLocal()));
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimVoid:
LOG(FATAL) << "Unexpected type " << load->GetType();
UNREACHABLE();
}
LOG(FATAL) << "Unreachable";
UNREACHABLE();
}
Location InvokeDexCallingConventionVisitorX86::GetNextLocation(Primitive::Type type) {
switch (type) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot: {
uint32_t index = gp_index_++;
stack_index_++;
if (index < calling_convention.GetNumberOfRegisters()) {
return Location::RegisterLocation(calling_convention.GetRegisterAt(index));
} else {
return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1));
}
}
case Primitive::kPrimLong: {
uint32_t index = gp_index_;
gp_index_ += 2;
stack_index_ += 2;
if (index + 1 < calling_convention.GetNumberOfRegisters()) {
X86ManagedRegister pair = X86ManagedRegister::FromRegisterPair(
calling_convention.GetRegisterPairAt(index));
return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh());
} else {
return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2));
}
}
case Primitive::kPrimFloat: {
uint32_t index = float_index_++;
stack_index_++;
if (index < calling_convention.GetNumberOfFpuRegisters()) {
return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index));
} else {
return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1));
}
}
case Primitive::kPrimDouble: {
uint32_t index = float_index_++;
stack_index_ += 2;
if (index < calling_convention.GetNumberOfFpuRegisters()) {
return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index));
} else {
return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2));
}
}
case Primitive::kPrimVoid:
LOG(FATAL) << "Unexpected parameter type " << type;
break;
}
return Location();
}
void CodeGeneratorX86::Move32(Location destination, Location source) {
if (source.Equals(destination)) {
return;
}
if (destination.IsRegister()) {
if (source.IsRegister()) {
__ movl(destination.AsRegister<Register>(), source.AsRegister<Register>());
} else if (source.IsFpuRegister()) {
__ movd(destination.AsRegister<Register>(), source.AsFpuRegister<XmmRegister>());
} else {
DCHECK(source.IsStackSlot());
__ movl(destination.AsRegister<Register>(), Address(ESP, source.GetStackIndex()));
}
} else if (destination.IsFpuRegister()) {
if (source.IsRegister()) {
__ movd(destination.AsFpuRegister<XmmRegister>(), source.AsRegister<Register>());
} else if (source.IsFpuRegister()) {
__ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
} else {
DCHECK(source.IsStackSlot());
__ movss(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
}
} else {
DCHECK(destination.IsStackSlot()) << destination;
if (source.IsRegister()) {
__ movl(Address(ESP, destination.GetStackIndex()), source.AsRegister<Register>());
} else if (source.IsFpuRegister()) {
__ movss(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
} else if (source.IsConstant()) {
HConstant* constant = source.GetConstant();
int32_t value = GetInt32ValueOf(constant);
__ movl(Address(ESP, destination.GetStackIndex()), Immediate(value));
} else {
DCHECK(source.IsStackSlot());
__ pushl(Address(ESP, source.GetStackIndex()));
__ popl(Address(ESP, destination.GetStackIndex()));
}
}
}
void CodeGeneratorX86::Move64(Location destination, Location source) {
if (source.Equals(destination)) {
return;
}
if (destination.IsRegisterPair()) {
if (source.IsRegisterPair()) {
EmitParallelMoves(
Location::RegisterLocation(source.AsRegisterPairHigh<Register>()),
Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()),
Primitive::kPrimInt,
Location::RegisterLocation(source.AsRegisterPairLow<Register>()),
Location::RegisterLocation(destination.AsRegisterPairLow<Register>()),
Primitive::kPrimInt);
} else if (source.IsFpuRegister()) {
LOG(FATAL) << "Unimplemented";
} else {
// No conflict possible, so just do the moves.
DCHECK(source.IsDoubleStackSlot());
__ movl(destination.AsRegisterPairLow<Register>(), Address(ESP, source.GetStackIndex()));
__ movl(destination.AsRegisterPairHigh<Register>(),
Address(ESP, source.GetHighStackIndex(kX86WordSize)));
}
} else if (destination.IsFpuRegister()) {
if (source.IsFpuRegister()) {
__ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
} else if (source.IsDoubleStackSlot()) {
__ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
} else {
LOG(FATAL) << "Unimplemented";
}
} else {
DCHECK(destination.IsDoubleStackSlot()) << destination;
if (source.IsRegisterPair()) {
// No conflict possible, so just do the moves.
__ movl(Address(ESP, destination.GetStackIndex()), source.AsRegisterPairLow<Register>());
__ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)),
source.AsRegisterPairHigh<Register>());
} else if (source.IsFpuRegister()) {
__ movsd(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
} else if (source.IsConstant()) {
HConstant* constant = source.GetConstant();
int64_t value;
if (constant->IsLongConstant()) {
value = constant->AsLongConstant()->GetValue();
} else {
DCHECK(constant->IsDoubleConstant());
value = bit_cast<int64_t, double>(constant->AsDoubleConstant()->GetValue());
}
__ movl(Address(ESP, destination.GetStackIndex()), Immediate(Low32Bits(value)));
__ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), Immediate(High32Bits(value)));
} else {
DCHECK(source.IsDoubleStackSlot()) << source;
EmitParallelMoves(
Location::StackSlot(source.GetStackIndex()),
Location::StackSlot(destination.GetStackIndex()),
Primitive::kPrimInt,
Location::StackSlot(source.GetHighStackIndex(kX86WordSize)),
Location::StackSlot(destination.GetHighStackIndex(kX86WordSize)),
Primitive::kPrimInt);
}
}
}
void CodeGeneratorX86::Move(HInstruction* instruction, Location location, HInstruction* move_for) {
LocationSummary* locations = instruction->GetLocations();
if (locations != nullptr && locations->Out().Equals(location)) {
return;
}
if (locations != nullptr && locations->Out().IsConstant()) {
HConstant* const_to_move = locations->Out().GetConstant();
if (const_to_move->IsIntConstant() || const_to_move->IsNullConstant()) {
Immediate imm(GetInt32ValueOf(const_to_move));
if (location.IsRegister()) {
__ movl(location.AsRegister<Register>(), imm);
} else if (location.IsStackSlot()) {
__ movl(Address(ESP, location.GetStackIndex()), imm);
} else {
DCHECK(location.IsConstant());
DCHECK_EQ(location.GetConstant(), const_to_move);
}
} else if (const_to_move->IsLongConstant()) {
int64_t value = const_to_move->AsLongConstant()->GetValue();
if (location.IsRegisterPair()) {
__ movl(location.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
__ movl(location.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
} else if (location.IsDoubleStackSlot()) {
__ movl(Address(ESP, location.GetStackIndex()), Immediate(Low32Bits(value)));
__ movl(Address(ESP, location.GetHighStackIndex(kX86WordSize)),
Immediate(High32Bits(value)));
} else {
DCHECK(location.IsConstant());
DCHECK_EQ(location.GetConstant(), instruction);
}
}
} else if (instruction->IsTemporary()) {
Location temp_location = GetTemporaryLocation(instruction->AsTemporary());
if (temp_location.IsStackSlot()) {
Move32(location, temp_location);
} else {
DCHECK(temp_location.IsDoubleStackSlot());
Move64(location, temp_location);
}
} else if (instruction->IsLoadLocal()) {
int slot = GetStackSlot(instruction->AsLoadLocal()->GetLocal());
switch (instruction->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
case Primitive::kPrimFloat:
Move32(location, Location::StackSlot(slot));
break;
case Primitive::kPrimLong:
case Primitive::kPrimDouble:
Move64(location, Location::DoubleStackSlot(slot));
break;
default:
LOG(FATAL) << "Unimplemented local type " << instruction->GetType();
}
} else {
DCHECK((instruction->GetNext() == move_for) || instruction->GetNext()->IsTemporary());
switch (instruction->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
case Primitive::kPrimFloat:
Move32(location, locations->Out());
break;
case Primitive::kPrimLong:
case Primitive::kPrimDouble:
Move64(location, locations->Out());
break;
default:
LOG(FATAL) << "Unexpected type " << instruction->GetType();
}
}
}
void LocationsBuilderX86::VisitGoto(HGoto* got) {
got->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitGoto(HGoto* got) {
HBasicBlock* successor = got->GetSuccessor();
DCHECK(!successor->IsExitBlock());
HBasicBlock* block = got->GetBlock();
HInstruction* previous = got->GetPrevious();
HLoopInformation* info = block->GetLoopInformation();
if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) {
GenerateSuspendCheck(info->GetSuspendCheck(), successor);
return;
}
if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) {
GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr);
}
if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) {
__ jmp(codegen_->GetLabelOf(successor));
}
}
void LocationsBuilderX86::VisitExit(HExit* exit) {
exit->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitExit(HExit* exit) {
UNUSED(exit);
}
void InstructionCodeGeneratorX86::GenerateTestAndBranch(HInstruction* instruction,
Label* true_target,
Label* false_target,
Label* always_true_target) {
HInstruction* cond = instruction->InputAt(0);
if (cond->IsIntConstant()) {
// Constant condition, statically compared against 1.
int32_t cond_value = cond->AsIntConstant()->GetValue();
if (cond_value == 1) {
if (always_true_target != nullptr) {
__ jmp(always_true_target);
}
return;
} else {
DCHECK_EQ(cond_value, 0);
}
} else {
bool materialized =
!cond->IsCondition() || cond->AsCondition()->NeedsMaterialization();
// Moves do not affect the eflags register, so if the condition is
// evaluated just before the if, we don't need to evaluate it
// again.
bool eflags_set = cond->IsCondition()
&& cond->AsCondition()->IsBeforeWhenDisregardMoves(instruction);
if (materialized) {
if (!eflags_set) {
// Materialized condition, compare against 0.
Location lhs = instruction->GetLocations()->InAt(0);
if (lhs.IsRegister()) {
__ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
} else {
__ cmpl(Address(ESP, lhs.GetStackIndex()), Immediate(0));
}
__ j(kNotEqual, true_target);
} else {
__ j(X86Condition(cond->AsCondition()->GetCondition()), true_target);
}
} else {
Location lhs = cond->GetLocations()->InAt(0);
Location rhs = cond->GetLocations()->InAt(1);
// LHS is guaranteed to be in a register (see
// LocationsBuilderX86::VisitCondition).
if (rhs.IsRegister()) {
__ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>());
} else if (rhs.IsConstant()) {
int32_t constant = CodeGenerator::GetInt32ValueOf(rhs.GetConstant());
if (constant == 0) {
__ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
} else {
__ cmpl(lhs.AsRegister<Register>(), Immediate(constant));
}
} else {
__ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex()));
}
__ j(X86Condition(cond->AsCondition()->GetCondition()), true_target);
}
}
if (false_target != nullptr) {
__ jmp(false_target);
}
}
void LocationsBuilderX86::VisitIf(HIf* if_instr) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(if_instr, LocationSummary::kNoCall);
HInstruction* cond = if_instr->InputAt(0);
if (!cond->IsCondition() || cond->AsCondition()->NeedsMaterialization()) {
locations->SetInAt(0, Location::Any());
}
}
void InstructionCodeGeneratorX86::VisitIf(HIf* if_instr) {
Label* true_target = codegen_->GetLabelOf(if_instr->IfTrueSuccessor());
Label* false_target = codegen_->GetLabelOf(if_instr->IfFalseSuccessor());
Label* always_true_target = true_target;
if (codegen_->GoesToNextBlock(if_instr->GetBlock(),
if_instr->IfTrueSuccessor())) {
always_true_target = nullptr;
}
if (codegen_->GoesToNextBlock(if_instr->GetBlock(),
if_instr->IfFalseSuccessor())) {
false_target = nullptr;
}
GenerateTestAndBranch(if_instr, true_target, false_target, always_true_target);
}
void LocationsBuilderX86::VisitDeoptimize(HDeoptimize* deoptimize) {
LocationSummary* locations = new (GetGraph()->GetArena())
LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath);
HInstruction* cond = deoptimize->InputAt(0);
DCHECK(cond->IsCondition());
if (cond->AsCondition()->NeedsMaterialization()) {
locations->SetInAt(0, Location::Any());
}
}
void InstructionCodeGeneratorX86::VisitDeoptimize(HDeoptimize* deoptimize) {
SlowPathCodeX86* slow_path = new (GetGraph()->GetArena())
DeoptimizationSlowPathX86(deoptimize);
codegen_->AddSlowPath(slow_path);
Label* slow_path_entry = slow_path->GetEntryLabel();
GenerateTestAndBranch(deoptimize, slow_path_entry, nullptr, slow_path_entry);
}
void LocationsBuilderX86::VisitLocal(HLocal* local) {
local->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitLocal(HLocal* local) {
DCHECK_EQ(local->GetBlock(), GetGraph()->GetEntryBlock());
}
void LocationsBuilderX86::VisitLoadLocal(HLoadLocal* local) {
local->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitLoadLocal(HLoadLocal* load) {
// Nothing to do, this is driven by the code generator.
UNUSED(load);
}
void LocationsBuilderX86::VisitStoreLocal(HStoreLocal* store) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(store, LocationSummary::kNoCall);
switch (store->InputAt(1)->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
case Primitive::kPrimFloat:
locations->SetInAt(1, Location::StackSlot(codegen_->GetStackSlot(store->GetLocal())));
break;
case Primitive::kPrimLong:
case Primitive::kPrimDouble:
locations->SetInAt(1, Location::DoubleStackSlot(codegen_->GetStackSlot(store->GetLocal())));
break;
default:
LOG(FATAL) << "Unknown local type " << store->InputAt(1)->GetType();
}
store->SetLocations(locations);
}
void InstructionCodeGeneratorX86::VisitStoreLocal(HStoreLocal* store) {
UNUSED(store);
}
void LocationsBuilderX86::VisitCondition(HCondition* comp) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(comp, LocationSummary::kNoCall);
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::Any());
if (comp->NeedsMaterialization()) {
// We need a byte register.
locations->SetOut(Location::RegisterLocation(ECX));
}
}
void InstructionCodeGeneratorX86::VisitCondition(HCondition* comp) {
if (comp->NeedsMaterialization()) {
LocationSummary* locations = comp->GetLocations();
Register reg = locations->Out().AsRegister<Register>();
// Clear register: setcc only sets the low byte.
__ xorl(reg, reg);
Location lhs = locations->InAt(0);
Location rhs = locations->InAt(1);
if (rhs.IsRegister()) {
__ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>());
} else if (rhs.IsConstant()) {
int32_t constant = CodeGenerator::GetInt32ValueOf(rhs.GetConstant());
if (constant == 0) {
__ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
} else {
__ cmpl(lhs.AsRegister<Register>(), Immediate(constant));
}
} else {
__ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex()));
}
__ setb(X86Condition(comp->GetCondition()), reg);
}
}
void LocationsBuilderX86::VisitEqual(HEqual* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitEqual(HEqual* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitNotEqual(HNotEqual* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitNotEqual(HNotEqual* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitLessThan(HLessThan* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitLessThan(HLessThan* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitGreaterThan(HGreaterThan* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitGreaterThan(HGreaterThan* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
VisitCondition(comp);
}
void InstructionCodeGeneratorX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
VisitCondition(comp);
}
void LocationsBuilderX86::VisitIntConstant(HIntConstant* constant) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
locations->SetOut(Location::ConstantLocation(constant));
}
void InstructionCodeGeneratorX86::VisitIntConstant(HIntConstant* constant) {
// Will be generated at use site.
UNUSED(constant);
}
void LocationsBuilderX86::VisitNullConstant(HNullConstant* constant) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
locations->SetOut(Location::ConstantLocation(constant));
}
void InstructionCodeGeneratorX86::VisitNullConstant(HNullConstant* constant) {
// Will be generated at use site.
UNUSED(constant);
}
void LocationsBuilderX86::VisitLongConstant(HLongConstant* constant) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
locations->SetOut(Location::ConstantLocation(constant));
}
void InstructionCodeGeneratorX86::VisitLongConstant(HLongConstant* constant) {
// Will be generated at use site.
UNUSED(constant);
}
void LocationsBuilderX86::VisitFloatConstant(HFloatConstant* constant) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
locations->SetOut(Location::ConstantLocation(constant));
}
void InstructionCodeGeneratorX86::VisitFloatConstant(HFloatConstant* constant) {
// Will be generated at use site.
UNUSED(constant);
}
void LocationsBuilderX86::VisitDoubleConstant(HDoubleConstant* constant) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
locations->SetOut(Location::ConstantLocation(constant));
}
void InstructionCodeGeneratorX86::VisitDoubleConstant(HDoubleConstant* constant) {
// Will be generated at use site.
UNUSED(constant);
}
void LocationsBuilderX86::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) {
memory_barrier->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) {
GenerateMemoryBarrier(memory_barrier->GetBarrierKind());
}
void LocationsBuilderX86::VisitReturnVoid(HReturnVoid* ret) {
ret->SetLocations(nullptr);
}
void InstructionCodeGeneratorX86::VisitReturnVoid(HReturnVoid* ret) {
UNUSED(ret);
codegen_->GenerateFrameExit();
}
void LocationsBuilderX86::VisitReturn(HReturn* ret) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall);
switch (ret->InputAt(0)->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
locations->SetInAt(0, Location::RegisterLocation(EAX));
break;
case Primitive::kPrimLong:
locations->SetInAt(
0, Location::RegisterPairLocation(EAX, EDX));
break;
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
locations->SetInAt(
0, Location::FpuRegisterLocation(XMM0));
break;
default:
LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType();
}
}
void InstructionCodeGeneratorX86::VisitReturn(HReturn* ret) {
if (kIsDebugBuild) {
switch (ret->InputAt(0)->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegister<Register>(), EAX);
break;
case Primitive::kPrimLong:
DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairLow<Register>(), EAX);
DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairHigh<Register>(), EDX);
break;
case Primitive::kPrimFloat:
case Primitive::kPrimDouble:
DCHECK_EQ(ret->GetLocations()->InAt(0).AsFpuRegister<XmmRegister>(), XMM0);
break;
default:
LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType();
}
}
codegen_->GenerateFrameExit();
}
void LocationsBuilderX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
// When we do not run baseline, explicit clinit checks triggered by static
// invokes must have been pruned by art::PrepareForRegisterAllocation.
DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck());
IntrinsicLocationsBuilderX86 intrinsic(codegen_);
if (intrinsic.TryDispatch(invoke)) {
return;
}
HandleInvoke(invoke);
}
static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorX86* codegen) {
if (invoke->GetLocations()->Intrinsified()) {
IntrinsicCodeGeneratorX86 intrinsic(codegen);
intrinsic.Dispatch(invoke);
return true;
}
return false;
}
void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
// When we do not run baseline, explicit clinit checks triggered by static
// invokes must have been pruned by art::PrepareForRegisterAllocation.
DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck());
if (TryGenerateIntrinsicCode(invoke, codegen_)) {
return;
}
codegen_->GenerateStaticOrDirectCall(
invoke, invoke->GetLocations()->GetTemp(0).AsRegister<Register>());
codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
}
void LocationsBuilderX86::VisitInvokeVirtual(HInvokeVirtual* invoke) {
HandleInvoke(invoke);
}
void LocationsBuilderX86::HandleInvoke(HInvoke* invoke) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(invoke, LocationSummary::kCall);
locations->AddTemp(Location::RegisterLocation(EAX));
InvokeDexCallingConventionVisitorX86 calling_convention_visitor;
for (size_t i = 0; i < invoke->GetNumberOfArguments(); i++) {
HInstruction* input = invoke->InputAt(i);
locations->SetInAt(i, calling_convention_visitor.GetNextLocation(input->GetType()));
}
switch (invoke->GetType()) {
case Primitive::kPrimBoolean:
case Primitive::kPrimByte:
case Primitive::kPrimChar:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimNot:
locations->SetOut(Location::RegisterLocation(EAX));
break;
case Primitive::kPrimLong:
locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
break;
case Primitive::kPrimVoid:
break;
case Primitive::kPrimDouble:
case Primitive::kPrimFloat:
locations->SetOut(Location::FpuRegisterLocation(XMM0));
break;
}
invoke->SetLocations(locations);
}
void InstructionCodeGeneratorX86::VisitInvokeVirtual(HInvokeVirtual* invoke) {
Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
uint32_t method_offset = mirror::Class::EmbeddedVTableOffset().Uint32Value() +
invoke->GetVTableIndex() * sizeof(mirror::Class::VTableEntry);
LocationSummary* locations = invoke->GetLocations();
Location receiver = locations->InAt(0);
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
// temp = object->GetClass();
if (receiver.IsStackSlot()) {
__ movl(temp, Address(ESP, receiver.GetStackIndex()));
__ movl(temp, Address(temp, class_offset));
} else {
__ movl(temp, Address(receiver.AsRegister<Register>(), class_offset));
}
codegen_->MaybeRecordImplicitNullCheck(invoke);
// temp = temp->GetMethodAt(method_offset);
__ movl(temp, Address(temp, method_offset));
// call temp->GetEntryPoint();
__ call(Address(
temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
DCHECK(!codegen_->IsLeafMethod());
codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
}
void LocationsBuilderX86::VisitInvokeInterface(HInvokeInterface* invoke) {
HandleInvoke(invoke);
// Add the hidden argument.
invoke->GetLocations()->AddTemp(Location::FpuRegisterLocation(XMM7));
}
void InstructionCodeGeneratorX86::VisitInvokeInterface(HInvokeInterface* invoke) {
// TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError.
Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
uint32_t method_offset = mirror::Class::EmbeddedImTableOffset().Uint32Value() +
(invoke->GetImtIndex() % mirror::Class::kImtSize) * sizeof(mirror::Class::ImTableEntry);
LocationSummary* locations = invoke->GetLocations();
Location receiver = locations->InAt(0);
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
// Set the hidden argument.
__ movl(temp, Immediate(invoke->GetDexMethodIndex()));
__ movd(invoke->GetLocations()->GetTemp(1).AsFpuRegister<XmmRegister>(), temp);
// temp = object->GetClass();
if (receiver.IsStackSlot()) {
__ movl(temp, Address(ESP, receiver.GetStackIndex()));
__ movl(temp, Address(temp, class_offset));
} else {
__ movl(temp, Address(receiver.AsRegister<Register>(), class_offset));
}
codegen_->MaybeRecordImplicitNullCheck(invoke);
// temp = temp->GetImtEntryAt(method_offset);
__ movl(temp, Address(temp, method_offset));
// call temp->GetEntryPoint();
__ call(Address(temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
kX86WordSize).Int32Value()));
DCHECK(!codegen_->IsLeafMethod());
codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
}
void LocationsBuilderX86::VisitNeg(HNeg* neg) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall);
switch (neg->GetResultType()) {
case Primitive::kPrimInt:
case Primitive::kPrimLong:
locations->SetInAt(0, Location::RequiresRegister());
locations->SetOut(Location::SameAsFirstInput());
break;
case Primitive::kPrimFloat:
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::SameAsFirstInput());
locations->AddTemp(Location::RequiresRegister());
locations->AddTemp(Location::RequiresFpuRegister());
break;
case Primitive::kPrimDouble:
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::SameAsFirstInput());
locations->AddTemp(Location::RequiresFpuRegister());
break;
default:
LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
}
}
void InstructionCodeGeneratorX86::VisitNeg(HNeg* neg) {
LocationSummary* locations = neg->GetLocations();
Location out = locations->Out();
Location in = locations->InAt(0);
switch (neg->GetResultType()) {
case Primitive::kPrimInt:
DCHECK(in.IsRegister());
DCHECK(in.Equals(out));
__ negl(out.AsRegister<Register>());
break;
case Primitive::kPrimLong:
DCHECK(in.IsRegisterPair());
DCHECK(in.Equals(out));
__ negl(out.AsRegisterPairLow<Register>());
// Negation is similar to subtraction from zero. The least
// significant byte triggers a borrow when it is different from
// zero; to take it into account, add 1 to the most significant
// byte if the carry flag (CF) is set to 1 after the first NEGL
// operation.
__ adcl(out.AsRegisterPairHigh<Register>(), Immediate(0));
__ negl(out.AsRegisterPairHigh<Register>());
break;
case Primitive::kPrimFloat: {
DCHECK(in.Equals(out));
Register constant = locations->GetTemp(0).AsRegister<Register>();
XmmRegister mask = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
// Implement float negation with an exclusive or with value
// 0x80000000 (mask for bit 31, representing the sign of a
// single-precision floating-point number).
__ movl(constant, Immediate(INT32_C(0x80000000)));
__ movd(mask, constant);
__ xorps(out.AsFpuRegister<XmmRegister>(), mask);
break;
}
case Primitive::kPrimDouble: {
DCHECK(in.Equals(out));
XmmRegister mask = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
// Implement double negation with an exclusive or with value
// 0x8000000000000000 (mask for bit 63, representing the sign of
// a double-precision floating-point number).
__ LoadLongConstant(mask, INT64_C(0x8000000000000000));
__ xorpd(out.AsFpuRegister<XmmRegister>(), mask);
break;
}
default:
LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
}
}
void LocationsBuilderX86::VisitTypeConversion(HTypeConversion* conversion) {
Primitive::Type result_type = conversion->GetResultType();
Primitive::Type input_type = conversion->GetInputType();
DCHECK_NE(result_type, input_type);
// The float-to-long and double-to-long type conversions rely on a
// call to the runtime.
LocationSummary::CallKind call_kind =
((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble)
&& result_type == Primitive::kPrimLong)
? LocationSummary::kCall
: LocationSummary::kNoCall;
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind);
// The Java language does not allow treating boolean as an integral type but
// our bit representation makes it safe.
switch (result_type) {
case Primitive::kPrimByte:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-byte' instruction.
locations->SetInAt(0, Location::ByteRegisterOrConstant(ECX, conversion->InputAt(0)));
// Make the output overlap to please the register allocator. This greatly simplifies
// the validation of the linear scan implementation
locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimShort:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-short' instruction.
locations->SetInAt(0, Location::Any());
locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimInt:
switch (input_type) {
case Primitive::kPrimLong:
// Processing a Dex `long-to-int' instruction.
locations->SetInAt(0, Location::Any());
locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
break;
case Primitive::kPrimFloat:
// Processing a Dex `float-to-int' instruction.
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::RequiresRegister());
locations->AddTemp(Location::RequiresFpuRegister());
break;
case Primitive::kPrimDouble:
// Processing a Dex `double-to-int' instruction.
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::RequiresRegister());
locations->AddTemp(Location::RequiresFpuRegister());
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimLong:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-long' instruction.
locations->SetInAt(0, Location::RegisterLocation(EAX));
locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
break;
case Primitive::kPrimFloat:
case Primitive::kPrimDouble: {
// Processing a Dex `float-to-long' or 'double-to-long' instruction.
InvokeRuntimeCallingConvention calling_convention;
XmmRegister parameter = calling_convention.GetFpuRegisterAt(0);
locations->SetInAt(0, Location::FpuRegisterLocation(parameter));
// The runtime helper puts the result in EAX, EDX.
locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
}
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimChar:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
// Processing a Dex `int-to-char' instruction.
locations->SetInAt(0, Location::Any());
locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimFloat:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-float' instruction.
locations->SetInAt(0, Location::RequiresRegister());
locations->SetOut(Location::RequiresFpuRegister());
break;
case Primitive::kPrimLong:
// Processing a Dex `long-to-float' instruction.
locations->SetInAt(0, Location::Any());
locations->SetOut(Location::Any());
break;
case Primitive::kPrimDouble:
// Processing a Dex `double-to-float' instruction.
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
};
break;
case Primitive::kPrimDouble:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-double' instruction.
locations->SetInAt(0, Location::RequiresRegister());
locations->SetOut(Location::RequiresFpuRegister());
break;
case Primitive::kPrimLong:
// Processing a Dex `long-to-double' instruction.
locations->SetInAt(0, Location::Any());
locations->SetOut(Location::Any());
break;
case Primitive::kPrimFloat:
// Processing a Dex `float-to-double' instruction.
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
}
void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversion) {
LocationSummary* locations = conversion->GetLocations();
Location out = locations->Out();
Location in = locations->InAt(0);
Primitive::Type result_type = conversion->GetResultType();
Primitive::Type input_type = conversion->GetInputType();
DCHECK_NE(result_type, input_type);
switch (result_type) {
case Primitive::kPrimByte:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-byte' instruction.
if (in.IsRegister()) {
__ movsxb(out.AsRegister<Register>(), in.AsRegister<ByteRegister>());
} else {
DCHECK(in.GetConstant()->IsIntConstant());
int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
__ movl(out.AsRegister<Register>(), Immediate(static_cast<int8_t>(value)));
}
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimShort:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-short' instruction.
if (in.IsRegister()) {
__ movsxw(out.AsRegister<Register>(), in.AsRegister<Register>());
} else if (in.IsStackSlot()) {
__ movsxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
} else {
DCHECK(in.GetConstant()->IsIntConstant());
int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
__ movl(out.AsRegister<Register>(), Immediate(static_cast<int16_t>(value)));
}
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimInt:
switch (input_type) {
case Primitive::kPrimLong:
// Processing a Dex `long-to-int' instruction.
if (in.IsRegisterPair()) {
__ movl(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>());
} else if (in.IsDoubleStackSlot()) {
__ movl(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
} else {
DCHECK(in.IsConstant());
DCHECK(in.GetConstant()->IsLongConstant());
int64_t value = in.GetConstant()->AsLongConstant()->GetValue();
__ movl(out.AsRegister<Register>(), Immediate(static_cast<int32_t>(value)));
}
break;
case Primitive::kPrimFloat: {
// Processing a Dex `float-to-int' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
Register output = out.AsRegister<Register>();
XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
Label done, nan;
__ movl(output, Immediate(kPrimIntMax));
// temp = int-to-float(output)
__ cvtsi2ss(temp, output);
// if input >= temp goto done
__ comiss(input, temp);
__ j(kAboveEqual, &done);
// if input == NaN goto nan
__ j(kUnordered, &nan);
// output = float-to-int-truncate(input)
__ cvttss2si(output, input);
__ jmp(&done);
__ Bind(&nan);
// output = 0
__ xorl(output, output);
__ Bind(&done);
break;
}
case Primitive::kPrimDouble: {
// Processing a Dex `double-to-int' instruction.
XmmRegister input = in.AsFpuRegister<XmmRegister>();
Register output = out.AsRegister<Register>();
XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
Label done, nan;
__ movl(output, Immediate(kPrimIntMax));
// temp = int-to-double(output)
__ cvtsi2sd(temp, output);
// if input >= temp goto done
__ comisd(input, temp);
__ j(kAboveEqual, &done);
// if input == NaN goto nan
__ j(kUnordered, &nan);
// output = double-to-int-truncate(input)
__ cvttsd2si(output, input);
__ jmp(&done);
__ Bind(&nan);
// output = 0
__ xorl(output, output);
__ Bind(&done);
break;
}
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimLong:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-long' instruction.
DCHECK_EQ(out.AsRegisterPairLow<Register>(), EAX);
DCHECK_EQ(out.AsRegisterPairHigh<Register>(), EDX);
DCHECK_EQ(in.AsRegister<Register>(), EAX);
__ cdq();
break;
case Primitive::kPrimFloat:
// Processing a Dex `float-to-long' instruction.
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pF2l)));
codegen_->RecordPcInfo(conversion, conversion->GetDexPc());
break;
case Primitive::kPrimDouble:
// Processing a Dex `double-to-long' instruction.
__ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pD2l)));
codegen_->RecordPcInfo(conversion, conversion->GetDexPc());
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimChar:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
// Processing a Dex `Process a Dex `int-to-char'' instruction.
if (in.IsRegister()) {
__ movzxw(out.AsRegister<Register>(), in.AsRegister<Register>());
} else if (in.IsStackSlot()) {
__ movzxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
} else {
DCHECK(in.GetConstant()->IsIntConstant());
int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
__ movl(out.AsRegister<Register>(), Immediate(static_cast<uint16_t>(value)));
}
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
break;
case Primitive::kPrimFloat:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-float' instruction.
__ cvtsi2ss(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>());
break;
case Primitive::kPrimLong: {
// Processing a Dex `long-to-float' instruction.
size_t adjustment = 0;
// Create stack space for the call to
// InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstps below.
// TODO: enhance register allocator to ask for stack temporaries.
if (!in.IsDoubleStackSlot() || !out.IsStackSlot()) {
adjustment = Primitive::ComponentSize(Primitive::kPrimLong);
__ subl(ESP, Immediate(adjustment));
}
// Load the value to the FP stack, using temporaries if needed.
PushOntoFPStack(in, 0, adjustment, false, true);
if (out.IsStackSlot()) {
__ fstps(Address(ESP, out.GetStackIndex() + adjustment));
} else {
__ fstps(Address(ESP, 0));
Location stack_temp = Location::StackSlot(0);
codegen_->Move32(out, stack_temp);
}
// Remove the temporary stack space we allocated.
if (adjustment != 0) {
__ addl(ESP, Immediate(adjustment));
}
break;
}
case Primitive::kPrimDouble:
// Processing a Dex `double-to-float' instruction.
__ cvtsd2ss(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>());
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
};
break;
case Primitive::kPrimDouble:
switch (input_type) {
case Primitive::kPrimBoolean:
// Boolean input is a result of code transformations.
case Primitive::kPrimByte:
case Primitive::kPrimShort:
case Primitive::kPrimInt:
case Primitive::kPrimChar:
// Processing a Dex `int-to-double' instruction.
__ cvtsi2sd(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>());
break;
case Primitive::kPrimLong: {
// Processing a Dex `long-to-double' instruction.
size_t adjustment = 0;
// Create stack space for the call to
// InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstpl below.
// TODO: enhance register allocator to ask for stack temporaries.
if (!in.IsDoubleStackSlot() || !out.IsDoubleStackSlot()) {
adjustment = Primitive::ComponentSize(Primitive::kPrimLong);
__ subl(ESP, Immediate(adjustment));
}
// Load the value to the FP stack, using temporaries if needed.
PushOntoFPStack(in, 0, adjustment, false, true);
if (out.IsDoubleStackSlot()) {
__ fstpl(Address(ESP, out.GetStackIndex() + adjustment));
} else {
__ fstpl(Address(ESP, 0));
Location stack_temp = Location::DoubleStackSlot(0);
codegen_->Move64(out, stack_temp);
}
// Remove the temporary stack space we allocated.
if (adjustment != 0) {
__ addl(ESP, Immediate(adjustment));
}
break;
}
case Primitive::kPrimFloat:
// Processing a Dex `float-to-double' instruction.
__ cvtss2sd(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>());
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
};
break;
default:
LOG(FATAL) << "Unexpected type conversion from " << input_type
<< " to " << result_type;
}
}
void LocationsBuilderX86::VisitAdd(HAdd* add) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall);
switch (add->GetResultType()) {
case Primitive::kPrimInt: {
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1)));
locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
break;
}
case Primitive::kPrimLong: {
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::Any());
locations->SetOut(Location::SameAsFirstInput());
break;
}
case Primitive::kPrimFloat:
case Primitive::kPrimDouble: {
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetInAt(1, Location::RequiresFpuRegister());
locations->SetOut(Location::SameAsFirstInput());
break;
}
default:
LOG(FATAL) << "Unexpected add type " << add->GetResultType();
break;
}
}
void InstructionCodeGeneratorX86::VisitAdd(HAdd* add) {
LocationSummary* locations = add->GetLocations();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
Location out = locations->Out();
switch (add->GetResultType()) {
case Primitive::kPrimInt: {
if (second.IsRegister()) {
if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
__ addl(out.AsRegister<Register>(), second.AsRegister<Register>());
} else {
__ leal(out.AsRegister<Register>(), Address(
first.AsRegister<Register>(), second.AsRegister<Register>(), TIMES_1, 0));
}
} else if (second.IsConstant()) {
int32_t value = second.GetConstant()->AsIntConstant()->GetValue();
if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
__ addl(out.AsRegister<Register>(), Immediate(value));
} else {
__ leal(out.AsRegister<Register>(), Address(first.AsRegister<Register>(), value));
}
} else {
DCHECK(first.Equals(locations->Out()));
__ addl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
}
break;
}
case Primitive::kPrimLong: {
if (second.IsRegisterPair()) {
__ addl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
__ adcl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
} else if (second.IsDoubleStackSlot()) {
__ addl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
__ adcl(first.AsRegisterPairHigh<Register>(),
Address(ESP, second.GetHighStackIndex(kX86WordSize)));
} else {
DCHECK(second.IsConstant()) << second;
int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
__ addl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
__ adcl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
}
break;
}
case Primitive::kPrimFloat: {
if (second.IsFpuRegister()) {
__ addss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
}
break;
}
case Primitive::kPrimDouble: {
if (second.IsFpuRegister()) {
__ addsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
}
break;
}
default:
LOG(FATAL) << "Unexpected add type " << add->GetResultType();
}
}
void LocationsBuilderX86::VisitSub(HSub* sub) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall);
switch (sub->GetResultType()) {
case Primitive::kPrimInt:
case Primitive::kPrimLong: {
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::Any());
locations->SetOut(Location::SameAsFirstInput());
break;
}
case Primitive::kPrimFloat:
case Primitive::kPrimDouble: {
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetInAt(1, Location::RequiresFpuRegister());
locations->SetOut(Location::SameAsFirstInput());
break;
}
default:
LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
}
}
void InstructionCodeGeneratorX86::VisitSub(HSub* sub) {
LocationSummary* locations = sub->GetLocations();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
DCHECK(first.Equals(locations->Out()));
switch (sub->GetResultType()) {
case Primitive::kPrimInt: {
if (second.IsRegister()) {
__ subl(first.AsRegister<Register>(), second.AsRegister<Register>());
} else if (second.IsConstant()) {
__ subl(first.AsRegister<Register>(),
Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
} else {
__ subl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
}
break;
}
case Primitive::kPrimLong: {
if (second.IsRegisterPair()) {
__ subl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
__ sbbl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
} else if (second.IsDoubleStackSlot()) {
__ subl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
__ sbbl(first.AsRegisterPairHigh<Register>(),
Address(ESP, second.GetHighStackIndex(kX86WordSize)));
} else {
DCHECK(second.IsConstant()) << second;
int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
__ subl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
__ sbbl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
}
break;
}
case Primitive::kPrimFloat: {
__ subss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
break;
}
case Primitive::kPrimDouble: {
__ subsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
break;
}
default:
LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
}
}
void LocationsBuilderX86::VisitMul(HMul* mul) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall);
switch (mul->GetResultType()) {
case Primitive::kPrimInt:
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::Any());
locations->SetOut(Location::SameAsFirstInput());
break;
case Primitive::kPrimLong: {
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::Any());
locations->SetOut(Location::SameAsFirstInput());
// Needed for imul on 32bits with 64bits output.
locations->AddTemp(Location::RegisterLocation(EAX));
locations->AddTemp(Location::RegisterLocation(EDX));
break;
}
case Primitive::kPrimFloat:
case Primitive::kPrimDouble: {
locations->SetInAt(0, Location::RequiresFpuRegister());
locations->SetInAt(1, Location::RequiresFpuRegister());
locations->SetOut(Location::SameAsFirstInput());
break;
}
default:
LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
}
}
void InstructionCodeGeneratorX86::VisitMul(HMul* mul) {
LocationSummary* locations = mul->GetLocations();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
DCHECK(first.Equals(locations->Out()));
switch (mul->GetResultType()) {
case Primitive::kPrimInt: {
if (second.IsRegister()) {
__ imull(first.AsRegister<Register>(), second.AsRegister<Register>());
} else if (second.IsConstant()) {
Immediate imm(second.GetConstant()->AsIntConstant()->GetValue());
__ imull(first.AsRegister<Register>(), imm);
} else {
DCHECK(second.IsStackSlot());
__ imull(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
}
break;
}
case Primitive::kPrimLong: {
Register in1_hi = first.AsRegisterPairHigh<Register>();
Register in1_lo = first.AsRegisterPairLow<Register>();
Register eax = locations->GetTemp(0).AsRegister<Register>();
Register edx = locations->GetTemp(1).AsRegister<Register>();
DCHECK_EQ(EAX, eax);
DCHECK_EQ(EDX, edx);
// input: in1 - 64 bits, in2 - 64 bits.
// output: in1
// formula: in1.hi : in1.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo
// parts: in1.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32]
// parts: in1.lo = (in1.lo * in2.lo)[31:0]
if (second.IsConstant()) {
DCHECK(second.GetConstant()->IsLongConstant());
int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
int32_t low_value = Low32Bits(value);
int32_t high_value = High32Bits(value);
Immediate low(low_value);
Immediate high(high_value);
__ movl(eax, high);
// eax <- in1.lo * in2.hi
__ imull(eax, in1_lo);
// in1.hi <- in1.hi * in2.lo
__ imull(in1_hi, low);
// in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
__ addl(in1_hi, eax);
// move in2_lo to eax to prepare for double precision
__ movl(eax, low);
// edx:eax <- in1.lo * in2.lo
__ mull(in1_lo);
// in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
__ addl(in1_hi, edx);
// in1.lo <- (in1.lo * in2.lo)[31:0];
__ movl(in1_lo, eax);
} else if (second.IsRegisterPair()) {
Register in2_hi = second.AsRegisterPairHigh<Register>();
Register in2_lo = second.AsRegisterPairLow<Register>();
__ movl(eax, in2_hi);
// eax <- in1.lo * in2.hi
__ imull(eax, in1_lo);
// in1.hi <- in1.hi * in2.lo
__ imull(in1_hi, in2_lo);
// in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
__ addl(in1_hi, eax);
// move in1_lo to eax to prepare for double precision
__ movl(eax, in1_lo);
// edx:eax <- in1.lo * in2.lo
__ mull(in2_lo);
// in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
__ addl(in1_hi, edx);
// in1.lo <- (in1.lo * in2.lo)[31:0];
__ movl(in1_lo, eax);
} else {
DCHECK(second.IsDoubleStackSlot()) << second;
Address in2_hi(ESP, second.GetHighStackIndex(kX86WordSize));
Address in2_lo(ESP, second.GetStackIndex());
__ movl(eax, in2_hi);
// eax <- in1.lo * in2.hi
__ imull(eax, in1_lo);
// in1.hi <- in1.hi * in2.lo
__ imull(in1_hi, in2_lo);
// in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
__ addl(in1_hi, eax);
// move in1_lo to eax to prepare for double precision
__ movl(eax, in1_lo);
// edx:eax <- in1.lo * in2.lo
__ mull(in2_lo);
// in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
__ addl(in1_hi, edx);
// in1.lo <- (in1.lo * in2.lo)[31:0];
__ movl(in1_lo, eax);
}
break;
}
case Primitive::kPrimFloat: {
__ mulss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
break;
}
case Primitive::kPrimDouble: {
__ mulsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
break;
}
default:
LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
}
}
void InstructionCodeGeneratorX86::PushOntoFPStack(Location source,
uint32_t temp_offset,
uint32_t stack_adjustment,
bool is_fp,
bool is_wide) {
if (source.IsStackSlot()) {
DCHECK(!is_wide);
if (is_fp) {
__ flds(Address(ESP, source.GetStackIndex() + stack_adjustment));
} else {
__ filds(Address(ESP, source.GetStackIndex() + stack_adjustment));
}
} else if (source.IsDoubleStackSlot()) {
DCHECK(is_wide);
if (is_fp) {
__ fldl(Address(ESP, source.GetStackIndex() + stack_adjustment));
} else {
__ fildl(Address(ESP, source.GetStackIndex() + stack_adjustment));
}
} else {
// Write the value to the temporary location on the stack and load to FP stack.
if (!is_wide) {
Location stack_temp = Location::StackSlot(temp_offset);
codegen_->Move32(stack_temp, source);
if (is_fp) {
__ flds(Address(ESP, temp_offset));
} else {
__ filds(Address(ESP, temp_offset));
}
} else {
Location stack_temp = Location::DoubleStackSlot(temp_offset);
codegen_->Move64(stack_temp, source);
if (is_fp) {
__ fldl(Address(ESP, temp_offset));
} else {
__ fildl(Address(ESP, temp_offset));
}
}
}
}
void InstructionCodeGeneratorX86::GenerateRemFP(HRem *rem) {
Primitive::Type type = rem->GetResultType();
bool is_float = type == Primitive::kPrimFloat;
size_t elem_size = Primitive::ComponentSize(type);
LocationSummary* locations = rem->GetLocations();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
Location out = locations->Out();
// Create stack space for 2 elements.
// TODO: enhance register allocator to ask for stack temporaries.
__ subl(ESP, Immediate(2 * elem_size));
// Load the values to the FP stack in reverse order, using temporaries if needed.
const bool is_wide = !is_float;
PushOntoFPStack(second, elem_size, 2 * elem_size, /* is_fp */ true, is_wide);
PushOntoFPStack(first, 0, 2 * elem_size, /* is_fp */ true, is_wide);
// Loop doing FPREM until we stabilize.
Label retry;
__ Bind(&retry);
__ fprem();
// Move FP status to AX.
__ fstsw();
// And see if the argument reduction is complete. This is signaled by the
// C2 FPU flag bit set to 0.
__ andl(EAX, Immediate(kC2ConditionMask));
__ j(kNotEqual, &retry);
// We have settled on the final value. Retrieve it into an XMM register.
// Store FP top of stack to real stack.
if (is_float) {
__ fsts(Address(ESP, 0));
} else {
__ fstl(Address(ESP, 0));
}
// Pop the 2 items from the FP stack.
__ fucompp();
// Load the value from the stack into an XMM register.
DCHECK(out.IsFpuRegister()) << out;
if (is_float) {
__ movss(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
} else {
__ movsd(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
}
// And remove the temporary stack space we allocated.
__ addl(ESP, Immediate(2 * elem_size));
}
void InstructionCodeGeneratorX86::DivRemOneOrMinusOne(HBinaryOperation* instruction) {
DCHECK(instruction->IsDiv() || instruction->IsRem());
LocationSummary* locations = instruction->GetLocations();
DCHECK(locations->InAt(1).IsConstant());
DCHECK(locations->InAt(1).GetConstant()->IsIntConstant());
Register out_register = locations->Out().AsRegister<Register>();
Register input_register = locations->InAt(0).AsRegister<Register>();
int32_t imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
DCHECK(imm == 1 || imm == -1);
if (instruction->IsRem()) {
__ xorl(out_register, out_register);
} else {
__ movl(out_register, input_register);
if (imm == -1) {
__ negl(out_register);
}
}
}
void InstructionCodeGeneratorX86::DivByPowerOfTwo(HDiv* instruction) {
LocationSummary* locations = instruction->GetLocations();
Register out_register = locations->Out().AsRegister<Register>();
Register input_register = locations->InAt(0).AsRegister<Register>();
int32_t imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
DCHECK(IsPowerOfTwo(std::abs(imm)));
Register num = locations->GetTemp(0).AsRegister<Register>();
__ leal(num, Address(input_register, std::abs(imm) - 1));
__ testl(input_register, input_register);
__ cmovl(kGreaterEqual, num, input_register);
int shift = CTZ(imm);
__ sarl(num, Immediate(shift));
if (imm < 0) {
__ negl(num);
}
__ movl(out_register, num);
}
void InstructionCodeGeneratorX86::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) {
DCHECK(instruction->IsDiv() || instruction->IsRem());
LocationSummary* locations = instruction->GetLocations();
int imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue();
Register eax = locations->InAt(0).AsRegister<Register>();
Register out = locations->Out().AsRegister<Register>();
Register num;
Register edx;
if (instruction->IsDiv()) {
edx = locations->GetTemp(0).AsRegister<Register>();
num = locations->GetTemp(1).AsRegister<Register>();
} else {
edx = locations->Out().AsRegister<Register>();
num = locations->GetTemp(0).AsRegister<Register>();
}
DCHECK_EQ(EAX, eax);
DCHECK_EQ(EDX, edx);
if (instruction->IsDiv()) {
DCHECK_EQ(EAX, out);
} else {
DCHECK_EQ(EDX, out);
}
int64_t magic;
int shift;
CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift);
Label ndiv;
Label end;
// If numerator is 0, the result is 0, no computation needed.
__ testl(eax, eax);
__ j(kNotEqual, &ndiv);
__ xorl(out, out);
__ jmp(&end);
__ Bind(&ndiv);
// Save the numerator.
__ movl(num, eax);
// EAX = magic
__ movl(eax, Immediate(magic));
// EDX:EAX = magic * numerator
__ imull(num);
if (imm > 0 && magic < 0) {
// EDX += num
__ addl(edx, num);
} else if (imm < 0 && magic > 0) {
__ subl(edx, num);
}
// Shift if needed.
if (shift != 0) {
__ sarl(edx, Immediate(shift));
}
// EDX += 1 if EDX < 0
__ movl(eax, edx);
__ shrl(edx, Immediate(31));
__ addl(edx, eax);
if (instruction->IsRem()) {
__ movl(eax, num);
__ imull(edx, Immediate(imm));
__ subl(eax, edx);
__ movl(edx, eax);
} else {
__ movl(eax, edx);
}
__ Bind(&end);
}
void InstructionCodeGeneratorX86::GenerateDivRemIntegral(HBinaryOperation* instruction) {
DCHECK(instruction->IsDiv() || instruction->IsRem());
LocationSummary* locations = instruction->GetLocations();
Location out = locations->Out();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
bool is_div = instruction->IsDiv();
switch (instruction->GetResultType()) {
case Primitive::kPrimInt: {
DCHECK_EQ(EAX, first.AsRegister<Register>());
DCHECK_EQ(is_div ? EAX : EDX, out.AsRegister<Register>());
if (instruction->InputAt(1)->IsIntConstant()) {
int32_t imm = second.GetConstant()->AsIntConstant()->GetValue();
if (imm == 0) {
// Do not generate anything for 0. DivZeroCheck would forbid any generated code.
} else if (imm == 1 || imm == -1) {
DivRemOneOrMinusOne(instruction);
} else if (is_div && IsPowerOfTwo(std::abs(imm))) {
DivByPowerOfTwo(instruction->AsDiv());
} else {
DCHECK(imm <= -2 || imm >= 2);