Google Git
Sign in
cobalt / cobalt / f309f9a11671768c59005e71c207e268484cbe9f / . / src / third_party / v8 / src / compiler / backend / arm / code-generator-arm.cc
blob: 9267cb1f0c353bf1a3f75d4081150115f5f4907d [file] [log] [blame]
// Copyright 2014 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/codegen/assembler-inl.h"
#include "src/codegen/macro-assembler.h"
#include "src/codegen/optimized-compilation-info.h"
#include "src/compiler/backend/code-generator-impl.h"
#include "src/compiler/backend/code-generator.h"
#include "src/compiler/backend/gap-resolver.h"
#include "src/compiler/backend/instruction-codes.h"
#include "src/compiler/node-matchers.h"
#include "src/compiler/osr.h"
#include "src/heap/memory-chunk.h"
#include "src/numbers/double.h"
#include "src/utils/boxed-float.h"
#include "src/wasm/wasm-code-manager.h"
#include "src/wasm/wasm-objects.h"
namespace v8 {
namespace internal {
namespace compiler {
#define __ tasm()->
// Adds Arm-specific methods to convert InstructionOperands.
class ArmOperandConverter final : public InstructionOperandConverter {
public:
ArmOperandConverter(CodeGenerator* gen, Instruction* instr)
: InstructionOperandConverter(gen, instr) {}
SBit OutputSBit() const {
switch (instr_->flags_mode()) {
case kFlags_branch:
case kFlags_branch_and_poison:
case kFlags_deoptimize:
case kFlags_deoptimize_and_poison:
case kFlags_set:
case kFlags_trap:
return SetCC;
case kFlags_none:
return LeaveCC;
}
UNREACHABLE();
}
Operand InputImmediate(size_t index) const {
return ToImmediate(instr_->InputAt(index));
}
Operand InputOperand2(size_t first_index) {
const size_t index = first_index;
switch (AddressingModeField::decode(instr_->opcode())) {
case kMode_None:
case kMode_Offset_RI:
case kMode_Offset_RR:
case kMode_Root:
break;
case kMode_Operand2_I:
return InputImmediate(index + 0);
case kMode_Operand2_R:
return Operand(InputRegister(index + 0));
case kMode_Operand2_R_ASR_I:
return Operand(InputRegister(index + 0), ASR, InputInt5(index + 1));
case kMode_Operand2_R_ASR_R:
return Operand(InputRegister(index + 0), ASR, InputRegister(index + 1));
case kMode_Operand2_R_LSL_I:
return Operand(InputRegister(index + 0), LSL, InputInt5(index + 1));
case kMode_Operand2_R_LSL_R:
return Operand(InputRegister(index + 0), LSL, InputRegister(index + 1));
case kMode_Operand2_R_LSR_I:
return Operand(InputRegister(index + 0), LSR, InputInt5(index + 1));
case kMode_Operand2_R_LSR_R:
return Operand(InputRegister(index + 0), LSR, InputRegister(index + 1));
case kMode_Operand2_R_ROR_I:
return Operand(InputRegister(index + 0), ROR, InputInt5(index + 1));
case kMode_Operand2_R_ROR_R:
return Operand(InputRegister(index + 0), ROR, InputRegister(index + 1));
}
UNREACHABLE();
}
MemOperand InputOffset(size_t* first_index) {
const size_t index = *first_index;
switch (AddressingModeField::decode(instr_->opcode())) {
case kMode_None:
case kMode_Operand2_I:
case kMode_Operand2_R:
case kMode_Operand2_R_ASR_I:
case kMode_Operand2_R_ASR_R:
case kMode_Operand2_R_LSL_R:
case kMode_Operand2_R_LSR_I:
case kMode_Operand2_R_LSR_R:
case kMode_Operand2_R_ROR_I:
case kMode_Operand2_R_ROR_R:
break;
case kMode_Operand2_R_LSL_I:
*first_index += 3;
return MemOperand(InputRegister(index + 0), InputRegister(index + 1),
LSL, InputInt32(index + 2));
case kMode_Offset_RI:
*first_index += 2;
return MemOperand(InputRegister(index + 0), InputInt32(index + 1));
case kMode_Offset_RR:
*first_index += 2;
return MemOperand(InputRegister(index + 0), InputRegister(index + 1));
case kMode_Root:
*first_index += 1;
return MemOperand(kRootRegister, InputInt32(index));
}
UNREACHABLE();
}
MemOperand InputOffset(size_t first_index = 0) {
return InputOffset(&first_index);
}
Operand ToImmediate(InstructionOperand* operand) const {
Constant constant = ToConstant(operand);
switch (constant.type()) {
case Constant::kInt32:
if (RelocInfo::IsWasmReference(constant.rmode())) {
return Operand(constant.ToInt32(), constant.rmode());
} else {
return Operand(constant.ToInt32());
}
case Constant::kFloat32:
return Operand::EmbeddedNumber(constant.ToFloat32());
case Constant::kFloat64:
return Operand::EmbeddedNumber(constant.ToFloat64().value());
case Constant::kExternalReference:
return Operand(constant.ToExternalReference());
case Constant::kDelayedStringConstant:
return Operand::EmbeddedStringConstant(
constant.ToDelayedStringConstant());
case Constant::kInt64:
case Constant::kCompressedHeapObject:
case Constant::kHeapObject:
// TODO(dcarney): loading RPO constants on arm.
case Constant::kRpoNumber:
break;
}
UNREACHABLE();
}
MemOperand ToMemOperand(InstructionOperand* op) const {
DCHECK_NOT_NULL(op);
DCHECK(op->IsStackSlot() || op->IsFPStackSlot());
return SlotToMemOperand(AllocatedOperand::cast(op)->index());
}
MemOperand SlotToMemOperand(int slot) const {
FrameOffset offset = frame_access_state()->GetFrameOffset(slot);
return MemOperand(offset.from_stack_pointer() ? sp : fp, offset.offset());
}
NeonMemOperand NeonInputOperand(size_t first_index) {
const size_t index = first_index;
switch (AddressingModeField::decode(instr_->opcode())) {
case kMode_Operand2_R:
return NeonMemOperand(InputRegister(index + 0));
default:
break;
}
UNREACHABLE();
}
};
namespace {
class OutOfLineRecordWrite final : public OutOfLineCode {
public:
OutOfLineRecordWrite(CodeGenerator* gen, Register object, Operand offset,
Register value, RecordWriteMode mode,
StubCallMode stub_mode,
UnwindingInfoWriter* unwinding_info_writer)
: OutOfLineCode(gen),
object_(object),
offset_(offset),
value_(value),
mode_(mode),
stub_mode_(stub_mode),
must_save_lr_(!gen->frame_access_state()->has_frame()),
unwinding_info_writer_(unwinding_info_writer),
zone_(gen->zone()) {}
void Generate() final {
if (mode_ > RecordWriteMode::kValueIsPointer) {
__ JumpIfSmi(value_, exit());
}
__ CheckPageFlag(value_, MemoryChunk::kPointersToHereAreInterestingMask, eq,
exit());
RememberedSetAction const remembered_set_action =
mode_ > RecordWriteMode::kValueIsMap ? EMIT_REMEMBERED_SET
: OMIT_REMEMBERED_SET;
SaveFPRegsMode const save_fp_mode =
frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
if (must_save_lr_) {
// We need to save and restore lr if the frame was elided.
__ Push(lr);
unwinding_info_writer_->MarkLinkRegisterOnTopOfStack(__ pc_offset());
}
if (mode_ == RecordWriteMode::kValueIsEphemeronKey) {
__ CallEphemeronKeyBarrier(object_, offset_, save_fp_mode);
} else if (stub_mode_ == StubCallMode::kCallWasmRuntimeStub) {
__ CallRecordWriteStub(object_, offset_, remembered_set_action,
save_fp_mode, wasm::WasmCode::kRecordWrite);
} else {
__ CallRecordWriteStub(object_, offset_, remembered_set_action,
save_fp_mode);
}
if (must_save_lr_) {
__ Pop(lr);
unwinding_info_writer_->MarkPopLinkRegisterFromTopOfStack(__ pc_offset());
}
}
private:
Register const object_;
Operand const offset_;
Register const value_;
RecordWriteMode const mode_;
StubCallMode stub_mode_;
bool must_save_lr_;
UnwindingInfoWriter* const unwinding_info_writer_;
Zone* zone_;
};
template <typename T>
class OutOfLineFloatMin final : public OutOfLineCode {
public:
OutOfLineFloatMin(CodeGenerator* gen, T result, T left, T right)
: OutOfLineCode(gen), result_(result), left_(left), right_(right) {}
void Generate() final { __ FloatMinOutOfLine(result_, left_, right_); }
private:
T const result_;
T const left_;
T const right_;
};
using OutOfLineFloat32Min = OutOfLineFloatMin<SwVfpRegister>;
using OutOfLineFloat64Min = OutOfLineFloatMin<DwVfpRegister>;
template <typename T>
class OutOfLineFloatMax final : public OutOfLineCode {
public:
OutOfLineFloatMax(CodeGenerator* gen, T result, T left, T right)
: OutOfLineCode(gen), result_(result), left_(left), right_(right) {}
void Generate() final { __ FloatMaxOutOfLine(result_, left_, right_); }
private:
T const result_;
T const left_;
T const right_;
};
using OutOfLineFloat32Max = OutOfLineFloatMax<SwVfpRegister>;
using OutOfLineFloat64Max = OutOfLineFloatMax<DwVfpRegister>;
Condition FlagsConditionToCondition(FlagsCondition condition) {
switch (condition) {
case kEqual:
return eq;
case kNotEqual:
return ne;
case kSignedLessThan:
return lt;
case kSignedGreaterThanOrEqual:
return ge;
case kSignedLessThanOrEqual:
return le;
case kSignedGreaterThan:
return gt;
case kUnsignedLessThan:
return lo;
case kUnsignedGreaterThanOrEqual:
return hs;
case kUnsignedLessThanOrEqual:
return ls;
case kUnsignedGreaterThan:
return hi;
case kFloatLessThanOrUnordered:
return lt;
case kFloatGreaterThanOrEqual:
return ge;
case kFloatLessThanOrEqual:
return ls;
case kFloatGreaterThanOrUnordered:
return hi;
case kFloatLessThan:
return lo;
case kFloatGreaterThanOrEqualOrUnordered:
return hs;
case kFloatLessThanOrEqualOrUnordered:
return le;
case kFloatGreaterThan:
return gt;
case kOverflow:
return vs;
case kNotOverflow:
return vc;
case kPositiveOrZero:
return pl;
case kNegative:
return mi;
default:
break;
}
UNREACHABLE();
}
void EmitWordLoadPoisoningIfNeeded(CodeGenerator* codegen,
InstructionCode opcode,
ArmOperandConverter const& i) {
const MemoryAccessMode access_mode =
static_cast<MemoryAccessMode>(MiscField::decode(opcode));
if (access_mode == kMemoryAccessPoisoned) {
Register value = i.OutputRegister();
codegen->tasm()->and_(value, value, Operand(kSpeculationPoisonRegister));
}
}
void ComputePoisonedAddressForLoad(CodeGenerator* codegen,
InstructionCode opcode,
ArmOperandConverter const& i,
Register address) {
DCHECK_EQ(kMemoryAccessPoisoned,
static_cast<MemoryAccessMode>(MiscField::decode(opcode)));
switch (AddressingModeField::decode(opcode)) {
case kMode_Offset_RI:
codegen->tasm()->mov(address, i.InputImmediate(1));
codegen->tasm()->add(address, address, i.InputRegister(0));
break;
case kMode_Offset_RR:
codegen->tasm()->add(address, i.InputRegister(0), i.InputRegister(1));
break;
default:
UNREACHABLE();
}
codegen->tasm()->and_(address, address, Operand(kSpeculationPoisonRegister));
}
} // namespace
#define ASSEMBLE_ATOMIC_LOAD_INTEGER(asm_instr) \
do { \
__ asm_instr(i.OutputRegister(), \
MemOperand(i.InputRegister(0), i.InputRegister(1))); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC_STORE_INTEGER(asm_instr) \
do { \
__ dmb(ISH); \
__ asm_instr(i.InputRegister(2), \
MemOperand(i.InputRegister(0), i.InputRegister(1))); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC_EXCHANGE_INTEGER(load_instr, store_instr) \
do { \
Label exchange; \
__ add(i.TempRegister(1), i.InputRegister(0), i.InputRegister(1)); \
__ dmb(ISH); \
__ bind(&exchange); \
__ load_instr(i.OutputRegister(0), i.TempRegister(1)); \
__ store_instr(i.TempRegister(0), i.InputRegister(2), i.TempRegister(1)); \
__ teq(i.TempRegister(0), Operand(0)); \
__ b(ne, &exchange); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC_COMPARE_EXCHANGE_INTEGER(load_instr, store_instr, \
cmp_reg) \
do { \
Label compareExchange; \
Label exit; \
__ dmb(ISH); \
__ bind(&compareExchange); \
__ load_instr(i.OutputRegister(0), i.TempRegister(1)); \
__ teq(cmp_reg, Operand(i.OutputRegister(0))); \
__ b(ne, &exit); \
__ store_instr(i.TempRegister(0), i.InputRegister(3), i.TempRegister(1)); \
__ teq(i.TempRegister(0), Operand(0)); \
__ b(ne, &compareExchange); \
__ bind(&exit); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC_BINOP(load_instr, store_instr, bin_instr) \
do { \
Label binop; \
__ add(i.TempRegister(1), i.InputRegister(0), i.InputRegister(1)); \
__ dmb(ISH); \
__ bind(&binop); \
__ load_instr(i.OutputRegister(0), i.TempRegister(1)); \
__ bin_instr(i.TempRegister(0), i.OutputRegister(0), \
Operand(i.InputRegister(2))); \
__ store_instr(i.TempRegister(2), i.TempRegister(0), i.TempRegister(1)); \
__ teq(i.TempRegister(2), Operand(0)); \
__ b(ne, &binop); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC64_ARITH_BINOP(instr1, instr2) \
do { \
Label binop; \
__ add(i.TempRegister(0), i.InputRegister(2), i.InputRegister(3)); \
__ dmb(ISH); \
__ bind(&binop); \
__ ldrexd(r2, r3, i.TempRegister(0)); \
__ instr1(i.TempRegister(1), r2, i.InputRegister(0), SBit::SetCC); \
__ instr2(i.TempRegister(2), r3, Operand(i.InputRegister(1))); \
DCHECK_EQ(LeaveCC, i.OutputSBit()); \
__ strexd(i.TempRegister(3), i.TempRegister(1), i.TempRegister(2), \
i.TempRegister(0)); \
__ teq(i.TempRegister(3), Operand(0)); \
__ b(ne, &binop); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_ATOMIC64_LOGIC_BINOP(instr) \
do { \
Label binop; \
__ add(i.TempRegister(0), i.InputRegister(2), i.InputRegister(3)); \
__ dmb(ISH); \
__ bind(&binop); \
__ ldrexd(r2, r3, i.TempRegister(0)); \
__ instr(i.TempRegister(1), r2, Operand(i.InputRegister(0))); \
__ instr(i.TempRegister(2), r3, Operand(i.InputRegister(1))); \
__ strexd(i.TempRegister(3), i.TempRegister(1), i.TempRegister(2), \
i.TempRegister(0)); \
__ teq(i.TempRegister(3), Operand(0)); \
__ b(ne, &binop); \
__ dmb(ISH); \
} while (0)
#define ASSEMBLE_IEEE754_BINOP(name) \
do { \
/* TODO(bmeurer): We should really get rid of this special instruction, */ \
/* and generate a CallAddress instruction instead. */ \
FrameScope scope(tasm(), StackFrame::MANUAL); \
__ PrepareCallCFunction(0, 2); \
__ MovToFloatParameters(i.InputDoubleRegister(0), \
i.InputDoubleRegister(1)); \
__ CallCFunction(ExternalReference::ieee754_##name##_function(), 0, 2); \
/* Move the result in the double result register. */ \
__ MovFromFloatResult(i.OutputDoubleRegister()); \
DCHECK_EQ(LeaveCC, i.OutputSBit()); \
} while (0)
#define ASSEMBLE_IEEE754_UNOP(name) \
do { \
/* TODO(bmeurer): We should really get rid of this special instruction, */ \
/* and generate a CallAddress instruction instead. */ \
FrameScope scope(tasm(), StackFrame::MANUAL); \
__ PrepareCallCFunction(0, 1); \
__ MovToFloatParameter(i.InputDoubleRegister(0)); \
__ CallCFunction(ExternalReference::ieee754_##name##_function(), 0, 1); \
/* Move the result in the double result register. */ \
__ MovFromFloatResult(i.OutputDoubleRegister()); \
DCHECK_EQ(LeaveCC, i.OutputSBit()); \
} while (0)
#define ASSEMBLE_NEON_NARROWING_OP(dt, sdt) \
do { \
Simd128Register dst = i.OutputSimd128Register(), \
src0 = i.InputSimd128Register(0), \
src1 = i.InputSimd128Register(1); \
if (dst == src0 && dst == src1) { \
__ vqmovn(dt, sdt, dst.low(), src0); \
__ vmov(dst.high(), dst.low()); \
} else if (dst == src0) { \
__ vqmovn(dt, sdt, dst.low(), src0); \
__ vqmovn(dt, sdt, dst.high(), src1); \
} else { \
__ vqmovn(dt, sdt, dst.high(), src1); \
__ vqmovn(dt, sdt, dst.low(), src0); \
} \
} while (0)
#define ASSEMBLE_NEON_PAIRWISE_OP(op, size) \
do { \
Simd128Register dst = i.OutputSimd128Register(), \
src0 = i.InputSimd128Register(0), \
src1 = i.InputSimd128Register(1); \
if (dst == src0) { \
__ op(size, dst.low(), src0.low(), src0.high()); \
if (dst == src1) { \
__ vmov(dst.high(), dst.low()); \
} else { \
__ op(size, dst.high(), src1.low(), src1.high()); \
} \
} else { \
__ op(size, dst.high(), src1.low(), src1.high()); \
__ op(size, dst.low(), src0.low(), src0.high()); \
} \
} while (0)
#define ASSEMBLE_F64X2_ARITHMETIC_BINOP(op) \
do { \
__ op(i.OutputSimd128Register().low(), i.InputSimd128Register(0).low(), \
i.InputSimd128Register(1).low()); \
__ op(i.OutputSimd128Register().high(), i.InputSimd128Register(0).high(), \
i.InputSimd128Register(1).high()); \
} while (0)
// If shift value is an immediate, we can call asm_imm, taking the shift value
// modulo 2^width. Otherwise, emit code to perform the modulus operation, and
// call vshl.
#define ASSEMBLE_SIMD_SHIFT_LEFT(asm_imm, width, sz, dt) \
do { \
QwNeonRegister dst = i.OutputSimd128Register(); \
QwNeonRegister src = i.InputSimd128Register(0); \
if (instr->InputAt(1)->IsImmediate()) { \
__ asm_imm(dt, dst, src, i.InputInt##width(1)); \
} else { \
QwNeonRegister tmp = i.TempSimd128Register(0); \
Register shift = i.TempRegister(1); \
constexpr int mask = (1 << width) - 1; \
__ and_(shift, i.InputRegister(1), Operand(mask)); \
__ vdup(sz, tmp, shift); \
__ vshl(dt, dst, src, tmp); \
} \
} while (0)
// If shift value is an immediate, we can call asm_imm, taking the shift value
// modulo 2^width. Otherwise, emit code to perform the modulus operation, and
// call vshl, passing in the negative shift value (treated as a right shift).
#define ASSEMBLE_SIMD_SHIFT_RIGHT(asm_imm, width, sz, dt) \
do { \
QwNeonRegister dst = i.OutputSimd128Register(); \
QwNeonRegister src = i.InputSimd128Register(0); \
if (instr->InputAt(1)->IsImmediate()) { \
__ asm_imm(dt, dst, src, i.InputInt##width(1)); \
} else { \
QwNeonRegister tmp = i.TempSimd128Register(0); \
Register shift = i.TempRegister(1); \
constexpr int mask = (1 << width) - 1; \
__ and_(shift, i.InputRegister(1), Operand(mask)); \
__ vdup(sz, tmp, shift); \
__ vneg(sz, tmp, tmp); \
__ vshl(dt, dst, src, tmp); \
} \
} while (0)
void CodeGenerator::AssembleDeconstructFrame() {
__ LeaveFrame(StackFrame::MANUAL);
unwinding_info_writer_.MarkFrameDeconstructed(__ pc_offset());
}
void CodeGenerator::AssemblePrepareTailCall() {
if (frame_access_state()->has_frame()) {
__ ldm(ia, fp, lr.bit() | fp.bit());
}
frame_access_state()->SetFrameAccessToSP();
}
void CodeGenerator::AssemblePopArgumentsAdaptorFrame(Register args_reg,
Register scratch1,
Register scratch2,
Register scratch3) {
DCHECK(!AreAliased(args_reg, scratch1, scratch2, scratch3));
Label done;
// Check if current frame is an arguments adaptor frame.
__ ldr(scratch1, MemOperand(fp, StandardFrameConstants::kContextOffset));
__ cmp(scratch1,
Operand(StackFrame::TypeToMarker(StackFrame::ARGUMENTS_ADAPTOR)));
__ b(ne, &done);
// Load arguments count from current arguments adaptor frame (note, it
// does not include receiver).
Register caller_args_count_reg = scratch1;
__ ldr(caller_args_count_reg,
MemOperand(fp, ArgumentsAdaptorFrameConstants::kLengthOffset));
__ SmiUntag(caller_args_count_reg);
__ PrepareForTailCall(args_reg, caller_args_count_reg, scratch2, scratch3);
__ bind(&done);
}
namespace {
void FlushPendingPushRegisters(TurboAssembler* tasm,
FrameAccessState* frame_access_state,
ZoneVector<Register>* pending_pushes) {
switch (pending_pushes->size()) {
case 0:
break;
case 1:
tasm->push((*pending_pushes)[0]);
break;
case 2:
tasm->Push((*pending_pushes)[0], (*pending_pushes)[1]);
break;
case 3:
tasm->Push((*pending_pushes)[0], (*pending_pushes)[1],
(*pending_pushes)[2]);
break;
default:
UNREACHABLE();
}
frame_access_state->IncreaseSPDelta(pending_pushes->size());
pending_pushes->clear();
}
void AdjustStackPointerForTailCall(
TurboAssembler* tasm, FrameAccessState* state, int new_slot_above_sp,
ZoneVector<Register>* pending_pushes = nullptr,
bool allow_shrinkage = true) {
int current_sp_offset = state->GetSPToFPSlotCount() +
StandardFrameConstants::kFixedSlotCountAboveFp;
int stack_slot_delta = new_slot_above_sp - current_sp_offset;
if (stack_slot_delta > 0) {
if (pending_pushes != nullptr) {
FlushPendingPushRegisters(tasm, state, pending_pushes);
}
tasm->AllocateStackSpace(stack_slot_delta * kSystemPointerSize);
state->IncreaseSPDelta(stack_slot_delta);
} else if (allow_shrinkage && stack_slot_delta < 0) {
if (pending_pushes != nullptr) {
FlushPendingPushRegisters(tasm, state, pending_pushes);
}
tasm->add(sp, sp, Operand(-stack_slot_delta * kSystemPointerSize));
state->IncreaseSPDelta(stack_slot_delta);
}
}
#if DEBUG
bool VerifyOutputOfAtomicPairInstr(ArmOperandConverter* converter,
const Instruction* instr, Register low,
Register high) {
DCHECK_GE(instr->OutputCount() + instr->TempCount(), 2);
if (instr->OutputCount() == 2) {
return (converter->OutputRegister(0) == low &&
converter->OutputRegister(1) == high);
}
if (instr->OutputCount() == 1) {
return (converter->OutputRegister(0) == low &&
converter->TempRegister(instr->TempCount() - 1) == high) ||
(converter->OutputRegister(0) == high &&
converter->TempRegister(instr->TempCount() - 1) == low);
}
DCHECK_EQ(instr->OutputCount(), 0);
return (converter->TempRegister(instr->TempCount() - 2) == low &&
converter->TempRegister(instr->TempCount() - 1) == high);
}
#endif
} // namespace
void CodeGenerator::AssembleTailCallBeforeGap(Instruction* instr,
int first_unused_stack_slot) {
ZoneVector<MoveOperands*> pushes(zone());
GetPushCompatibleMoves(instr, kRegisterPush, &pushes);
if (!pushes.empty() &&
(LocationOperand::cast(pushes.back()->destination()).index() + 1 ==
first_unused_stack_slot)) {
ArmOperandConverter g(this, instr);
ZoneVector<Register> pending_pushes(zone());
for (auto move : pushes) {
LocationOperand destination_location(
LocationOperand::cast(move->destination()));
InstructionOperand source(move->source());
AdjustStackPointerForTailCall(
tasm(), frame_access_state(),
destination_location.index() - pending_pushes.size(),
&pending_pushes);
// Pushes of non-register data types are not supported.
DCHECK(source.IsRegister());
LocationOperand source_location(LocationOperand::cast(source));
pending_pushes.push_back(source_location.GetRegister());
// TODO(arm): We can push more than 3 registers at once. Add support in
// the macro-assembler for pushing a list of registers.
if (pending_pushes.size() == 3) {
FlushPendingPushRegisters(tasm(), frame_access_state(),
&pending_pushes);
}
move->Eliminate();
}
FlushPendingPushRegisters(tasm(), frame_access_state(), &pending_pushes);
}
AdjustStackPointerForTailCall(tasm(), frame_access_state(),
first_unused_stack_slot, nullptr, false);
}
void CodeGenerator::AssembleTailCallAfterGap(Instruction* instr,
int first_unused_stack_slot) {
AdjustStackPointerForTailCall(tasm(), frame_access_state(),
first_unused_stack_slot);
}
// Check that {kJavaScriptCallCodeStartRegister} is correct.
void CodeGenerator::AssembleCodeStartRegisterCheck() {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ ComputeCodeStartAddress(scratch);
__ cmp(scratch, kJavaScriptCallCodeStartRegister);
__ Assert(eq, AbortReason::kWrongFunctionCodeStart);
}
// Check if the code object is marked for deoptimization. If it is, then it
// jumps to the CompileLazyDeoptimizedCode builtin. In order to do this we need
// to:
// 1. read from memory the word that contains that bit, which can be found in
// the flags in the referenced {CodeDataContainer} object;
// 2. test kMarkedForDeoptimizationBit in those flags; and
// 3. if it is not zero then it jumps to the builtin.
void CodeGenerator::BailoutIfDeoptimized() {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
int offset = Code::kCodeDataContainerOffset - Code::kHeaderSize;
__ ldr(scratch, MemOperand(kJavaScriptCallCodeStartRegister, offset));
__ ldr(scratch,
FieldMemOperand(scratch, CodeDataContainer::kKindSpecificFlagsOffset));
__ tst(scratch, Operand(1 << Code::kMarkedForDeoptimizationBit));
__ Jump(BUILTIN_CODE(isolate(), CompileLazyDeoptimizedCode),
RelocInfo::CODE_TARGET, ne);
}
void CodeGenerator::GenerateSpeculationPoisonFromCodeStartRegister() {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
// Set a mask which has all bits set in the normal case, but has all
// bits cleared if we are speculatively executing the wrong PC.
__ ComputeCodeStartAddress(scratch);
__ cmp(kJavaScriptCallCodeStartRegister, scratch);
__ mov(kSpeculationPoisonRegister, Operand(-1), SBit::LeaveCC, eq);
__ mov(kSpeculationPoisonRegister, Operand(0), SBit::LeaveCC, ne);
__ csdb();
}
void CodeGenerator::AssembleRegisterArgumentPoisoning() {
__ and_(kJSFunctionRegister, kJSFunctionRegister, kSpeculationPoisonRegister);
__ and_(kContextRegister, kContextRegister, kSpeculationPoisonRegister);
__ and_(sp, sp, kSpeculationPoisonRegister);
}
// Assembles an instruction after register allocation, producing machine code.
CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
Instruction* instr) {
ArmOperandConverter i(this, instr);
__ MaybeCheckConstPool();
InstructionCode opcode = instr->opcode();
ArchOpcode arch_opcode = ArchOpcodeField::decode(opcode);
switch (arch_opcode) {
case kArchCallCodeObject: {
if (instr->InputAt(0)->IsImmediate()) {
__ Call(i.InputCode(0), RelocInfo::CODE_TARGET);
} else {
Register reg = i.InputRegister(0);
DCHECK_IMPLIES(
instr->HasCallDescriptorFlag(CallDescriptor::kFixedTargetRegister),
reg == kJavaScriptCallCodeStartRegister);
__ CallCodeObject(reg);
}
RecordCallPosition(instr);
DCHECK_EQ(LeaveCC, i.OutputSBit());
frame_access_state()->ClearSPDelta();
break;
}
case kArchCallBuiltinPointer: {
DCHECK(!instr->InputAt(0)->IsImmediate());
Register builtin_index = i.InputRegister(0);
__ CallBuiltinByIndex(builtin_index);
RecordCallPosition(instr);
frame_access_state()->ClearSPDelta();
break;
}
case kArchCallWasmFunction: {
if (instr->InputAt(0)->IsImmediate()) {
Constant constant = i.ToConstant(instr->InputAt(0));
Address wasm_code = static_cast<Address>(constant.ToInt32());
__ Call(wasm_code, constant.rmode());
} else {
__ Call(i.InputRegister(0));
}
RecordCallPosition(instr);
DCHECK_EQ(LeaveCC, i.OutputSBit());
frame_access_state()->ClearSPDelta();
break;
}
case kArchTailCallCodeObjectFromJSFunction:
case kArchTailCallCodeObject: {
if (arch_opcode == kArchTailCallCodeObjectFromJSFunction) {
AssemblePopArgumentsAdaptorFrame(kJavaScriptCallArgCountRegister,
i.TempRegister(0), i.TempRegister(1),
i.TempRegister(2));
}
if (instr->InputAt(0)->IsImmediate()) {
__ Jump(i.InputCode(0), RelocInfo::CODE_TARGET);
} else {
Register reg = i.InputRegister(0);
DCHECK_IMPLIES(
instr->HasCallDescriptorFlag(CallDescriptor::kFixedTargetRegister),
reg == kJavaScriptCallCodeStartRegister);
__ JumpCodeObject(reg);
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
unwinding_info_writer_.MarkBlockWillExit();
frame_access_state()->ClearSPDelta();
frame_access_state()->SetFrameAccessToDefault();
break;
}
case kArchTailCallWasm: {
if (instr->InputAt(0)->IsImmediate()) {
Constant constant = i.ToConstant(instr->InputAt(0));
Address wasm_code = static_cast<Address>(constant.ToInt32());
__ Jump(wasm_code, constant.rmode());
} else {
__ Jump(i.InputRegister(0));
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
unwinding_info_writer_.MarkBlockWillExit();
frame_access_state()->ClearSPDelta();
frame_access_state()->SetFrameAccessToDefault();
break;
}
case kArchTailCallAddress: {
CHECK(!instr->InputAt(0)->IsImmediate());
Register reg = i.InputRegister(0);
DCHECK_IMPLIES(
instr->HasCallDescriptorFlag(CallDescriptor::kFixedTargetRegister),
reg == kJavaScriptCallCodeStartRegister);
__ Jump(reg);
unwinding_info_writer_.MarkBlockWillExit();
frame_access_state()->ClearSPDelta();
frame_access_state()->SetFrameAccessToDefault();
break;
}
case kArchCallJSFunction: {
Register func = i.InputRegister(0);
if (FLAG_debug_code) {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
// Check the function's context matches the context argument.
__ ldr(scratch, FieldMemOperand(func, JSFunction::kContextOffset));
__ cmp(cp, scratch);
__ Assert(eq, AbortReason::kWrongFunctionContext);
}
static_assert(kJavaScriptCallCodeStartRegister == r2, "ABI mismatch");
__ ldr(r2, FieldMemOperand(func, JSFunction::kCodeOffset));
__ CallCodeObject(r2);
RecordCallPosition(instr);
DCHECK_EQ(LeaveCC, i.OutputSBit());
frame_access_state()->ClearSPDelta();
break;
}
case kArchPrepareCallCFunction: {
int const num_parameters = MiscField::decode(instr->opcode());
__ PrepareCallCFunction(num_parameters);
// Frame alignment requires using FP-relative frame addressing.
frame_access_state()->SetFrameAccessToFP();
break;
}
case kArchSaveCallerRegisters: {
fp_mode_ =
static_cast<SaveFPRegsMode>(MiscField::decode(instr->opcode()));
DCHECK(fp_mode_ == kDontSaveFPRegs || fp_mode_ == kSaveFPRegs);
// kReturnRegister0 should have been saved before entering the stub.
int bytes = __ PushCallerSaved(fp_mode_, kReturnRegister0);
DCHECK(IsAligned(bytes, kSystemPointerSize));
DCHECK_EQ(0, frame_access_state()->sp_delta());
frame_access_state()->IncreaseSPDelta(bytes / kSystemPointerSize);
DCHECK(!caller_registers_saved_);
caller_registers_saved_ = true;
break;
}
case kArchRestoreCallerRegisters: {
DCHECK(fp_mode_ ==
static_cast<SaveFPRegsMode>(MiscField::decode(instr->opcode())));
DCHECK(fp_mode_ == kDontSaveFPRegs || fp_mode_ == kSaveFPRegs);
// Don't overwrite the returned value.
int bytes = __ PopCallerSaved(fp_mode_, kReturnRegister0);
frame_access_state()->IncreaseSPDelta(-(bytes / kSystemPointerSize));
DCHECK_EQ(0, frame_access_state()->sp_delta());
DCHECK(caller_registers_saved_);
caller_registers_saved_ = false;
break;
}
case kArchPrepareTailCall:
AssemblePrepareTailCall();
break;
case kArchCallCFunction: {
int const num_parameters = MiscField::decode(instr->opcode());
if (linkage()->GetIncomingDescriptor()->IsWasmCapiFunction()) {
// Put the return address in a stack slot.
__ str(pc, MemOperand(fp, WasmExitFrameConstants::kCallingPCOffset));
}
if (instr->InputAt(0)->IsImmediate()) {
ExternalReference ref = i.InputExternalReference(0);
__ CallCFunction(ref, num_parameters);
} else {
Register func = i.InputRegister(0);
__ CallCFunction(func, num_parameters);
}
if (linkage()->GetIncomingDescriptor()->IsWasmCapiFunction()) {
RecordSafepoint(instr->reference_map(), Safepoint::kNoLazyDeopt);
}
frame_access_state()->SetFrameAccessToDefault();
// Ideally, we should decrement SP delta to match the change of stack
// pointer in CallCFunction. However, for certain architectures (e.g.
// ARM), there may be more strict alignment requirement, causing old SP
// to be saved on the stack. In those cases, we can not calculate the SP
// delta statically.
frame_access_state()->ClearSPDelta();
if (caller_registers_saved_) {
// Need to re-sync SP delta introduced in kArchSaveCallerRegisters.
// Here, we assume the sequence to be:
// kArchSaveCallerRegisters;
// kArchCallCFunction;
// kArchRestoreCallerRegisters;
int bytes =
__ RequiredStackSizeForCallerSaved(fp_mode_, kReturnRegister0);
frame_access_state()->IncreaseSPDelta(bytes / kSystemPointerSize);
}
break;
}
case kArchJmp:
AssembleArchJump(i.InputRpo(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchBinarySearchSwitch:
AssembleArchBinarySearchSwitch(instr);
break;
case kArchTableSwitch:
AssembleArchTableSwitch(instr);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchAbortCSAAssert:
DCHECK(i.InputRegister(0) == r1);
{
// We don't actually want to generate a pile of code for this, so just
// claim there is a stack frame, without generating one.
FrameScope scope(tasm(), StackFrame::NONE);
__ Call(
isolate()->builtins()->builtin_handle(Builtins::kAbortCSAAssert),
RelocInfo::CODE_TARGET);
}
__ stop();
unwinding_info_writer_.MarkBlockWillExit();
break;
case kArchDebugBreak:
__ DebugBreak();
break;
case kArchComment:
__ RecordComment(reinterpret_cast<const char*>(i.InputInt32(0)));
break;
case kArchThrowTerminator:
DCHECK_EQ(LeaveCC, i.OutputSBit());
unwinding_info_writer_.MarkBlockWillExit();
break;
case kArchNop:
// don't emit code for nops.
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchDeoptimize: {
DeoptimizationExit* exit =
BuildTranslation(instr, -1, 0, OutputFrameStateCombine::Ignore());
__ b(exit->label());
break;
}
case kArchRet:
AssembleReturn(instr->InputAt(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchFramePointer:
__ mov(i.OutputRegister(), fp);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchParentFramePointer:
if (frame_access_state()->has_frame()) {
__ ldr(i.OutputRegister(), MemOperand(fp, 0));
} else {
__ mov(i.OutputRegister(), fp);
}
break;
case kArchStackPointerGreaterThan: {
// Potentially apply an offset to the current stack pointer before the
// comparison to consider the size difference of an optimized frame versus
// the contained unoptimized frames.
Register lhs_register = sp;
uint32_t offset;
if (ShouldApplyOffsetToStackCheck(instr, &offset)) {
lhs_register = i.TempRegister(0);
__ sub(lhs_register, sp, Operand(offset));
}
constexpr size_t kValueIndex = 0;
DCHECK(instr->InputAt(kValueIndex)->IsRegister());
__ cmp(lhs_register, i.InputRegister(kValueIndex));
break;
}
case kArchStackCheckOffset:
__ Move(i.OutputRegister(), Smi::FromInt(GetStackCheckOffset()));
break;
case kArchTruncateDoubleToI:
__ TruncateDoubleToI(isolate(), zone(), i.OutputRegister(),
i.InputDoubleRegister(0), DetermineStubCallMode());
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArchStoreWithWriteBarrier: {
RecordWriteMode mode =
static_cast<RecordWriteMode>(MiscField::decode(instr->opcode()));
Register object = i.InputRegister(0);
Register value = i.InputRegister(2);
AddressingMode addressing_mode =
AddressingModeField::decode(instr->opcode());
Operand offset(0);
if (addressing_mode == kMode_Offset_RI) {
int32_t immediate = i.InputInt32(1);
offset = Operand(immediate);
__ str(value, MemOperand(object, immediate));
} else {
DCHECK_EQ(kMode_Offset_RR, addressing_mode);
Register reg = i.InputRegister(1);
offset = Operand(reg);
__ str(value, MemOperand(object, reg));
}
auto ool = zone()->New<OutOfLineRecordWrite>(
this, object, offset, value, mode, DetermineStubCallMode(),
&unwinding_info_writer_);
__ CheckPageFlag(object, MemoryChunk::kPointersFromHereAreInterestingMask,
ne, ool->entry());
__ bind(ool->exit());
break;
}
case kArchStackSlot: {
FrameOffset offset =
frame_access_state()->GetFrameOffset(i.InputInt32(0));
Register base = offset.from_stack_pointer() ? sp : fp;
__ add(i.OutputRegister(0), base, Operand(offset.offset()));
break;
}
case kIeee754Float64Acos:
ASSEMBLE_IEEE754_UNOP(acos);
break;
case kIeee754Float64Acosh:
ASSEMBLE_IEEE754_UNOP(acosh);
break;
case kIeee754Float64Asin:
ASSEMBLE_IEEE754_UNOP(asin);
break;
case kIeee754Float64Asinh:
ASSEMBLE_IEEE754_UNOP(asinh);
break;
case kIeee754Float64Atan:
ASSEMBLE_IEEE754_UNOP(atan);
break;
case kIeee754Float64Atanh:
ASSEMBLE_IEEE754_UNOP(atanh);
break;
case kIeee754Float64Atan2:
ASSEMBLE_IEEE754_BINOP(atan2);
break;
case kIeee754Float64Cbrt:
ASSEMBLE_IEEE754_UNOP(cbrt);
break;
case kIeee754Float64Cos:
ASSEMBLE_IEEE754_UNOP(cos);
break;
case kIeee754Float64Cosh:
ASSEMBLE_IEEE754_UNOP(cosh);
break;
case kIeee754Float64Exp:
ASSEMBLE_IEEE754_UNOP(exp);
break;
case kIeee754Float64Expm1:
ASSEMBLE_IEEE754_UNOP(expm1);
break;
case kIeee754Float64Log:
ASSEMBLE_IEEE754_UNOP(log);
break;
case kIeee754Float64Log1p:
ASSEMBLE_IEEE754_UNOP(log1p);
break;
case kIeee754Float64Log2:
ASSEMBLE_IEEE754_UNOP(log2);
break;
case kIeee754Float64Log10:
ASSEMBLE_IEEE754_UNOP(log10);
break;
case kIeee754Float64Pow:
ASSEMBLE_IEEE754_BINOP(pow);
break;
case kIeee754Float64Sin:
ASSEMBLE_IEEE754_UNOP(sin);
break;
case kIeee754Float64Sinh:
ASSEMBLE_IEEE754_UNOP(sinh);
break;
case kIeee754Float64Tan:
ASSEMBLE_IEEE754_UNOP(tan);
break;
case kIeee754Float64Tanh:
ASSEMBLE_IEEE754_UNOP(tanh);
break;
case kArmAdd:
__ add(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmAnd:
__ and_(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmBic:
__ bic(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmMul:
__ mul(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.OutputSBit());
break;
case kArmMla:
__ mla(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputRegister(2), i.OutputSBit());
break;
case kArmMls: {
CpuFeatureScope scope(tasm(), ARMv7);
__ mls(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmSmull:
__ smull(i.OutputRegister(0), i.OutputRegister(1), i.InputRegister(0),
i.InputRegister(1));
break;
case kArmSmmul:
__ smmul(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmSmmla:
__ smmla(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmUmull:
__ umull(i.OutputRegister(0), i.OutputRegister(1), i.InputRegister(0),
i.InputRegister(1), i.OutputSBit());
break;
case kArmSdiv: {
CpuFeatureScope scope(tasm(), SUDIV);
__ sdiv(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmUdiv: {
CpuFeatureScope scope(tasm(), SUDIV);
__ udiv(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmMov:
__ Move(i.OutputRegister(), i.InputOperand2(0), i.OutputSBit());
break;
case kArmMvn:
__ mvn(i.OutputRegister(), i.InputOperand2(0), i.OutputSBit());
break;
case kArmOrr:
__ orr(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmEor:
__ eor(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmSub:
__ sub(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmRsb:
__ rsb(i.OutputRegister(), i.InputRegister(0), i.InputOperand2(1),
i.OutputSBit());
break;
case kArmBfc: {
CpuFeatureScope scope(tasm(), ARMv7);
__ bfc(i.OutputRegister(), i.InputInt8(1), i.InputInt8(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmUbfx: {
CpuFeatureScope scope(tasm(), ARMv7);
__ ubfx(i.OutputRegister(), i.InputRegister(0), i.InputInt8(1),
i.InputInt8(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmSbfx: {
CpuFeatureScope scope(tasm(), ARMv7);
__ sbfx(i.OutputRegister(), i.InputRegister(0), i.InputInt8(1),
i.InputInt8(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmSxtb:
__ sxtb(i.OutputRegister(), i.InputRegister(0), i.InputInt32(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmSxth:
__ sxth(i.OutputRegister(), i.InputRegister(0), i.InputInt32(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmSxtab:
__ sxtab(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputInt32(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmSxtah:
__ sxtah(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputInt32(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmUxtb:
__ uxtb(i.OutputRegister(), i.InputRegister(0), i.InputInt32(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmUxth:
__ uxth(i.OutputRegister(), i.InputRegister(0), i.InputInt32(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmUxtab:
__ uxtab(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputInt32(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmUxtah:
__ uxtah(i.OutputRegister(), i.InputRegister(0), i.InputRegister(1),
i.InputInt32(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmRbit: {
CpuFeatureScope scope(tasm(), ARMv7);
__ rbit(i.OutputRegister(), i.InputRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmRev:
__ rev(i.OutputRegister(), i.InputRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmClz:
__ clz(i.OutputRegister(), i.InputRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmCmp:
__ cmp(i.InputRegister(0), i.InputOperand2(1));
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmCmn:
__ cmn(i.InputRegister(0), i.InputOperand2(1));
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmTst:
__ tst(i.InputRegister(0), i.InputOperand2(1));
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmTeq:
__ teq(i.InputRegister(0), i.InputOperand2(1));
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmAddPair:
// i.InputRegister(0) ... left low word.
// i.InputRegister(1) ... left high word.
// i.InputRegister(2) ... right low word.
// i.InputRegister(3) ... right high word.
__ add(i.OutputRegister(0), i.InputRegister(0), i.InputRegister(2),
SBit::SetCC);
__ adc(i.OutputRegister(1), i.InputRegister(1),
Operand(i.InputRegister(3)));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmSubPair:
// i.InputRegister(0) ... left low word.
// i.InputRegister(1) ... left high word.
// i.InputRegister(2) ... right low word.
// i.InputRegister(3) ... right high word.
__ sub(i.OutputRegister(0), i.InputRegister(0), i.InputRegister(2),
SBit::SetCC);
__ sbc(i.OutputRegister(1), i.InputRegister(1),
Operand(i.InputRegister(3)));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmMulPair:
// i.InputRegister(0) ... left low word.
// i.InputRegister(1) ... left high word.
// i.InputRegister(2) ... right low word.
// i.InputRegister(3) ... right high word.
__ umull(i.OutputRegister(0), i.OutputRegister(1), i.InputRegister(0),
i.InputRegister(2));
__ mla(i.OutputRegister(1), i.InputRegister(0), i.InputRegister(3),
i.OutputRegister(1));
__ mla(i.OutputRegister(1), i.InputRegister(2), i.InputRegister(1),
i.OutputRegister(1));
break;
case kArmLslPair: {
Register second_output =
instr->OutputCount() >= 2 ? i.OutputRegister(1) : i.TempRegister(0);
if (instr->InputAt(2)->IsImmediate()) {
__ LslPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputInt32(2));
} else {
__ LslPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputRegister(2));
}
break;
}
case kArmLsrPair: {
Register second_output =
instr->OutputCount() >= 2 ? i.OutputRegister(1) : i.TempRegister(0);
if (instr->InputAt(2)->IsImmediate()) {
__ LsrPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputInt32(2));
} else {
__ LsrPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputRegister(2));
}
break;
}
case kArmAsrPair: {
Register second_output =
instr->OutputCount() >= 2 ? i.OutputRegister(1) : i.TempRegister(0);
if (instr->InputAt(2)->IsImmediate()) {
__ AsrPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputInt32(2));
} else {
__ AsrPair(i.OutputRegister(0), second_output, i.InputRegister(0),
i.InputRegister(1), i.InputRegister(2));
}
break;
}
case kArmVcmpF32:
if (instr->InputAt(1)->IsFPRegister()) {
__ VFPCompareAndSetFlags(i.InputFloatRegister(0),
i.InputFloatRegister(1));
} else {
DCHECK(instr->InputAt(1)->IsImmediate());
// 0.0 is the only immediate supported by vcmp instructions.
DCHECK_EQ(0.0f, i.InputFloat32(1));
__ VFPCompareAndSetFlags(i.InputFloatRegister(0), i.InputFloat32(1));
}
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmVaddF32:
__ vadd(i.OutputFloatRegister(), i.InputFloatRegister(0),
i.InputFloatRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVsubF32:
__ vsub(i.OutputFloatRegister(), i.InputFloatRegister(0),
i.InputFloatRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmulF32:
__ vmul(i.OutputFloatRegister(), i.InputFloatRegister(0),
i.InputFloatRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmlaF32:
__ vmla(i.OutputFloatRegister(), i.InputFloatRegister(1),
i.InputFloatRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmlsF32:
__ vmls(i.OutputFloatRegister(), i.InputFloatRegister(1),
i.InputFloatRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVdivF32:
__ vdiv(i.OutputFloatRegister(), i.InputFloatRegister(0),
i.InputFloatRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVsqrtF32:
__ vsqrt(i.OutputFloatRegister(), i.InputFloatRegister(0));
break;
case kArmVabsF32:
__ vabs(i.OutputFloatRegister(), i.InputFloatRegister(0));
break;
case kArmVnegF32:
__ vneg(i.OutputFloatRegister(), i.InputFloatRegister(0));
break;
case kArmVcmpF64:
if (instr->InputAt(1)->IsFPRegister()) {
__ VFPCompareAndSetFlags(i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
} else {
DCHECK(instr->InputAt(1)->IsImmediate());
// 0.0 is the only immediate supported by vcmp instructions.
DCHECK_EQ(0.0, i.InputDouble(1));
__ VFPCompareAndSetFlags(i.InputDoubleRegister(0), i.InputDouble(1));
}
DCHECK_EQ(SetCC, i.OutputSBit());
break;
case kArmVaddF64:
__ vadd(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVsubF64:
__ vsub(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmulF64:
__ vmul(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmlaF64:
__ vmla(i.OutputDoubleRegister(), i.InputDoubleRegister(1),
i.InputDoubleRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmlsF64:
__ vmls(i.OutputDoubleRegister(), i.InputDoubleRegister(1),
i.InputDoubleRegister(2));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVdivF64:
__ vdiv(i.OutputDoubleRegister(), i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmodF64: {
// TODO(bmeurer): We should really get rid of this special instruction,
// and generate a CallAddress instruction instead.
FrameScope scope(tasm(), StackFrame::MANUAL);
__ PrepareCallCFunction(0, 2);
__ MovToFloatParameters(i.InputDoubleRegister(0),
i.InputDoubleRegister(1));
__ CallCFunction(ExternalReference::mod_two_doubles_operation(), 0, 2);
// Move the result in the double result register.
__ MovFromFloatResult(i.OutputDoubleRegister());
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVsqrtF64:
__ vsqrt(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
case kArmVabsF64:
__ vabs(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
case kArmVnegF64:
__ vneg(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
case kArmVrintmF32: {
CpuFeatureScope scope(tasm(), ARMv8);
if (instr->InputAt(0)->IsSimd128Register()) {
__ vrintm(NeonS32, i.OutputSimd128Register(),
i.InputSimd128Register(0));
} else {
__ vrintm(i.OutputFloatRegister(), i.InputFloatRegister(0));
}
break;
}
case kArmVrintmF64: {
CpuFeatureScope scope(tasm(), ARMv8);
__ vrintm(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
}
case kArmVrintpF32: {
CpuFeatureScope scope(tasm(), ARMv8);
if (instr->InputAt(0)->IsSimd128Register()) {
__ vrintp(NeonS32, i.OutputSimd128Register(),
i.InputSimd128Register(0));
} else {
__ vrintp(i.OutputFloatRegister(), i.InputFloatRegister(0));
}
break;
}
case kArmVrintpF64: {
CpuFeatureScope scope(tasm(), ARMv8);
__ vrintp(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
}
case kArmVrintzF32: {
CpuFeatureScope scope(tasm(), ARMv8);
if (instr->InputAt(0)->IsSimd128Register()) {
__ vrintz(NeonS32, i.OutputSimd128Register(),
i.InputSimd128Register(0));
} else {
__ vrintz(i.OutputFloatRegister(), i.InputFloatRegister(0));
}
break;
}
case kArmVrintzF64: {
CpuFeatureScope scope(tasm(), ARMv8);
__ vrintz(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
}
case kArmVrintaF64: {
CpuFeatureScope scope(tasm(), ARMv8);
__ vrinta(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
}
case kArmVrintnF32: {
CpuFeatureScope scope(tasm(), ARMv8);
if (instr->InputAt(0)->IsSimd128Register()) {
__ vrintn(NeonS32, i.OutputSimd128Register(),
i.InputSimd128Register(0));
} else {
__ vrintn(i.OutputFloatRegister(), i.InputFloatRegister(0));
}
break;
}
case kArmVrintnF64: {
CpuFeatureScope scope(tasm(), ARMv8);
__ vrintn(i.OutputDoubleRegister(), i.InputDoubleRegister(0));
break;
}
case kArmVcvtF32F64: {
__ vcvt_f32_f64(i.OutputFloatRegister(), i.InputDoubleRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtF64F32: {
__ vcvt_f64_f32(i.OutputDoubleRegister(), i.InputFloatRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtF32S32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vmov(scratch, i.InputRegister(0));
__ vcvt_f32_s32(i.OutputFloatRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtF32U32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vmov(scratch, i.InputRegister(0));
__ vcvt_f32_u32(i.OutputFloatRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtF64S32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vmov(scratch, i.InputRegister(0));
__ vcvt_f64_s32(i.OutputDoubleRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtF64U32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vmov(scratch, i.InputRegister(0));
__ vcvt_f64_u32(i.OutputDoubleRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtS32F32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vcvt_s32_f32(scratch, i.InputFloatRegister(0));
__ vmov(i.OutputRegister(), scratch);
bool set_overflow_to_min_i32 = MiscField::decode(instr->opcode());
if (set_overflow_to_min_i32) {
// Avoid INT32_MAX as an overflow indicator and use INT32_MIN instead,
// because INT32_MIN allows easier out-of-bounds detection.
__ cmn(i.OutputRegister(), Operand(1));
__ mov(i.OutputRegister(), Operand(INT32_MIN), SBit::LeaveCC, vs);
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtU32F32: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vcvt_u32_f32(scratch, i.InputFloatRegister(0));
__ vmov(i.OutputRegister(), scratch);
bool set_overflow_to_min_u32 = MiscField::decode(instr->opcode());
if (set_overflow_to_min_u32) {
// Avoid UINT32_MAX as an overflow indicator and use 0 instead,
// because 0 allows easier out-of-bounds detection.
__ cmn(i.OutputRegister(), Operand(1));
__ adc(i.OutputRegister(), i.OutputRegister(), Operand::Zero());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtS32F64: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vcvt_s32_f64(scratch, i.InputDoubleRegister(0));
__ vmov(i.OutputRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVcvtU32F64: {
UseScratchRegisterScope temps(tasm());
SwVfpRegister scratch = temps.AcquireS();
__ vcvt_u32_f64(scratch, i.InputDoubleRegister(0));
__ vmov(i.OutputRegister(), scratch);
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVmovU32F32:
__ vmov(i.OutputRegister(), i.InputFloatRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovF32U32:
__ vmov(i.OutputFloatRegister(), i.InputRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovLowU32F64:
__ VmovLow(i.OutputRegister(), i.InputDoubleRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovLowF64U32:
__ VmovLow(i.OutputDoubleRegister(), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovHighU32F64:
__ VmovHigh(i.OutputRegister(), i.InputDoubleRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovHighF64U32:
__ VmovHigh(i.OutputDoubleRegister(), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovF64U32U32:
__ vmov(i.OutputDoubleRegister(), i.InputRegister(0), i.InputRegister(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVmovU32U32F64:
__ vmov(i.OutputRegister(0), i.OutputRegister(1),
i.InputDoubleRegister(0));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmLdrb:
__ ldrb(i.OutputRegister(), i.InputOffset());
DCHECK_EQ(LeaveCC, i.OutputSBit());
EmitWordLoadPoisoningIfNeeded(this, opcode, i);
break;
case kArmLdrsb:
__ ldrsb(i.OutputRegister(), i.InputOffset());
DCHECK_EQ(LeaveCC, i.OutputSBit());
EmitWordLoadPoisoningIfNeeded(this, opcode, i);
break;
case kArmStrb:
__ strb(i.InputRegister(0), i.InputOffset(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmLdrh:
__ ldrh(i.OutputRegister(), i.InputOffset());
EmitWordLoadPoisoningIfNeeded(this, opcode, i);
break;
case kArmLdrsh:
__ ldrsh(i.OutputRegister(), i.InputOffset());
EmitWordLoadPoisoningIfNeeded(this, opcode, i);
break;
case kArmStrh:
__ strh(i.InputRegister(0), i.InputOffset(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmLdr:
__ ldr(i.OutputRegister(), i.InputOffset());
EmitWordLoadPoisoningIfNeeded(this, opcode, i);
break;
case kArmStr:
__ str(i.InputRegister(0), i.InputOffset(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVldrF32: {
const MemoryAccessMode access_mode =
static_cast<MemoryAccessMode>(MiscField::decode(opcode));
if (access_mode == kMemoryAccessPoisoned) {
UseScratchRegisterScope temps(tasm());
Register address = temps.Acquire();
ComputePoisonedAddressForLoad(this, opcode, i, address);
__ vldr(i.OutputFloatRegister(), address, 0);
} else {
__ vldr(i.OutputFloatRegister(), i.InputOffset());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVstrF32:
__ vstr(i.InputFloatRegister(0), i.InputOffset(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmVld1F64: {
__ vld1(Neon8, NeonListOperand(i.OutputDoubleRegister()),
i.NeonInputOperand(0));
break;
}
case kArmVst1F64: {
__ vst1(Neon8, NeonListOperand(i.InputDoubleRegister(0)),
i.NeonInputOperand(1));
break;
}
case kArmVld1S128: {
__ vld1(Neon8, NeonListOperand(i.OutputSimd128Register()),
i.NeonInputOperand(0));
break;
}
case kArmVst1S128: {
__ vst1(Neon8, NeonListOperand(i.InputSimd128Register(0)),
i.NeonInputOperand(1));
break;
}
case kArmVldrF64: {
const MemoryAccessMode access_mode =
static_cast<MemoryAccessMode>(MiscField::decode(opcode));
if (access_mode == kMemoryAccessPoisoned) {
UseScratchRegisterScope temps(tasm());
Register address = temps.Acquire();
ComputePoisonedAddressForLoad(this, opcode, i, address);
__ vldr(i.OutputDoubleRegister(), address, 0);
} else {
__ vldr(i.OutputDoubleRegister(), i.InputOffset());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmVstrF64:
__ vstr(i.InputDoubleRegister(0), i.InputOffset(1));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmFloat32Max: {
SwVfpRegister result = i.OutputFloatRegister();
SwVfpRegister left = i.InputFloatRegister(0);
SwVfpRegister right = i.InputFloatRegister(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool = zone()->New<OutOfLineFloat32Max>(this, result, left, right);
__ FloatMax(result, left, right, ool->entry());
__ bind(ool->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmFloat64Max: {
DwVfpRegister result = i.OutputDoubleRegister();
DwVfpRegister left = i.InputDoubleRegister(0);
DwVfpRegister right = i.InputDoubleRegister(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool = zone()->New<OutOfLineFloat64Max>(this, result, left, right);
__ FloatMax(result, left, right, ool->entry());
__ bind(ool->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmFloat32Min: {
SwVfpRegister result = i.OutputFloatRegister();
SwVfpRegister left = i.InputFloatRegister(0);
SwVfpRegister right = i.InputFloatRegister(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool = zone()->New<OutOfLineFloat32Min>(this, result, left, right);
__ FloatMin(result, left, right, ool->entry());
__ bind(ool->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmFloat64Min: {
DwVfpRegister result = i.OutputDoubleRegister();
DwVfpRegister left = i.InputDoubleRegister(0);
DwVfpRegister right = i.InputDoubleRegister(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool = zone()->New<OutOfLineFloat64Min>(this, result, left, right);
__ FloatMin(result, left, right, ool->entry());
__ bind(ool->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmFloat64SilenceNaN: {
DwVfpRegister value = i.InputDoubleRegister(0);
DwVfpRegister result = i.OutputDoubleRegister();
__ VFPCanonicalizeNaN(result, value);
break;
}
case kArmPush:
if (instr->InputAt(0)->IsFPRegister()) {
LocationOperand* op = LocationOperand::cast(instr->InputAt(0));
switch (op->representation()) {
case MachineRepresentation::kFloat32:
__ vpush(i.InputFloatRegister(0));
frame_access_state()->IncreaseSPDelta(1);
break;
case MachineRepresentation::kFloat64:
__ vpush(i.InputDoubleRegister(0));
frame_access_state()->IncreaseSPDelta(kDoubleSize /
kSystemPointerSize);
break;
case MachineRepresentation::kSimd128: {
__ vpush(i.InputSimd128Register(0));
frame_access_state()->IncreaseSPDelta(kSimd128Size /
kSystemPointerSize);
break;
}
default:
UNREACHABLE();
break;
}
} else {
__ push(i.InputRegister(0));
frame_access_state()->IncreaseSPDelta(1);
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
case kArmPoke: {
int const slot = MiscField::decode(instr->opcode());
__ str(i.InputRegister(0), MemOperand(sp, slot * kSystemPointerSize));
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmPeek: {
int reverse_slot = i.InputInt32(0);
int offset =
FrameSlotToFPOffset(frame()->GetTotalFrameSlotCount() - reverse_slot);
if (instr->OutputAt(0)->IsFPRegister()) {
LocationOperand* op = LocationOperand::cast(instr->OutputAt(0));
if (op->representation() == MachineRepresentation::kFloat64) {
__ vldr(i.OutputDoubleRegister(), MemOperand(fp, offset));
} else if (op->representation() == MachineRepresentation::kFloat32) {
__ vldr(i.OutputFloatRegister(), MemOperand(fp, offset));
} else {
DCHECK_EQ(MachineRepresentation::kSimd128, op->representation());
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ add(scratch, fp, Operand(offset));
__ vld1(Neon8, NeonListOperand(i.OutputSimd128Register()),
NeonMemOperand(scratch));
}
} else {
__ ldr(i.OutputRegister(), MemOperand(fp, offset));
}
break;
}
case kArmDmbIsh: {
__ dmb(ISH);
break;
}
case kArmDsbIsb: {
__ dsb(SY);
__ isb(SY);
break;
}
case kArchWordPoisonOnSpeculation:
__ and_(i.OutputRegister(0), i.InputRegister(0),
Operand(kSpeculationPoisonRegister));
break;
case kArmF64x2Splat: {
Simd128Register dst = i.OutputSimd128Register();
DoubleRegister src = i.InputDoubleRegister(0);
__ Move(dst.low(), src);
__ Move(dst.high(), src);
break;
}
case kArmF64x2ExtractLane: {
__ ExtractLane(i.OutputDoubleRegister(), i.InputSimd128Register(0),
i.InputInt8(1));
break;
}
case kArmF64x2ReplaceLane: {
__ ReplaceLane(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputDoubleRegister(2), i.InputInt8(1));
break;
}
case kArmF64x2Abs: {
__ vabs(i.OutputSimd128Register().low(), i.InputSimd128Register(0).low());
__ vabs(i.OutputSimd128Register().high(),
i.InputSimd128Register(0).high());
break;
}
case kArmF64x2Neg: {
__ vneg(i.OutputSimd128Register().low(), i.InputSimd128Register(0).low());
__ vneg(i.OutputSimd128Register().high(),
i.InputSimd128Register(0).high());
break;
}
case kArmF64x2Sqrt: {
__ vsqrt(i.OutputSimd128Register().low(),
i.InputSimd128Register(0).low());
__ vsqrt(i.OutputSimd128Register().high(),
i.InputSimd128Register(0).high());
break;
}
case kArmF64x2Add: {
ASSEMBLE_F64X2_ARITHMETIC_BINOP(vadd);
break;
}
case kArmF64x2Sub: {
ASSEMBLE_F64X2_ARITHMETIC_BINOP(vsub);
break;
}
case kArmF64x2Mul: {
ASSEMBLE_F64X2_ARITHMETIC_BINOP(vmul);
break;
}
case kArmF64x2Div: {
ASSEMBLE_F64X2_ARITHMETIC_BINOP(vdiv);
break;
}
case kArmF64x2Min: {
Simd128Register result = i.OutputSimd128Register();
Simd128Register left = i.InputSimd128Register(0);
Simd128Register right = i.InputSimd128Register(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool_low = zone()->New<OutOfLineFloat64Min>(
this, result.low(), left.low(), right.low());
auto ool_high = zone()->New<OutOfLineFloat64Min>(
this, result.high(), left.high(), right.high());
__ FloatMin(result.low(), left.low(), right.low(), ool_low->entry());
__ bind(ool_low->exit());
__ FloatMin(result.high(), left.high(), right.high(),
ool_high->entry());
__ bind(ool_high->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
case kArmF64x2Max: {
Simd128Register result = i.OutputSimd128Register();
Simd128Register left = i.InputSimd128Register(0);
Simd128Register right = i.InputSimd128Register(1);
if (left == right) {
__ Move(result, left);
} else {
auto ool_low = zone()->New<OutOfLineFloat64Max>(
this, result.low(), left.low(), right.low());
auto ool_high = zone()->New<OutOfLineFloat64Max>(
this, result.high(), left.high(), right.high());
__ FloatMax(result.low(), left.low(), right.low(), ool_low->entry());
__ bind(ool_low->exit());
__ FloatMax(result.high(), left.high(), right.high(),
ool_high->entry());
__ bind(ool_high->exit());
}
DCHECK_EQ(LeaveCC, i.OutputSBit());
break;
}
#undef ASSEMBLE_F64X2_ARITHMETIC_BINOP
case kArmF64x2Eq: {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ mov(scratch, Operand(0));
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).low(),
i.InputSimd128Register(1).low());
__ mov(scratch, Operand(-1), LeaveCC, eq);
__ vmov(i.OutputSimd128Register().low(), scratch, scratch);
__ mov(scratch, Operand(0));
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).high(),
i.InputSimd128Register(1).high());
__ mov(scratch, Operand(-1), LeaveCC, eq);
__ vmov(i.OutputSimd128Register().high(), scratch, scratch);
break;
}
case kArmF64x2Ne: {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ mov(scratch, Operand(0));
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).low(),
i.InputSimd128Register(1).low());
__ mov(scratch, Operand(-1), LeaveCC, ne);
__ vmov(i.OutputSimd128Register().low(), scratch, scratch);
__ mov(scratch, Operand(0));
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).high(),
i.InputSimd128Register(1).high());
__ mov(scratch, Operand(-1), LeaveCC, ne);
__ vmov(i.OutputSimd128Register().high(), scratch, scratch);
break;
}
case kArmF64x2Lt: {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).low(),
i.InputSimd128Register(1).low());
__ mov(scratch, Operand(0), LeaveCC, cs);
__ mov(scratch, Operand(-1), LeaveCC, mi);
__ vmov(i.OutputSimd128Register().low(), scratch, scratch);
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).high(),
i.InputSimd128Register(1).high());
__ mov(scratch, Operand(0), LeaveCC, cs);
__ mov(scratch, Operand(-1), LeaveCC, mi);
__ vmov(i.OutputSimd128Register().high(), scratch, scratch);
break;
}
case kArmF64x2Le: {
UseScratchRegisterScope temps(tasm());
Register scratch = temps.Acquire();
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).low(),
i.InputSimd128Register(1).low());
__ mov(scratch, Operand(0), LeaveCC, hi);
__ mov(scratch, Operand(-1), LeaveCC, ls);
__ vmov(i.OutputSimd128Register().low(), scratch, scratch);
__ VFPCompareAndSetFlags(i.InputSimd128Register(0).high(),
i.InputSimd128Register(1).high());
__ mov(scratch, Operand(0), LeaveCC, hi);
__ mov(scratch, Operand(-1), LeaveCC, ls);
__ vmov(i.OutputSimd128Register().high(), scratch, scratch);
break;
}
case kArmF64x2Pmin: {
Simd128Register dst = i.OutputSimd128Register();
Simd128Register lhs = i.InputSimd128Register(0);
Simd128Register rhs = i.InputSimd128Register(1);
DCHECK_EQ(dst, lhs);
// Move rhs only when rhs is strictly lesser (mi).
__ VFPCompareAndSetFlags(rhs.low(), lhs.low());
__ vmov(dst.low(), rhs.low(), mi);
__ VFPCompareAndSetFlags(rhs.high(), lhs.high());
__ vmov(dst.high(), rhs.high(), mi);
break;
}
case kArmF64x2Pmax: {
Simd128Register dst = i.OutputSimd128Register();
Simd128Register lhs = i.InputSimd128Register(0);
Simd128Register rhs = i.InputSimd128Register(1);
DCHECK_EQ(dst, lhs);
// Move rhs only when rhs is strictly greater (gt).
__ VFPCompareAndSetFlags(rhs.low(), lhs.low());
__ vmov(dst.low(), rhs.low(), gt);
__ VFPCompareAndSetFlags(rhs.high(), lhs.high());
__ vmov(dst.high(), rhs.high(), gt);
break;
}
case kArmF64x2Ceil: {
CpuFeatureScope scope(tasm(), ARMv8);
Simd128Register dst = i.OutputSimd128Register();
Simd128Register src = i.InputSimd128Register(0);
__ vrintp(dst.low(), src.low());
__ vrintp(dst.high(), src.high());
break;
}
case kArmF64x2Floor: {
CpuFeatureScope scope(tasm(), ARMv8);
Simd128Register dst = i.OutputSimd128Register();
Simd128Register src = i.InputSimd128Register(0);
__ vrintm(dst.low(), src.low());
__ vrintm(dst.high(), src.high());
break;
}
case kArmF64x2Trunc: {
CpuFeatureScope scope(tasm(), ARMv8);
Simd128Register dst = i.OutputSimd128Register();
Simd128Register src = i.InputSimd128Register(0);
__ vrintz(dst.low(), src.low());
__ vrintz(dst.high(), src.high());
break;
}
case kArmF64x2NearestInt: {
CpuFeatureScope scope(tasm(), ARMv8);
Simd128Register dst = i.OutputSimd128Register();
Simd128Register src = i.InputSimd128Register(0);
__ vrintn(dst.low(), src.low());
__ vrintn(dst.high(), src.high());
break;
}
case kArmI64x2SplatI32Pair: {
Simd128Register dst = i.OutputSimd128Register();
__ vdup(Neon32, dst, i.InputRegister(0));
__ ReplaceLane(dst, dst, i.InputRegister(1), NeonS32, 1);
__ ReplaceLane(dst, dst, i.InputRegister(1), NeonS32, 3);
break;
}
case kArmI64x2ReplaceLaneI32Pair: {
Simd128Register dst = i.OutputSimd128Register();
int8_t lane = i.InputInt8(1);
__ ReplaceLane(dst, dst, i.InputRegister(2), NeonS32, lane * 2);
__ ReplaceLane(dst, dst, i.InputRegister(3), NeonS32, lane * 2 + 1);
break;
}
case kArmI64x2Add: {
__ vadd(Neon64, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI64x2Sub: {
__ vsub(Neon64, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI64x2Mul: {
QwNeonRegister dst = i.OutputSimd128Register();
QwNeonRegister left = i.InputSimd128Register(0);
QwNeonRegister right = i.InputSimd128Register(1);
QwNeonRegister tmp1 = i.TempSimd128Register(0);
QwNeonRegister tmp2 = i.TempSimd128Register(1);
// This algorithm uses vector operations to perform 64-bit integer
// multiplication by splitting it into a high and low 32-bit integers.
// The tricky part is getting the low and high integers in the correct
// place inside a NEON register, so that we can use as little vmull and
// vmlal as possible.
// Move left and right into temporaries, they will be modified by vtrn.
__ vmov(tmp1, left);
__ vmov(tmp2, right);
// This diagram shows how the 64-bit integers fit into NEON registers.
//
// [q.high()| q.low()]
// left/tmp1: [ a3, a2 | a1, a0 ]
// right/tmp2: [ b3, b2 | b1, b0 ]
//
// We want to multiply the low 32 bits of left with high 32 bits of right,
// for each lane, i.e. a2 * b3, a0 * b1. However, vmull takes two input d
// registers, and multiply the corresponding low/high 32 bits, to get a
// 64-bit integer: a1 * b1, a0 * b0. In order to make it work we transpose
// the vectors, so that we get the low 32 bits of each 64-bit integer into
// the same lane, similarly for high 32 bits.
__ vtrn(Neon32, tmp1.low(), tmp1.high());
// tmp1: [ a3, a1 | a2, a0 ]
__ vtrn(Neon32, tmp2.low(), tmp2.high());
// tmp2: [ b3, b1 | b2, b0 ]
__ vmull(NeonU32, dst, tmp1.low(), tmp2.high());
// dst: [ a2*b3 | a0*b1 ]
__ vmlal(NeonU32, dst, tmp1.high(), tmp2.low());
// dst: [ a2*b3 + a3*b2 | a0*b1 + a1*b0 ]
__ vshl(NeonU64, dst, dst, 32);
// dst: [ (a2*b3 + a3*b2) << 32 | (a0*b1 + a1*b0) << 32 ]
__ vmlal(NeonU32, dst, tmp1.low(), tmp2.low());
// dst: [ (a2*b3 + a3*b2)<<32 + (a2*b2) | (a0*b1 + a1*b0)<<32 + (a0*b0) ]
break;
}
case kArmI64x2Neg: {
Simd128Register dst = i.OutputSimd128Register();
__ vmov(dst, uint64_t{0});
__ vsub(Neon64, dst, dst, i.InputSimd128Register(0));
break;
}
case kArmI64x2Shl: {
ASSEMBLE_SIMD_SHIFT_LEFT(vshl, 6, Neon32, NeonS64);
break;
}
case kArmI64x2ShrS: {
// Only the least significant byte of each lane is used, so we can use
// Neon32 as the size.
ASSEMBLE_SIMD_SHIFT_RIGHT(vshr, 6, Neon32, NeonS64);
break;
}
case kArmI64x2ShrU: {
// Only the least significant byte of each lane is used, so we can use
// Neon32 as the size.
ASSEMBLE_SIMD_SHIFT_RIGHT(vshr, 6, Neon32, NeonU64);
break;
}
case kArmF32x4Splat: {
int src_code = i.InputFloatRegister(0).code();
__ vdup(Neon32, i.OutputSimd128Register(),
DwVfpRegister::from_code(src_code / 2), src_code % 2);
break;
}
case kArmF32x4ExtractLane: {
__ ExtractLane(i.OutputFloatRegister(), i.InputSimd128Register(0),
i.InputInt8(1));
break;
}
case kArmF32x4ReplaceLane: {
__ ReplaceLane(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputFloatRegister(2), i.InputInt8(1));
break;
}
case kArmF32x4SConvertI32x4: {
__ vcvt_f32_s32(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4UConvertI32x4: {
__ vcvt_f32_u32(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4Abs: {
__ vabs(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4Neg: {
__ vneg(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4Sqrt: {
QwNeonRegister dst = i.OutputSimd128Register();
QwNeonRegister src1 = i.InputSimd128Register(0);
DCHECK_EQ(dst, q0);
DCHECK_EQ(src1, q0);
#define S_FROM_Q(reg, lane) SwVfpRegister::from_code(reg.code() * 4 + lane)
__ vsqrt(S_FROM_Q(dst, 0), S_FROM_Q(src1, 0));
__ vsqrt(S_FROM_Q(dst, 1), S_FROM_Q(src1, 1));
__ vsqrt(S_FROM_Q(dst, 2), S_FROM_Q(src1, 2));
__ vsqrt(S_FROM_Q(dst, 3), S_FROM_Q(src1, 3));
#undef S_FROM_Q
break;
}
case kArmF32x4RecipApprox: {
__ vrecpe(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4RecipSqrtApprox: {
__ vrsqrte(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmF32x4Add: {
__ vadd(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4AddHoriz: {
Simd128Register dst = i.OutputSimd128Register(),
src0 = i.InputSimd128Register(0),
src1 = i.InputSimd128Register(1);
// Make sure we don't overwrite source data before it's used.
if (dst == src0) {
__ vpadd(dst.low(), src0.low(), src0.high());
if (dst == src1) {
__ vmov(dst.high(), dst.low());
} else {
__ vpadd(dst.high(), src1.low(), src1.high());
}
} else {
__ vpadd(dst.high(), src1.low(), src1.high());
__ vpadd(dst.low(), src0.low(), src0.high());
}
break;
}
case kArmF32x4Sub: {
__ vsub(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4Mul: {
__ vmul(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4Div: {
QwNeonRegister dst = i.OutputSimd128Register();
QwNeonRegister src1 = i.InputSimd128Register(0);
QwNeonRegister src2 = i.InputSimd128Register(1);
DCHECK_EQ(dst, q0);
DCHECK_EQ(src1, q0);
DCHECK_EQ(src2, q1);
#define S_FROM_Q(reg, lane) SwVfpRegister::from_code(reg.code() * 4 + lane)
__ vdiv(S_FROM_Q(dst, 0), S_FROM_Q(src1, 0), S_FROM_Q(src2, 0));
__ vdiv(S_FROM_Q(dst, 1), S_FROM_Q(src1, 1), S_FROM_Q(src2, 1));
__ vdiv(S_FROM_Q(dst, 2), S_FROM_Q(src1, 2), S_FROM_Q(src2, 2));
__ vdiv(S_FROM_Q(dst, 3), S_FROM_Q(src1, 3), S_FROM_Q(src2, 3));
#undef S_FROM_Q
break;
}
case kArmF32x4Min: {
__ vmin(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4Max: {
__ vmax(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4Eq: {
__ vceq(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmF32x4Ne: {
Simd128Register dst = i.OutputSimd128Register();
__ vceq(dst, i.InputSimd128Register(0), i.InputSimd128Register(1));
__ vmvn(dst, dst);
break;
}
case kArmF32x4Lt: {
__ vcgt(i.OutputSimd128Register(), i.InputSimd128Register(1),
i.InputSimd128Register(0));
break;
}
case kArmF32x4Le: {
__ vcge(i.OutputSimd128Register(), i.InputSimd128Register(1),
i.InputSimd128Register(0));
break;
}
case kArmF32x4Pmin: {
Simd128Register dst = i.OutputSimd128Register();
Simd128Register lhs = i.InputSimd128Register(0);
Simd128Register rhs = i.InputSimd128Register(1);
DCHECK_NE(dst, lhs);
DCHECK_NE(dst, rhs);
// f32x4.pmin(lhs, rhs)
// = v128.bitselect(rhs, lhs, f32x4.lt(rhs, lhs))
// = v128.bitselect(rhs, lhs, f32x4.gt(lhs, rhs))
__ vcgt(dst, lhs, rhs);
__ vbsl(dst, rhs, lhs);
break;
}
case kArmF32x4Pmax: {
Simd128Register dst = i.OutputSimd128Register();
Simd128Register lhs = i.InputSimd128Register(0);
Simd128Register rhs = i.InputSimd128Register(1);
DCHECK_NE(dst, lhs);
DCHECK_NE(dst, rhs);
// f32x4.pmax(lhs, rhs)
// = v128.bitselect(rhs, lhs, f32x4.gt(rhs, lhs))
__ vcgt(dst, rhs, lhs);
__ vbsl(dst, rhs, lhs);
break;
}
case kArmI32x4Splat: {
__ vdup(Neon32, i.OutputSimd128Register(), i.InputRegister(0));
break;
}
case kArmI32x4ExtractLane: {
__ ExtractLane(i.OutputRegister(), i.InputSimd128Register(0), NeonS32,
i.InputInt8(1));
break;
}
case kArmI32x4ReplaceLane: {
__ ReplaceLane(i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputRegister(2), NeonS32, i.InputInt8(1));
break;
}
case kArmI32x4SConvertF32x4: {
__ vcvt_s32_f32(i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmI32x4SConvertI16x8Low: {
__ vmovl(NeonS16, i.OutputSimd128Register(),
i.InputSimd128Register(0).low());
break;
}
case kArmI32x4SConvertI16x8High: {
__ vmovl(NeonS16, i.OutputSimd128Register(),
i.InputSimd128Register(0).high());
break;
}
case kArmI32x4Neg: {
__ vneg(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0));
break;
}
case kArmI32x4Shl: {
ASSEMBLE_SIMD_SHIFT_LEFT(vshl, 5, Neon32, NeonS32);
break;
}
case kArmI32x4ShrS: {
ASSEMBLE_SIMD_SHIFT_RIGHT(vshr, 5, Neon32, NeonS32);
break;
}
case kArmI32x4Add: {
__ vadd(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4AddHoriz:
ASSEMBLE_NEON_PAIRWISE_OP(vpadd, Neon32);
break;
case kArmI32x4Sub: {
__ vsub(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4Mul: {
__ vmul(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4MinS: {
__ vmin(NeonS32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4MaxS: {
__ vmax(NeonS32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4Eq: {
__ vceq(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4Ne: {
Simd128Register dst = i.OutputSimd128Register();
__ vceq(Neon32, dst, i.InputSimd128Register(0),
i.InputSimd128Register(1));
__ vmvn(dst, dst);
break;
}
case kArmI32x4GtS: {
__ vcgt(NeonS32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register(1));
break;
}
case kArmI32x4GeS: {
__ vcge(NeonS32, i.OutputSimd128Register(), i.InputSimd128Register(0),
i.InputSimd128Register