| // 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 <algorithm> |
| |
| #include "src/base/iterator.h" |
| #include "src/base/logging.h" |
| #include "src/base/overflowing-math.h" |
| #include "src/base/platform/wrappers.h" |
| #include "src/codegen/cpu-features.h" |
| #include "src/codegen/machine-type.h" |
| #include "src/compiler/backend/instruction-codes.h" |
| #include "src/compiler/backend/instruction-selector-impl.h" |
| #include "src/compiler/backend/instruction.h" |
| #include "src/compiler/machine-operator.h" |
| #include "src/compiler/node-matchers.h" |
| #include "src/compiler/node-properties.h" |
| #include "src/compiler/opcodes.h" |
| #include "src/roots/roots-inl.h" |
| |
| #if V8_ENABLE_WEBASSEMBLY |
| #include "src/wasm/simd-shuffle.h" |
| #endif // V8_ENABLE_WEBASSEMBLY |
| |
| namespace v8 { |
| namespace internal { |
| namespace compiler { |
| |
| // Adds X64-specific methods for generating operands. |
| class X64OperandGenerator final : public OperandGenerator { |
| public: |
| explicit X64OperandGenerator(InstructionSelector* selector) |
| : OperandGenerator(selector) {} |
| |
| bool CanBeImmediate(Node* node) { |
| switch (node->opcode()) { |
| case IrOpcode::kInt32Constant: |
| case IrOpcode::kRelocatableInt32Constant: { |
| const int32_t value = OpParameter<int32_t>(node->op()); |
| // int32_t min will overflow if displacement mode is |
| // kNegativeDisplacement. |
| return value != std::numeric_limits<int32_t>::min(); |
| } |
| case IrOpcode::kInt64Constant: { |
| const int64_t value = OpParameter<int64_t>(node->op()); |
| return std::numeric_limits<int32_t>::min() < value && |
| value <= std::numeric_limits<int32_t>::max(); |
| } |
| case IrOpcode::kNumberConstant: { |
| const double value = OpParameter<double>(node->op()); |
| return bit_cast<int64_t>(value) == 0; |
| } |
| default: |
| return false; |
| } |
| } |
| |
| int32_t GetImmediateIntegerValue(Node* node) { |
| DCHECK(CanBeImmediate(node)); |
| if (node->opcode() == IrOpcode::kInt32Constant) { |
| return OpParameter<int32_t>(node->op()); |
| } |
| DCHECK_EQ(IrOpcode::kInt64Constant, node->opcode()); |
| return static_cast<int32_t>(OpParameter<int64_t>(node->op())); |
| } |
| |
| bool CanBeMemoryOperand(InstructionCode opcode, Node* node, Node* input, |
| int effect_level) { |
| if ((input->opcode() != IrOpcode::kLoad && |
| input->opcode() != IrOpcode::kLoadImmutable) || |
| !selector()->CanCover(node, input)) { |
| return false; |
| } |
| if (effect_level != selector()->GetEffectLevel(input)) { |
| return false; |
| } |
| MachineRepresentation rep = |
| LoadRepresentationOf(input->op()).representation(); |
| switch (opcode) { |
| case kX64And: |
| case kX64Or: |
| case kX64Xor: |
| case kX64Add: |
| case kX64Sub: |
| case kX64Push: |
| case kX64Cmp: |
| case kX64Test: |
| // When pointer compression is enabled 64-bit memory operands can't be |
| // used for tagged values. |
| return rep == MachineRepresentation::kWord64 || |
| (!COMPRESS_POINTERS_BOOL && IsAnyTagged(rep)); |
| case kX64And32: |
| case kX64Or32: |
| case kX64Xor32: |
| case kX64Add32: |
| case kX64Sub32: |
| case kX64Cmp32: |
| case kX64Test32: |
| // When pointer compression is enabled 32-bit memory operands can be |
| // used for tagged values. |
| return rep == MachineRepresentation::kWord32 || |
| (COMPRESS_POINTERS_BOOL && |
| (IsAnyTagged(rep) || IsAnyCompressed(rep))); |
| case kAVXFloat64Add: |
| case kAVXFloat64Sub: |
| case kAVXFloat64Mul: |
| DCHECK_EQ(MachineRepresentation::kFloat64, rep); |
| return true; |
| case kAVXFloat32Add: |
| case kAVXFloat32Sub: |
| case kAVXFloat32Mul: |
| DCHECK_EQ(MachineRepresentation::kFloat32, rep); |
| return true; |
| case kX64Cmp16: |
| case kX64Test16: |
| return rep == MachineRepresentation::kWord16; |
| case kX64Cmp8: |
| case kX64Test8: |
| return rep == MachineRepresentation::kWord8; |
| default: |
| break; |
| } |
| return false; |
| } |
| |
| AddressingMode GenerateMemoryOperandInputs( |
| Node* index, int scale_exponent, Node* base, Node* displacement, |
| DisplacementMode displacement_mode, InstructionOperand inputs[], |
| size_t* input_count, |
| RegisterUseKind reg_kind = RegisterUseKind::kUseRegister) { |
| AddressingMode mode = kMode_MRI; |
| if (base != nullptr && (index != nullptr || displacement != nullptr)) { |
| if (base->opcode() == IrOpcode::kInt32Constant && |
| OpParameter<int32_t>(base->op()) == 0) { |
| base = nullptr; |
| } else if (base->opcode() == IrOpcode::kInt64Constant && |
| OpParameter<int64_t>(base->op()) == 0) { |
| base = nullptr; |
| } |
| } |
| if (base != nullptr) { |
| inputs[(*input_count)++] = UseRegister(base, reg_kind); |
| if (index != nullptr) { |
| DCHECK(scale_exponent >= 0 && scale_exponent <= 3); |
| inputs[(*input_count)++] = UseRegister(index, reg_kind); |
| if (displacement != nullptr) { |
| inputs[(*input_count)++] = displacement_mode == kNegativeDisplacement |
| ? UseNegatedImmediate(displacement) |
| : UseImmediate(displacement); |
| static const AddressingMode kMRnI_modes[] = {kMode_MR1I, kMode_MR2I, |
| kMode_MR4I, kMode_MR8I}; |
| mode = kMRnI_modes[scale_exponent]; |
| } else { |
| static const AddressingMode kMRn_modes[] = {kMode_MR1, kMode_MR2, |
| kMode_MR4, kMode_MR8}; |
| mode = kMRn_modes[scale_exponent]; |
| } |
| } else { |
| if (displacement == nullptr) { |
| mode = kMode_MR; |
| } else { |
| inputs[(*input_count)++] = displacement_mode == kNegativeDisplacement |
| ? UseNegatedImmediate(displacement) |
| : UseImmediate(displacement); |
| mode = kMode_MRI; |
| } |
| } |
| } else { |
| DCHECK(scale_exponent >= 0 && scale_exponent <= 3); |
| if (displacement != nullptr) { |
| if (index == nullptr) { |
| inputs[(*input_count)++] = UseRegister(displacement, reg_kind); |
| mode = kMode_MR; |
| } else { |
| inputs[(*input_count)++] = UseRegister(index, reg_kind); |
| inputs[(*input_count)++] = displacement_mode == kNegativeDisplacement |
| ? UseNegatedImmediate(displacement) |
| : UseImmediate(displacement); |
| static const AddressingMode kMnI_modes[] = {kMode_MRI, kMode_M2I, |
| kMode_M4I, kMode_M8I}; |
| mode = kMnI_modes[scale_exponent]; |
| } |
| } else { |
| inputs[(*input_count)++] = UseRegister(index, reg_kind); |
| static const AddressingMode kMn_modes[] = {kMode_MR, kMode_MR1, |
| kMode_M4, kMode_M8}; |
| mode = kMn_modes[scale_exponent]; |
| if (mode == kMode_MR1) { |
| // [%r1 + %r1*1] has a smaller encoding than [%r1*2+0] |
| inputs[(*input_count)++] = UseRegister(index, reg_kind); |
| } |
| } |
| } |
| return mode; |
| } |
| |
| AddressingMode GetEffectiveAddressMemoryOperand( |
| Node* operand, InstructionOperand inputs[], size_t* input_count, |
| RegisterUseKind reg_kind = RegisterUseKind::kUseRegister) { |
| { |
| LoadMatcher<ExternalReferenceMatcher> m(operand); |
| if (m.index().HasResolvedValue() && m.object().HasResolvedValue() && |
| selector()->CanAddressRelativeToRootsRegister( |
| m.object().ResolvedValue())) { |
| ptrdiff_t const delta = |
| m.index().ResolvedValue() + |
| TurboAssemblerBase::RootRegisterOffsetForExternalReference( |
| selector()->isolate(), m.object().ResolvedValue()); |
| if (is_int32(delta)) { |
| inputs[(*input_count)++] = TempImmediate(static_cast<int32_t>(delta)); |
| return kMode_Root; |
| } |
| } |
| } |
| BaseWithIndexAndDisplacement64Matcher m(operand, AddressOption::kAllowAll); |
| DCHECK(m.matches()); |
| if (m.displacement() == nullptr || CanBeImmediate(m.displacement())) { |
| return GenerateMemoryOperandInputs( |
| m.index(), m.scale(), m.base(), m.displacement(), |
| m.displacement_mode(), inputs, input_count, reg_kind); |
| } else if (m.base() == nullptr && |
| m.displacement_mode() == kPositiveDisplacement) { |
| // The displacement cannot be an immediate, but we can use the |
| // displacement as base instead and still benefit from addressing |
| // modes for the scale. |
| return GenerateMemoryOperandInputs(m.index(), m.scale(), m.displacement(), |
| nullptr, m.displacement_mode(), inputs, |
| input_count, reg_kind); |
| } else { |
| inputs[(*input_count)++] = UseRegister(operand->InputAt(0), reg_kind); |
| inputs[(*input_count)++] = UseRegister(operand->InputAt(1), reg_kind); |
| return kMode_MR1; |
| } |
| } |
| |
| InstructionOperand GetEffectiveIndexOperand(Node* index, |
| AddressingMode* mode) { |
| if (CanBeImmediate(index)) { |
| *mode = kMode_MRI; |
| return UseImmediate(index); |
| } else { |
| *mode = kMode_MR1; |
| return UseUniqueRegister(index); |
| } |
| } |
| |
| bool CanBeBetterLeftOperand(Node* node) const { |
| return !selector()->IsLive(node); |
| } |
| }; |
| |
| namespace { |
| |
| ArchOpcode GetLoadOpcode(LoadRepresentation load_rep) { |
| ArchOpcode opcode; |
| switch (load_rep.representation()) { |
| case MachineRepresentation::kFloat32: |
| opcode = kX64Movss; |
| break; |
| case MachineRepresentation::kFloat64: |
| opcode = kX64Movsd; |
| break; |
| case MachineRepresentation::kBit: // Fall through. |
| case MachineRepresentation::kWord8: |
| opcode = load_rep.IsSigned() ? kX64Movsxbl : kX64Movzxbl; |
| break; |
| case MachineRepresentation::kWord16: |
| opcode = load_rep.IsSigned() ? kX64Movsxwl : kX64Movzxwl; |
| break; |
| case MachineRepresentation::kWord32: |
| opcode = kX64Movl; |
| break; |
| case MachineRepresentation::kCompressedPointer: // Fall through. |
| case MachineRepresentation::kCompressed: |
| #ifdef V8_COMPRESS_POINTERS |
| opcode = kX64Movl; |
| break; |
| #else |
| UNREACHABLE(); |
| #endif |
| #ifdef V8_COMPRESS_POINTERS |
| case MachineRepresentation::kTaggedSigned: |
| opcode = kX64MovqDecompressTaggedSigned; |
| break; |
| case MachineRepresentation::kTaggedPointer: |
| opcode = kX64MovqDecompressTaggedPointer; |
| break; |
| case MachineRepresentation::kTagged: |
| opcode = kX64MovqDecompressAnyTagged; |
| break; |
| #else |
| case MachineRepresentation::kTaggedSigned: // Fall through. |
| case MachineRepresentation::kTaggedPointer: // Fall through. |
| case MachineRepresentation::kTagged: // Fall through. |
| #endif |
| case MachineRepresentation::kWord64: |
| opcode = kX64Movq; |
| break; |
| case MachineRepresentation::kSandboxedPointer: |
| opcode = kX64MovqDecodeSandboxedPointer; |
| break; |
| case MachineRepresentation::kSimd128: |
| opcode = kX64Movdqu; |
| break; |
| case MachineRepresentation::kNone: // Fall through. |
| case MachineRepresentation::kMapWord: |
| UNREACHABLE(); |
| } |
| return opcode; |
| } |
| |
| ArchOpcode GetStoreOpcode(StoreRepresentation store_rep) { |
| switch (store_rep.representation()) { |
| case MachineRepresentation::kFloat32: |
| return kX64Movss; |
| case MachineRepresentation::kFloat64: |
| return kX64Movsd; |
| case MachineRepresentation::kBit: // Fall through. |
| case MachineRepresentation::kWord8: |
| return kX64Movb; |
| case MachineRepresentation::kWord16: |
| return kX64Movw; |
| case MachineRepresentation::kWord32: |
| return kX64Movl; |
| case MachineRepresentation::kCompressedPointer: // Fall through. |
| case MachineRepresentation::kCompressed: |
| #ifdef V8_COMPRESS_POINTERS |
| return kX64MovqCompressTagged; |
| #else |
| UNREACHABLE(); |
| #endif |
| case MachineRepresentation::kTaggedSigned: // Fall through. |
| case MachineRepresentation::kTaggedPointer: // Fall through. |
| case MachineRepresentation::kTagged: |
| return kX64MovqCompressTagged; |
| case MachineRepresentation::kWord64: |
| return kX64Movq; |
| case MachineRepresentation::kSandboxedPointer: |
| return kX64MovqEncodeSandboxedPointer; |
| case MachineRepresentation::kSimd128: |
| return kX64Movdqu; |
| case MachineRepresentation::kNone: // Fall through. |
| case MachineRepresentation::kMapWord: |
| UNREACHABLE(); |
| } |
| UNREACHABLE(); |
| } |
| |
| ArchOpcode GetSeqCstStoreOpcode(StoreRepresentation store_rep) { |
| switch (store_rep.representation()) { |
| case MachineRepresentation::kWord8: |
| return kAtomicStoreWord8; |
| case MachineRepresentation::kWord16: |
| return kAtomicStoreWord16; |
| case MachineRepresentation::kWord32: |
| return kAtomicStoreWord32; |
| case MachineRepresentation::kWord64: |
| return kX64Word64AtomicStoreWord64; |
| case MachineRepresentation::kTaggedSigned: // Fall through. |
| case MachineRepresentation::kTaggedPointer: // Fall through. |
| case MachineRepresentation::kTagged: |
| if (COMPRESS_POINTERS_BOOL) return kAtomicStoreWord32; |
| return kX64Word64AtomicStoreWord64; |
| case MachineRepresentation::kCompressedPointer: // Fall through. |
| case MachineRepresentation::kCompressed: |
| CHECK(COMPRESS_POINTERS_BOOL); |
| return kAtomicStoreWord32; |
| default: |
| UNREACHABLE(); |
| } |
| } |
| |
| } // namespace |
| |
| void InstructionSelector::VisitStackSlot(Node* node) { |
| StackSlotRepresentation rep = StackSlotRepresentationOf(node->op()); |
| int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment()); |
| OperandGenerator g(this); |
| |
| Emit(kArchStackSlot, g.DefineAsRegister(node), |
| sequence()->AddImmediate(Constant(slot)), 0, nullptr); |
| } |
| |
| void InstructionSelector::VisitAbortCSADcheck(Node* node) { |
| X64OperandGenerator g(this); |
| Emit(kArchAbortCSADcheck, g.NoOutput(), g.UseFixed(node->InputAt(0), rdx)); |
| } |
| |
| void InstructionSelector::VisitLoadLane(Node* node) { |
| LoadLaneParameters params = LoadLaneParametersOf(node->op()); |
| InstructionCode opcode = kArchNop; |
| if (params.rep == MachineType::Int8()) { |
| opcode = kX64Pinsrb; |
| } else if (params.rep == MachineType::Int16()) { |
| opcode = kX64Pinsrw; |
| } else if (params.rep == MachineType::Int32()) { |
| opcode = kX64Pinsrd; |
| } else if (params.rep == MachineType::Int64()) { |
| opcode = kX64Pinsrq; |
| } else { |
| UNREACHABLE(); |
| } |
| |
| X64OperandGenerator g(this); |
| InstructionOperand outputs[] = {g.DefineAsRegister(node)}; |
| // Input 0 is value node, 1 is lane idx, and GetEffectiveAddressMemoryOperand |
| // uses up to 3 inputs. This ordering is consistent with other operations that |
| // use the same opcode. |
| InstructionOperand inputs[5]; |
| size_t input_count = 0; |
| |
| inputs[input_count++] = g.UseRegister(node->InputAt(2)); |
| inputs[input_count++] = g.UseImmediate(params.laneidx); |
| |
| AddressingMode mode = |
| g.GetEffectiveAddressMemoryOperand(node, inputs, &input_count); |
| opcode |= AddressingModeField::encode(mode); |
| |
| DCHECK_GE(5, input_count); |
| |
| // x64 supports unaligned loads. |
| DCHECK_NE(params.kind, MemoryAccessKind::kUnaligned); |
| if (params.kind == MemoryAccessKind::kProtected) { |
| opcode |= AccessModeField::encode(kMemoryAccessProtected); |
| } |
| Emit(opcode, 1, outputs, input_count, inputs); |
| } |
| |
| void InstructionSelector::VisitLoadTransform(Node* node) { |
| LoadTransformParameters params = LoadTransformParametersOf(node->op()); |
| ArchOpcode opcode; |
| switch (params.transformation) { |
| case LoadTransformation::kS128Load8Splat: |
| opcode = kX64S128Load8Splat; |
| break; |
| case LoadTransformation::kS128Load16Splat: |
| opcode = kX64S128Load16Splat; |
| break; |
| case LoadTransformation::kS128Load32Splat: |
| opcode = kX64S128Load32Splat; |
| break; |
| case LoadTransformation::kS128Load64Splat: |
| opcode = kX64S128Load64Splat; |
| break; |
| case LoadTransformation::kS128Load8x8S: |
| opcode = kX64S128Load8x8S; |
| break; |
| case LoadTransformation::kS128Load8x8U: |
| opcode = kX64S128Load8x8U; |
| break; |
| case LoadTransformation::kS128Load16x4S: |
| opcode = kX64S128Load16x4S; |
| break; |
| case LoadTransformation::kS128Load16x4U: |
| opcode = kX64S128Load16x4U; |
| break; |
| case LoadTransformation::kS128Load32x2S: |
| opcode = kX64S128Load32x2S; |
| break; |
| case LoadTransformation::kS128Load32x2U: |
| opcode = kX64S128Load32x2U; |
| break; |
| case LoadTransformation::kS128Load32Zero: |
| opcode = kX64Movss; |
| break; |
| case LoadTransformation::kS128Load64Zero: |
| opcode = kX64Movsd; |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| // x64 supports unaligned loads |
| DCHECK_NE(params.kind, MemoryAccessKind::kUnaligned); |
| InstructionCode code = opcode; |
| if (params.kind == MemoryAccessKind::kProtected) { |
| code |= AccessModeField::encode(kMemoryAccessProtected); |
| } |
| VisitLoad(node, node, code); |
| } |
| |
| void InstructionSelector::VisitLoad(Node* node, Node* value, |
| InstructionCode opcode) { |
| X64OperandGenerator g(this); |
| #ifdef V8_IS_TSAN |
| // On TSAN builds we require one scratch register. Because of this we also |
| // have to modify the inputs to take into account possible aliasing and use |
| // UseUniqueRegister which is not required for non-TSAN builds. |
| InstructionOperand temps[] = {g.TempRegister()}; |
| size_t temp_count = arraysize(temps); |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseUniqueRegister; |
| #else |
| InstructionOperand* temps = nullptr; |
| size_t temp_count = 0; |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseRegister; |
| #endif // V8_IS_TSAN |
| InstructionOperand outputs[] = {g.DefineAsRegister(node)}; |
| InstructionOperand inputs[3]; |
| size_t input_count = 0; |
| AddressingMode mode = |
| g.GetEffectiveAddressMemoryOperand(value, inputs, &input_count, reg_kind); |
| InstructionCode code = opcode | AddressingModeField::encode(mode); |
| if (node->opcode() == IrOpcode::kProtectedLoad) { |
| code |= AccessModeField::encode(kMemoryAccessProtected); |
| } |
| Emit(code, 1, outputs, input_count, inputs, temp_count, temps); |
| } |
| |
| void InstructionSelector::VisitLoad(Node* node) { |
| LoadRepresentation load_rep = LoadRepresentationOf(node->op()); |
| DCHECK(!load_rep.IsMapWord()); |
| VisitLoad(node, node, GetLoadOpcode(load_rep)); |
| } |
| |
| void InstructionSelector::VisitProtectedLoad(Node* node) { VisitLoad(node); } |
| |
| namespace { |
| |
| // Shared routine for Word32/Word64 Atomic Exchange |
| void VisitAtomicExchange(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode, AtomicWidth width) { |
| X64OperandGenerator g(selector); |
| Node* base = node->InputAt(0); |
| Node* index = node->InputAt(1); |
| Node* value = node->InputAt(2); |
| AddressingMode addressing_mode; |
| InstructionOperand inputs[] = { |
| g.UseUniqueRegister(value), g.UseUniqueRegister(base), |
| g.GetEffectiveIndexOperand(index, &addressing_mode)}; |
| InstructionOperand outputs[] = {g.DefineSameAsFirst(node)}; |
| InstructionCode code = opcode | AddressingModeField::encode(addressing_mode) | |
| AtomicWidthField::encode(width); |
| selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs); |
| } |
| |
| void VisitStoreCommon(InstructionSelector* selector, Node* node, |
| StoreRepresentation store_rep, |
| base::Optional<AtomicMemoryOrder> atomic_order) { |
| X64OperandGenerator g(selector); |
| Node* base = node->InputAt(0); |
| Node* index = node->InputAt(1); |
| Node* value = node->InputAt(2); |
| |
| DCHECK_NE(store_rep.representation(), MachineRepresentation::kMapWord); |
| WriteBarrierKind write_barrier_kind = store_rep.write_barrier_kind(); |
| const bool is_seqcst = |
| atomic_order && *atomic_order == AtomicMemoryOrder::kSeqCst; |
| |
| if (FLAG_enable_unconditional_write_barriers && |
| CanBeTaggedOrCompressedPointer(store_rep.representation())) { |
| write_barrier_kind = kFullWriteBarrier; |
| } |
| |
| if (write_barrier_kind != kNoWriteBarrier && !FLAG_disable_write_barriers) { |
| DCHECK(CanBeTaggedOrCompressedPointer(store_rep.representation())); |
| AddressingMode addressing_mode; |
| InstructionOperand inputs[] = { |
| g.UseUniqueRegister(base), |
| g.GetEffectiveIndexOperand(index, &addressing_mode), |
| g.UseUniqueRegister(value)}; |
| RecordWriteMode record_write_mode = |
| WriteBarrierKindToRecordWriteMode(write_barrier_kind); |
| InstructionOperand temps[] = {g.TempRegister(), g.TempRegister()}; |
| InstructionCode code = is_seqcst ? kArchAtomicStoreWithWriteBarrier |
| : kArchStoreWithWriteBarrier; |
| code |= AddressingModeField::encode(addressing_mode); |
| code |= MiscField::encode(static_cast<int>(record_write_mode)); |
| selector->Emit(code, 0, nullptr, arraysize(inputs), inputs, |
| arraysize(temps), temps); |
| } else { |
| #ifdef V8_IS_TSAN |
| // On TSAN builds we require two scratch registers. Because of this we also |
| // have to modify the inputs to take into account possible aliasing and use |
| // UseUniqueRegister which is not required for non-TSAN builds. |
| InstructionOperand temps[] = {g.TempRegister(), g.TempRegister()}; |
| size_t temp_count = arraysize(temps); |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseUniqueRegister; |
| #else |
| InstructionOperand* temps = nullptr; |
| size_t temp_count = 0; |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseRegister; |
| #endif // V8_IS_TSAN |
| |
| // Release and non-atomic stores emit MOV and sequentially consistent stores |
| // emit XCHG. |
| // https://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html |
| |
| ArchOpcode opcode; |
| AddressingMode addressing_mode; |
| InstructionOperand inputs[4]; |
| size_t input_count = 0; |
| |
| if (is_seqcst) { |
| // SeqCst stores emit XCHG instead of MOV, so encode the inputs as we |
| // would for XCHG. XCHG can't encode the value as an immediate and has |
| // fewer addressing modes available. |
| inputs[input_count++] = g.UseUniqueRegister(value); |
| inputs[input_count++] = g.UseUniqueRegister(base); |
| inputs[input_count++] = |
| g.GetEffectiveIndexOperand(index, &addressing_mode); |
| opcode = GetSeqCstStoreOpcode(store_rep); |
| } else { |
| if ((ElementSizeLog2Of(store_rep.representation()) < |
| kSystemPointerSizeLog2) && |
| value->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| value = value->InputAt(0); |
| } |
| |
| addressing_mode = g.GetEffectiveAddressMemoryOperand( |
| node, inputs, &input_count, reg_kind); |
| InstructionOperand value_operand = g.CanBeImmediate(value) |
| ? g.UseImmediate(value) |
| : g.UseRegister(value, reg_kind); |
| inputs[input_count++] = value_operand; |
| opcode = GetStoreOpcode(store_rep); |
| } |
| |
| InstructionCode code = |
| opcode | AddressingModeField::encode(addressing_mode); |
| selector->Emit(code, 0, static_cast<InstructionOperand*>(nullptr), |
| input_count, inputs, temp_count, temps); |
| } |
| } |
| |
| } // namespace |
| |
| void InstructionSelector::VisitStore(Node* node) { |
| return VisitStoreCommon(this, node, StoreRepresentationOf(node->op()), |
| base::nullopt); |
| } |
| |
| void InstructionSelector::VisitProtectedStore(Node* node) { |
| X64OperandGenerator g(this); |
| Node* value = node->InputAt(2); |
| StoreRepresentation store_rep = StoreRepresentationOf(node->op()); |
| |
| #ifdef V8_IS_TSAN |
| // On TSAN builds we require two scratch registers. Because of this we also |
| // have to modify the inputs to take into account possible aliasing and use |
| // UseUniqueRegister which is not required for non-TSAN builds. |
| InstructionOperand temps[] = {g.TempRegister(), g.TempRegister()}; |
| size_t temp_count = arraysize(temps); |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseUniqueRegister; |
| #else |
| InstructionOperand* temps = nullptr; |
| size_t temp_count = 0; |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseRegister; |
| #endif // V8_IS_TSAN |
| |
| InstructionOperand inputs[4]; |
| size_t input_count = 0; |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(node, inputs, &input_count, reg_kind); |
| InstructionOperand value_operand = g.CanBeImmediate(value) |
| ? g.UseImmediate(value) |
| : g.UseRegister(value, reg_kind); |
| inputs[input_count++] = value_operand; |
| ArchOpcode opcode = GetStoreOpcode(store_rep); |
| InstructionCode code = opcode | AddressingModeField::encode(addressing_mode) | |
| AccessModeField::encode(kMemoryAccessProtected); |
| Emit(code, 0, static_cast<InstructionOperand*>(nullptr), input_count, inputs, |
| temp_count, temps); |
| } |
| |
| // Architecture supports unaligned access, therefore VisitLoad is used instead |
| void InstructionSelector::VisitUnalignedLoad(Node* node) { UNREACHABLE(); } |
| |
| // Architecture supports unaligned access, therefore VisitStore is used instead |
| void InstructionSelector::VisitUnalignedStore(Node* node) { UNREACHABLE(); } |
| |
| void InstructionSelector::VisitStoreLane(Node* node) { |
| X64OperandGenerator g(this); |
| |
| StoreLaneParameters params = StoreLaneParametersOf(node->op()); |
| InstructionCode opcode = kArchNop; |
| if (params.rep == MachineRepresentation::kWord8) { |
| opcode = kX64Pextrb; |
| } else if (params.rep == MachineRepresentation::kWord16) { |
| opcode = kX64Pextrw; |
| } else if (params.rep == MachineRepresentation::kWord32) { |
| opcode = kX64S128Store32Lane; |
| } else if (params.rep == MachineRepresentation::kWord64) { |
| opcode = kX64S128Store64Lane; |
| } else { |
| UNREACHABLE(); |
| } |
| |
| InstructionOperand inputs[4]; |
| size_t input_count = 0; |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(node, inputs, &input_count); |
| opcode |= AddressingModeField::encode(addressing_mode); |
| |
| if (params.kind == MemoryAccessKind::kProtected) { |
| opcode |= AccessModeField::encode(kMemoryAccessProtected); |
| } |
| |
| InstructionOperand value_operand = g.UseRegister(node->InputAt(2)); |
| inputs[input_count++] = value_operand; |
| inputs[input_count++] = g.UseImmediate(params.laneidx); |
| DCHECK_GE(4, input_count); |
| Emit(opcode, 0, nullptr, input_count, inputs); |
| } |
| |
| // Shared routine for multiple binary operations. |
| static void VisitBinop(InstructionSelector* selector, Node* node, |
| InstructionCode opcode, FlagsContinuation* cont) { |
| X64OperandGenerator g(selector); |
| Int32BinopMatcher m(node); |
| Node* left = m.left().node(); |
| Node* right = m.right().node(); |
| InstructionOperand inputs[8]; |
| size_t input_count = 0; |
| InstructionOperand outputs[1]; |
| size_t output_count = 0; |
| |
| // TODO(turbofan): match complex addressing modes. |
| if (left == right) { |
| // If both inputs refer to the same operand, enforce allocating a register |
| // for both of them to ensure that we don't end up generating code like |
| // this: |
| // |
| // mov rax, [rbp-0x10] |
| // add rax, [rbp-0x10] |
| // jo label |
| InstructionOperand const input = g.UseRegister(left); |
| inputs[input_count++] = input; |
| inputs[input_count++] = input; |
| } else if (g.CanBeImmediate(right)) { |
| inputs[input_count++] = g.UseRegister(left); |
| inputs[input_count++] = g.UseImmediate(right); |
| } else { |
| int effect_level = selector->GetEffectLevel(node, cont); |
| if (node->op()->HasProperty(Operator::kCommutative) && |
| g.CanBeBetterLeftOperand(right) && |
| (!g.CanBeBetterLeftOperand(left) || |
| !g.CanBeMemoryOperand(opcode, node, right, effect_level))) { |
| std::swap(left, right); |
| } |
| if (g.CanBeMemoryOperand(opcode, node, right, effect_level)) { |
| inputs[input_count++] = g.UseRegister(left); |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(right, inputs, &input_count); |
| opcode |= AddressingModeField::encode(addressing_mode); |
| } else { |
| inputs[input_count++] = g.UseRegister(left); |
| inputs[input_count++] = g.Use(right); |
| } |
| } |
| |
| if (cont->IsBranch()) { |
| inputs[input_count++] = g.Label(cont->true_block()); |
| inputs[input_count++] = g.Label(cont->false_block()); |
| } |
| |
| outputs[output_count++] = g.DefineSameAsFirst(node); |
| |
| DCHECK_NE(0u, input_count); |
| DCHECK_EQ(1u, output_count); |
| DCHECK_GE(arraysize(inputs), input_count); |
| DCHECK_GE(arraysize(outputs), output_count); |
| |
| selector->EmitWithContinuation(opcode, output_count, outputs, input_count, |
| inputs, cont); |
| } |
| |
| // Shared routine for multiple binary operations. |
| static void VisitBinop(InstructionSelector* selector, Node* node, |
| InstructionCode opcode) { |
| FlagsContinuation cont; |
| VisitBinop(selector, node, opcode, &cont); |
| } |
| |
| void InstructionSelector::VisitWord32And(Node* node) { |
| X64OperandGenerator g(this); |
| Uint32BinopMatcher m(node); |
| if (m.right().Is(0xFF)) { |
| Emit(kX64Movzxbl, g.DefineAsRegister(node), g.Use(m.left().node())); |
| } else if (m.right().Is(0xFFFF)) { |
| Emit(kX64Movzxwl, g.DefineAsRegister(node), g.Use(m.left().node())); |
| } else { |
| VisitBinop(this, node, kX64And32); |
| } |
| } |
| |
| void InstructionSelector::VisitWord64And(Node* node) { |
| VisitBinop(this, node, kX64And); |
| } |
| |
| void InstructionSelector::VisitWord32Or(Node* node) { |
| VisitBinop(this, node, kX64Or32); |
| } |
| |
| void InstructionSelector::VisitWord64Or(Node* node) { |
| VisitBinop(this, node, kX64Or); |
| } |
| |
| void InstructionSelector::VisitWord32Xor(Node* node) { |
| X64OperandGenerator g(this); |
| Uint32BinopMatcher m(node); |
| if (m.right().Is(-1)) { |
| Emit(kX64Not32, g.DefineSameAsFirst(node), g.UseRegister(m.left().node())); |
| } else { |
| VisitBinop(this, node, kX64Xor32); |
| } |
| } |
| |
| void InstructionSelector::VisitWord64Xor(Node* node) { |
| X64OperandGenerator g(this); |
| Uint64BinopMatcher m(node); |
| if (m.right().Is(-1)) { |
| Emit(kX64Not, g.DefineSameAsFirst(node), g.UseRegister(m.left().node())); |
| } else { |
| VisitBinop(this, node, kX64Xor); |
| } |
| } |
| |
| void InstructionSelector::VisitStackPointerGreaterThan( |
| Node* node, FlagsContinuation* cont) { |
| StackCheckKind kind = StackCheckKindOf(node->op()); |
| InstructionCode opcode = |
| kArchStackPointerGreaterThan | MiscField::encode(static_cast<int>(kind)); |
| |
| int effect_level = GetEffectLevel(node, cont); |
| |
| X64OperandGenerator g(this); |
| Node* const value = node->InputAt(0); |
| if (g.CanBeMemoryOperand(kX64Cmp, node, value, effect_level)) { |
| DCHECK(IrOpcode::kLoad == value->opcode() || |
| IrOpcode::kLoadImmutable == value->opcode()); |
| |
| // GetEffectiveAddressMemoryOperand can create at most 3 inputs. |
| static constexpr int kMaxInputCount = 3; |
| |
| size_t input_count = 0; |
| InstructionOperand inputs[kMaxInputCount]; |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(value, inputs, &input_count); |
| opcode |= AddressingModeField::encode(addressing_mode); |
| DCHECK_LE(input_count, kMaxInputCount); |
| |
| EmitWithContinuation(opcode, 0, nullptr, input_count, inputs, cont); |
| } else { |
| EmitWithContinuation(opcode, g.UseRegister(value), cont); |
| } |
| } |
| |
| namespace { |
| |
| bool TryMergeTruncateInt64ToInt32IntoLoad(InstructionSelector* selector, |
| Node* node, Node* load) { |
| if ((load->opcode() == IrOpcode::kLoad || |
| load->opcode() == IrOpcode::kLoadImmutable) && |
| selector->CanCover(node, load)) { |
| LoadRepresentation load_rep = LoadRepresentationOf(load->op()); |
| MachineRepresentation rep = load_rep.representation(); |
| InstructionCode opcode; |
| switch (rep) { |
| case MachineRepresentation::kBit: // Fall through. |
| case MachineRepresentation::kWord8: |
| opcode = load_rep.IsSigned() ? kX64Movsxbl : kX64Movzxbl; |
| break; |
| case MachineRepresentation::kWord16: |
| opcode = load_rep.IsSigned() ? kX64Movsxwl : kX64Movzxwl; |
| break; |
| case MachineRepresentation::kWord32: |
| case MachineRepresentation::kWord64: |
| case MachineRepresentation::kTaggedSigned: |
| case MachineRepresentation::kTagged: |
| case MachineRepresentation::kCompressed: // Fall through. |
| opcode = kX64Movl; |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| X64OperandGenerator g(selector); |
| #ifdef V8_IS_TSAN |
| // On TSAN builds we require one scratch register. Because of this we also |
| // have to modify the inputs to take into account possible aliasing and use |
| // UseUniqueRegister which is not required for non-TSAN builds. |
| InstructionOperand temps[] = {g.TempRegister()}; |
| size_t temp_count = arraysize(temps); |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseUniqueRegister; |
| #else |
| InstructionOperand* temps = nullptr; |
| size_t temp_count = 0; |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseRegister; |
| #endif // V8_IS_TSAN |
| InstructionOperand outputs[] = {g.DefineAsRegister(node)}; |
| size_t input_count = 0; |
| InstructionOperand inputs[3]; |
| AddressingMode mode = g.GetEffectiveAddressMemoryOperand( |
| node->InputAt(0), inputs, &input_count, reg_kind); |
| opcode |= AddressingModeField::encode(mode); |
| selector->Emit(opcode, 1, outputs, input_count, inputs, temp_count, temps); |
| return true; |
| } |
| return false; |
| } |
| |
| // Shared routine for multiple 32-bit shift operations. |
| // TODO(bmeurer): Merge this with VisitWord64Shift using template magic? |
| void VisitWord32Shift(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| Int32BinopMatcher m(node); |
| Node* left = m.left().node(); |
| Node* right = m.right().node(); |
| |
| if (left->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| left = left->InputAt(0); |
| } |
| |
| if (g.CanBeImmediate(right)) { |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(left), |
| g.UseImmediate(right)); |
| } else { |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(left), |
| g.UseFixed(right, rcx)); |
| } |
| } |
| |
| // Shared routine for multiple 64-bit shift operations. |
| // TODO(bmeurer): Merge this with VisitWord32Shift using template magic? |
| void VisitWord64Shift(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| Int64BinopMatcher m(node); |
| Node* left = m.left().node(); |
| Node* right = m.right().node(); |
| |
| if (g.CanBeImmediate(right)) { |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(left), |
| g.UseImmediate(right)); |
| } else { |
| if (m.right().IsWord64And()) { |
| Int64BinopMatcher mright(right); |
| if (mright.right().Is(0x3F)) { |
| right = mright.left().node(); |
| } |
| } |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(left), |
| g.UseFixed(right, rcx)); |
| } |
| } |
| |
| // Shared routine for multiple shift operations with continuation. |
| template <typename BinopMatcher, int Bits> |
| bool TryVisitWordShift(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode, FlagsContinuation* cont) { |
| X64OperandGenerator g(selector); |
| BinopMatcher m(node); |
| Node* left = m.left().node(); |
| Node* right = m.right().node(); |
| |
| // If the shift count is 0, the flags are not affected. |
| if (!g.CanBeImmediate(right) || |
| (g.GetImmediateIntegerValue(right) & (Bits - 1)) == 0) { |
| return false; |
| } |
| InstructionOperand output = g.DefineSameAsFirst(node); |
| InstructionOperand inputs[2]; |
| inputs[0] = g.UseRegister(left); |
| inputs[1] = g.UseImmediate(right); |
| selector->EmitWithContinuation(opcode, 1, &output, 2, inputs, cont); |
| return true; |
| } |
| |
| void EmitLea(InstructionSelector* selector, InstructionCode opcode, |
| Node* result, Node* index, int scale, Node* base, |
| Node* displacement, DisplacementMode displacement_mode) { |
| X64OperandGenerator g(selector); |
| |
| InstructionOperand inputs[4]; |
| size_t input_count = 0; |
| AddressingMode mode = |
| g.GenerateMemoryOperandInputs(index, scale, base, displacement, |
| displacement_mode, inputs, &input_count); |
| |
| DCHECK_NE(0u, input_count); |
| DCHECK_GE(arraysize(inputs), input_count); |
| |
| InstructionOperand outputs[1]; |
| outputs[0] = g.DefineAsRegister(result); |
| |
| opcode = AddressingModeField::encode(mode) | opcode; |
| |
| selector->Emit(opcode, 1, outputs, input_count, inputs); |
| } |
| |
| } // namespace |
| |
| void InstructionSelector::VisitWord32Shl(Node* node) { |
| Int32ScaleMatcher m(node, true); |
| if (m.matches()) { |
| Node* index = node->InputAt(0); |
| Node* base = m.power_of_two_plus_one() ? index : nullptr; |
| EmitLea(this, kX64Lea32, node, index, m.scale(), base, nullptr, |
| kPositiveDisplacement); |
| return; |
| } |
| VisitWord32Shift(this, node, kX64Shl32); |
| } |
| |
| void InstructionSelector::VisitWord64Shl(Node* node) { |
| X64OperandGenerator g(this); |
| Int64ScaleMatcher m(node, true); |
| if (m.matches()) { |
| Node* index = node->InputAt(0); |
| Node* base = m.power_of_two_plus_one() ? index : nullptr; |
| EmitLea(this, kX64Lea, node, index, m.scale(), base, nullptr, |
| kPositiveDisplacement); |
| return; |
| } else { |
| Int64BinopMatcher bm(node); |
| if ((bm.left().IsChangeInt32ToInt64() || |
| bm.left().IsChangeUint32ToUint64()) && |
| bm.right().IsInRange(32, 63)) { |
| // There's no need to sign/zero-extend to 64-bit if we shift out the upper |
| // 32 bits anyway. |
| Emit(kX64Shl, g.DefineSameAsFirst(node), |
| g.UseRegister(bm.left().node()->InputAt(0)), |
| g.UseImmediate(bm.right().node())); |
| return; |
| } |
| } |
| VisitWord64Shift(this, node, kX64Shl); |
| } |
| |
| void InstructionSelector::VisitWord32Shr(Node* node) { |
| VisitWord32Shift(this, node, kX64Shr32); |
| } |
| |
| namespace { |
| |
| inline AddressingMode AddDisplacementToAddressingMode(AddressingMode mode) { |
| switch (mode) { |
| case kMode_MR: |
| return kMode_MRI; |
| case kMode_MR1: |
| return kMode_MR1I; |
| case kMode_MR2: |
| return kMode_MR2I; |
| case kMode_MR4: |
| return kMode_MR4I; |
| case kMode_MR8: |
| return kMode_MR8I; |
| case kMode_M1: |
| return kMode_M1I; |
| case kMode_M2: |
| return kMode_M2I; |
| case kMode_M4: |
| return kMode_M4I; |
| case kMode_M8: |
| return kMode_M8I; |
| case kMode_None: |
| case kMode_MRI: |
| case kMode_MR1I: |
| case kMode_MR2I: |
| case kMode_MR4I: |
| case kMode_MR8I: |
| case kMode_M1I: |
| case kMode_M2I: |
| case kMode_M4I: |
| case kMode_M8I: |
| case kMode_Root: |
| UNREACHABLE(); |
| } |
| UNREACHABLE(); |
| } |
| |
| bool TryMatchLoadWord64AndShiftRight(InstructionSelector* selector, Node* node, |
| InstructionCode opcode) { |
| DCHECK(IrOpcode::kWord64Sar == node->opcode() || |
| IrOpcode::kWord64Shr == node->opcode()); |
| X64OperandGenerator g(selector); |
| Int64BinopMatcher m(node); |
| if (selector->CanCover(m.node(), m.left().node()) && m.left().IsLoad() && |
| m.right().Is(32)) { |
| DCHECK_EQ(selector->GetEffectLevel(node), |
| selector->GetEffectLevel(m.left().node())); |
| // Just load and sign-extend the interesting 4 bytes instead. This happens, |
| // for example, when we're loading and untagging SMIs. |
| BaseWithIndexAndDisplacement64Matcher mleft(m.left().node(), |
| AddressOption::kAllowAll); |
| if (mleft.matches() && (mleft.displacement() == nullptr || |
| g.CanBeImmediate(mleft.displacement()))) { |
| #ifdef V8_IS_TSAN |
| // On TSAN builds we require one scratch register. Because of this we also |
| // have to modify the inputs to take into account possible aliasing and |
| // use UseUniqueRegister which is not required for non-TSAN builds. |
| InstructionOperand temps[] = {g.TempRegister()}; |
| size_t temp_count = arraysize(temps); |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseUniqueRegister; |
| #else |
| InstructionOperand* temps = nullptr; |
| size_t temp_count = 0; |
| auto reg_kind = OperandGenerator::RegisterUseKind::kUseRegister; |
| #endif // V8_IS_TSAN |
| size_t input_count = 0; |
| InstructionOperand inputs[3]; |
| AddressingMode mode = g.GetEffectiveAddressMemoryOperand( |
| m.left().node(), inputs, &input_count, reg_kind); |
| if (mleft.displacement() == nullptr) { |
| // Make sure that the addressing mode indicates the presence of an |
| // immediate displacement. It seems that we never use M1 and M2, but we |
| // handle them here anyways. |
| mode = AddDisplacementToAddressingMode(mode); |
| inputs[input_count++] = |
| ImmediateOperand(ImmediateOperand::INLINE_INT32, 4); |
| } else { |
| // In the case that the base address was zero, the displacement will be |
| // in a register and replacing it with an immediate is not allowed. This |
| // usually only happens in dead code anyway. |
| if (!inputs[input_count - 1].IsImmediate()) return false; |
| int32_t displacement = g.GetImmediateIntegerValue(mleft.displacement()); |
| inputs[input_count - 1] = |
| ImmediateOperand(ImmediateOperand::INLINE_INT32, displacement + 4); |
| } |
| InstructionOperand outputs[] = {g.DefineAsRegister(node)}; |
| InstructionCode code = opcode | AddressingModeField::encode(mode); |
| selector->Emit(code, 1, outputs, input_count, inputs, temp_count, temps); |
| return true; |
| } |
| } |
| return false; |
| } |
| |
| } // namespace |
| |
| void InstructionSelector::VisitWord64Shr(Node* node) { |
| if (TryMatchLoadWord64AndShiftRight(this, node, kX64Movl)) return; |
| VisitWord64Shift(this, node, kX64Shr); |
| } |
| |
| void InstructionSelector::VisitWord32Sar(Node* node) { |
| X64OperandGenerator g(this); |
| Int32BinopMatcher m(node); |
| if (CanCover(m.node(), m.left().node()) && m.left().IsWord32Shl()) { |
| Int32BinopMatcher mleft(m.left().node()); |
| if (mleft.right().Is(16) && m.right().Is(16)) { |
| Emit(kX64Movsxwl, g.DefineAsRegister(node), g.Use(mleft.left().node())); |
| return; |
| } else if (mleft.right().Is(24) && m.right().Is(24)) { |
| Emit(kX64Movsxbl, g.DefineAsRegister(node), g.Use(mleft.left().node())); |
| return; |
| } |
| } |
| VisitWord32Shift(this, node, kX64Sar32); |
| } |
| |
| void InstructionSelector::VisitWord64Sar(Node* node) { |
| if (TryMatchLoadWord64AndShiftRight(this, node, kX64Movsxlq)) return; |
| VisitWord64Shift(this, node, kX64Sar); |
| } |
| |
| void InstructionSelector::VisitWord32Rol(Node* node) { |
| VisitWord32Shift(this, node, kX64Rol32); |
| } |
| |
| void InstructionSelector::VisitWord64Rol(Node* node) { |
| VisitWord64Shift(this, node, kX64Rol); |
| } |
| |
| void InstructionSelector::VisitWord32Ror(Node* node) { |
| VisitWord32Shift(this, node, kX64Ror32); |
| } |
| |
| void InstructionSelector::VisitWord64Ror(Node* node) { |
| VisitWord64Shift(this, node, kX64Ror); |
| } |
| |
| void InstructionSelector::VisitWord32ReverseBits(Node* node) { UNREACHABLE(); } |
| |
| void InstructionSelector::VisitWord64ReverseBits(Node* node) { UNREACHABLE(); } |
| |
| void InstructionSelector::VisitWord64ReverseBytes(Node* node) { |
| X64OperandGenerator g(this); |
| Emit(kX64Bswap, g.DefineSameAsFirst(node), g.UseRegister(node->InputAt(0))); |
| } |
| |
| void InstructionSelector::VisitWord32ReverseBytes(Node* node) { |
| X64OperandGenerator g(this); |
| Emit(kX64Bswap32, g.DefineSameAsFirst(node), g.UseRegister(node->InputAt(0))); |
| } |
| |
| void InstructionSelector::VisitSimd128ReverseBytes(Node* node) { |
| UNREACHABLE(); |
| } |
| |
| void InstructionSelector::VisitInt32Add(Node* node) { |
| X64OperandGenerator g(this); |
| |
| // No need to truncate the values before Int32Add. |
| DCHECK_EQ(node->InputCount(), 2); |
| Node* left = node->InputAt(0); |
| Node* right = node->InputAt(1); |
| if (left->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| node->ReplaceInput(0, left->InputAt(0)); |
| } |
| if (right->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| node->ReplaceInput(1, right->InputAt(0)); |
| } |
| |
| // Try to match the Add to a leal pattern |
| BaseWithIndexAndDisplacement32Matcher m(node); |
| if (m.matches() && |
| (m.displacement() == nullptr || g.CanBeImmediate(m.displacement()))) { |
| EmitLea(this, kX64Lea32, node, m.index(), m.scale(), m.base(), |
| m.displacement(), m.displacement_mode()); |
| return; |
| } |
| |
| // No leal pattern match, use addl |
| VisitBinop(this, node, kX64Add32); |
| } |
| |
| void InstructionSelector::VisitInt64Add(Node* node) { |
| X64OperandGenerator g(this); |
| |
| // Try to match the Add to a leaq pattern |
| BaseWithIndexAndDisplacement64Matcher m(node); |
| if (m.matches() && |
| (m.displacement() == nullptr || g.CanBeImmediate(m.displacement()))) { |
| EmitLea(this, kX64Lea, node, m.index(), m.scale(), m.base(), |
| m.displacement(), m.displacement_mode()); |
| return; |
| } |
| |
| // No leal pattern match, use addq |
| VisitBinop(this, node, kX64Add); |
| } |
| |
| void InstructionSelector::VisitInt64AddWithOverflow(Node* node) { |
| if (Node* ovf = NodeProperties::FindProjection(node, 1)) { |
| FlagsContinuation cont = FlagsContinuation::ForSet(kOverflow, ovf); |
| return VisitBinop(this, node, kX64Add, &cont); |
| } |
| FlagsContinuation cont; |
| VisitBinop(this, node, kX64Add, &cont); |
| } |
| |
| void InstructionSelector::VisitInt32Sub(Node* node) { |
| X64OperandGenerator g(this); |
| DCHECK_EQ(node->InputCount(), 2); |
| Node* input1 = node->InputAt(0); |
| Node* input2 = node->InputAt(1); |
| if (input1->opcode() == IrOpcode::kTruncateInt64ToInt32 && |
| g.CanBeImmediate(input2)) { |
| int32_t imm = g.GetImmediateIntegerValue(input2); |
| InstructionOperand int64_input = g.UseRegister(input1->InputAt(0)); |
| if (imm == 0) { |
| // Emit "movl" for subtraction of 0. |
| Emit(kX64Movl, g.DefineAsRegister(node), int64_input); |
| } else { |
| // Omit truncation and turn subtractions of constant values into immediate |
| // "leal" instructions by negating the value. |
| Emit(kX64Lea32 | AddressingModeField::encode(kMode_MRI), |
| g.DefineAsRegister(node), int64_input, |
| g.TempImmediate(base::NegateWithWraparound(imm))); |
| } |
| return; |
| } |
| |
| Int32BinopMatcher m(node); |
| if (m.left().Is(0)) { |
| Emit(kX64Neg32, g.DefineSameAsFirst(node), g.UseRegister(m.right().node())); |
| } else if (m.right().Is(0)) { |
| // {EmitIdentity} reuses the virtual register of the first input |
| // for the output. This is exactly what we want here. |
| EmitIdentity(node); |
| } else if (m.right().HasResolvedValue() && |
| g.CanBeImmediate(m.right().node())) { |
| // Turn subtractions of constant values into immediate "leal" instructions |
| // by negating the value. |
| Emit( |
| kX64Lea32 | AddressingModeField::encode(kMode_MRI), |
| g.DefineAsRegister(node), g.UseRegister(m.left().node()), |
| g.TempImmediate(base::NegateWithWraparound(m.right().ResolvedValue()))); |
| } else { |
| VisitBinop(this, node, kX64Sub32); |
| } |
| } |
| |
| void InstructionSelector::VisitInt64Sub(Node* node) { |
| X64OperandGenerator g(this); |
| Int64BinopMatcher m(node); |
| if (m.left().Is(0)) { |
| Emit(kX64Neg, g.DefineSameAsFirst(node), g.UseRegister(m.right().node())); |
| } else { |
| if (m.right().HasResolvedValue() && g.CanBeImmediate(m.right().node())) { |
| // Turn subtractions of constant values into immediate "leaq" instructions |
| // by negating the value. |
| Emit(kX64Lea | AddressingModeField::encode(kMode_MRI), |
| g.DefineAsRegister(node), g.UseRegister(m.left().node()), |
| g.TempImmediate(-static_cast<int32_t>(m.right().ResolvedValue()))); |
| return; |
| } |
| VisitBinop(this, node, kX64Sub); |
| } |
| } |
| |
| void InstructionSelector::VisitInt64SubWithOverflow(Node* node) { |
| if (Node* ovf = NodeProperties::FindProjection(node, 1)) { |
| FlagsContinuation cont = FlagsContinuation::ForSet(kOverflow, ovf); |
| return VisitBinop(this, node, kX64Sub, &cont); |
| } |
| FlagsContinuation cont; |
| VisitBinop(this, node, kX64Sub, &cont); |
| } |
| |
| namespace { |
| |
| void VisitMul(InstructionSelector* selector, Node* node, ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| Int32BinopMatcher m(node); |
| Node* left = m.left().node(); |
| Node* right = m.right().node(); |
| if (g.CanBeImmediate(right)) { |
| selector->Emit(opcode, g.DefineAsRegister(node), g.Use(left), |
| g.UseImmediate(right)); |
| } else { |
| if (g.CanBeBetterLeftOperand(right)) { |
| std::swap(left, right); |
| } |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(left), |
| g.Use(right)); |
| } |
| } |
| |
| void VisitMulHigh(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| Node* left = node->InputAt(0); |
| Node* right = node->InputAt(1); |
| if (selector->IsLive(left) && !selector->IsLive(right)) { |
| std::swap(left, right); |
| } |
| InstructionOperand temps[] = {g.TempRegister(rax)}; |
| // TODO(turbofan): We use UseUniqueRegister here to improve register |
| // allocation. |
| selector->Emit(opcode, g.DefineAsFixed(node, rdx), g.UseFixed(left, rax), |
| g.UseUniqueRegister(right), arraysize(temps), temps); |
| } |
| |
| void VisitDiv(InstructionSelector* selector, Node* node, ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| InstructionOperand temps[] = {g.TempRegister(rdx)}; |
| selector->Emit( |
| opcode, g.DefineAsFixed(node, rax), g.UseFixed(node->InputAt(0), rax), |
| g.UseUniqueRegister(node->InputAt(1)), arraysize(temps), temps); |
| } |
| |
| void VisitMod(InstructionSelector* selector, Node* node, ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| InstructionOperand temps[] = {g.TempRegister(rax)}; |
| selector->Emit( |
| opcode, g.DefineAsFixed(node, rdx), g.UseFixed(node->InputAt(0), rax), |
| g.UseUniqueRegister(node->InputAt(1)), arraysize(temps), temps); |
| } |
| |
| } // namespace |
| |
| void InstructionSelector::VisitInt32Mul(Node* node) { |
| Int32ScaleMatcher m(node, true); |
| if (m.matches()) { |
| Node* index = node->InputAt(0); |
| Node* base = m.power_of_two_plus_one() ? index : nullptr; |
| EmitLea(this, kX64Lea32, node, index, m.scale(), base, nullptr, |
| kPositiveDisplacement); |
| return; |
| } |
| VisitMul(this, node, kX64Imul32); |
| } |
| |
| void InstructionSelector::VisitInt32MulWithOverflow(Node* node) { |
| // TODO(mvstanton): Use Int32ScaleMatcher somehow. |
| if (Node* ovf = NodeProperties::FindProjection(node, 1)) { |
| FlagsContinuation cont = FlagsContinuation::ForSet(kOverflow, ovf); |
| return VisitBinop(this, node, kX64Imul32, &cont); |
| } |
| FlagsContinuation cont; |
| VisitBinop(this, node, kX64Imul32, &cont); |
| } |
| |
| void InstructionSelector::VisitInt64Mul(Node* node) { |
| VisitMul(this, node, kX64Imul); |
| } |
| |
| void InstructionSelector::VisitInt32MulHigh(Node* node) { |
| VisitMulHigh(this, node, kX64ImulHigh32); |
| } |
| |
| void InstructionSelector::VisitInt32Div(Node* node) { |
| VisitDiv(this, node, kX64Idiv32); |
| } |
| |
| void InstructionSelector::VisitInt64Div(Node* node) { |
| VisitDiv(this, node, kX64Idiv); |
| } |
| |
| void InstructionSelector::VisitUint32Div(Node* node) { |
| VisitDiv(this, node, kX64Udiv32); |
| } |
| |
| void InstructionSelector::VisitUint64Div(Node* node) { |
| VisitDiv(this, node, kX64Udiv); |
| } |
| |
| void InstructionSelector::VisitInt32Mod(Node* node) { |
| VisitMod(this, node, kX64Idiv32); |
| } |
| |
| void InstructionSelector::VisitInt64Mod(Node* node) { |
| VisitMod(this, node, kX64Idiv); |
| } |
| |
| void InstructionSelector::VisitUint32Mod(Node* node) { |
| VisitMod(this, node, kX64Udiv32); |
| } |
| |
| void InstructionSelector::VisitUint64Mod(Node* node) { |
| VisitMod(this, node, kX64Udiv); |
| } |
| |
| void InstructionSelector::VisitUint32MulHigh(Node* node) { |
| VisitMulHigh(this, node, kX64UmulHigh32); |
| } |
| |
| void InstructionSelector::VisitTryTruncateFloat32ToInt64(Node* node) { |
| X64OperandGenerator g(this); |
| InstructionOperand inputs[] = {g.UseRegister(node->InputAt(0))}; |
| InstructionOperand outputs[2]; |
| size_t output_count = 0; |
| outputs[output_count++] = g.DefineAsRegister(node); |
| |
| Node* success_output = NodeProperties::FindProjection(node, 1); |
| if (success_output) { |
| outputs[output_count++] = g.DefineAsRegister(success_output); |
| } |
| |
| Emit(kSSEFloat32ToInt64, output_count, outputs, 1, inputs); |
| } |
| |
| void InstructionSelector::VisitTryTruncateFloat64ToInt64(Node* node) { |
| X64OperandGenerator g(this); |
| InstructionOperand inputs[] = {g.UseRegister(node->InputAt(0))}; |
| InstructionOperand outputs[2]; |
| size_t output_count = 0; |
| outputs[output_count++] = g.DefineAsRegister(node); |
| |
| Node* success_output = NodeProperties::FindProjection(node, 1); |
| if (success_output) { |
| outputs[output_count++] = g.DefineAsRegister(success_output); |
| } |
| |
| Emit(kSSEFloat64ToInt64, output_count, outputs, 1, inputs); |
| } |
| |
| void InstructionSelector::VisitTryTruncateFloat32ToUint64(Node* node) { |
| X64OperandGenerator g(this); |
| InstructionOperand inputs[] = {g.UseRegister(node->InputAt(0))}; |
| InstructionOperand outputs[2]; |
| size_t output_count = 0; |
| outputs[output_count++] = g.DefineAsRegister(node); |
| |
| Node* success_output = NodeProperties::FindProjection(node, 1); |
| if (success_output) { |
| outputs[output_count++] = g.DefineAsRegister(success_output); |
| } |
| |
| Emit(kSSEFloat32ToUint64, output_count, outputs, 1, inputs); |
| } |
| |
| void InstructionSelector::VisitTryTruncateFloat64ToUint64(Node* node) { |
| X64OperandGenerator g(this); |
| InstructionOperand inputs[] = {g.UseRegister(node->InputAt(0))}; |
| InstructionOperand outputs[2]; |
| size_t output_count = 0; |
| outputs[output_count++] = g.DefineAsRegister(node); |
| |
| Node* success_output = NodeProperties::FindProjection(node, 1); |
| if (success_output) { |
| outputs[output_count++] = g.DefineAsRegister(success_output); |
| } |
| |
| Emit(kSSEFloat64ToUint64, output_count, outputs, 1, inputs); |
| } |
| |
| void InstructionSelector::VisitBitcastWord32ToWord64(Node* node) { |
| DCHECK(SmiValuesAre31Bits()); |
| DCHECK(COMPRESS_POINTERS_BOOL); |
| EmitIdentity(node); |
| } |
| |
| void InstructionSelector::VisitChangeInt32ToInt64(Node* node) { |
| DCHECK_EQ(node->InputCount(), 1); |
| Node* input = node->InputAt(0); |
| if (input->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| node->ReplaceInput(0, input->InputAt(0)); |
| } |
| |
| X64OperandGenerator g(this); |
| Node* const value = node->InputAt(0); |
| if ((value->opcode() == IrOpcode::kLoad || |
| value->opcode() == IrOpcode::kLoadImmutable) && |
| CanCover(node, value)) { |
| LoadRepresentation load_rep = LoadRepresentationOf(value->op()); |
| MachineRepresentation rep = load_rep.representation(); |
| InstructionCode opcode; |
| switch (rep) { |
| case MachineRepresentation::kBit: // Fall through. |
| case MachineRepresentation::kWord8: |
| opcode = load_rep.IsSigned() ? kX64Movsxbq : kX64Movzxbq; |
| break; |
| case MachineRepresentation::kWord16: |
| opcode = load_rep.IsSigned() ? kX64Movsxwq : kX64Movzxwq; |
| break; |
| case MachineRepresentation::kWord32: |
| // ChangeInt32ToInt64 must interpret its input as a _signed_ 32-bit |
| // integer, so here we must sign-extend the loaded value in any case. |
| opcode = kX64Movsxlq; |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| InstructionOperand outputs[] = {g.DefineAsRegister(node)}; |
| size_t input_count = 0; |
| InstructionOperand inputs[3]; |
| AddressingMode mode = g.GetEffectiveAddressMemoryOperand( |
| node->InputAt(0), inputs, &input_count); |
| opcode |= AddressingModeField::encode(mode); |
| Emit(opcode, 1, outputs, input_count, inputs); |
| } else { |
| Emit(kX64Movsxlq, g.DefineAsRegister(node), g.Use(node->InputAt(0))); |
| } |
| } |
| |
| bool InstructionSelector::ZeroExtendsWord32ToWord64NoPhis(Node* node) { |
| X64OperandGenerator g(this); |
| DCHECK_NE(node->opcode(), IrOpcode::kPhi); |
| switch (node->opcode()) { |
| case IrOpcode::kWord32And: |
| case IrOpcode::kWord32Or: |
| case IrOpcode::kWord32Xor: |
| case IrOpcode::kWord32Shl: |
| case IrOpcode::kWord32Shr: |
| case IrOpcode::kWord32Sar: |
| case IrOpcode::kWord32Rol: |
| case IrOpcode::kWord32Ror: |
| case IrOpcode::kWord32Equal: |
| case IrOpcode::kInt32Add: |
| case IrOpcode::kInt32Sub: |
| case IrOpcode::kInt32Mul: |
| case IrOpcode::kInt32MulHigh: |
| case IrOpcode::kInt32Div: |
| case IrOpcode::kInt32LessThan: |
| case IrOpcode::kInt32LessThanOrEqual: |
| case IrOpcode::kInt32Mod: |
| case IrOpcode::kUint32Div: |
| case IrOpcode::kUint32LessThan: |
| case IrOpcode::kUint32LessThanOrEqual: |
| case IrOpcode::kUint32Mod: |
| case IrOpcode::kUint32MulHigh: |
| case IrOpcode::kTruncateInt64ToInt32: |
| // These 32-bit operations implicitly zero-extend to 64-bit on x64, so the |
| // zero-extension is a no-op. |
| return true; |
| case IrOpcode::kProjection: { |
| Node* const value = node->InputAt(0); |
| switch (value->opcode()) { |
| case IrOpcode::kInt32AddWithOverflow: |
| case IrOpcode::kInt32SubWithOverflow: |
| case IrOpcode::kInt32MulWithOverflow: |
| return true; |
| default: |
| return false; |
| } |
| } |
| case IrOpcode::kLoad: |
| case IrOpcode::kLoadImmutable: |
| case IrOpcode::kProtectedLoad: { |
| // The movzxbl/movsxbl/movzxwl/movsxwl/movl operations implicitly |
| // zero-extend to 64-bit on x64, so the zero-extension is a no-op. |
| LoadRepresentation load_rep = LoadRepresentationOf(node->op()); |
| switch (load_rep.representation()) { |
| case MachineRepresentation::kWord8: |
| case MachineRepresentation::kWord16: |
| case MachineRepresentation::kWord32: |
| return true; |
| default: |
| return false; |
| } |
| } |
| case IrOpcode::kInt32Constant: |
| case IrOpcode::kInt64Constant: |
| // Constants are loaded with movl or movq, or xorl for zero; see |
| // CodeGenerator::AssembleMove. So any non-negative constant that fits |
| // in a 32-bit signed integer is zero-extended to 64 bits. |
| if (g.CanBeImmediate(node)) { |
| return g.GetImmediateIntegerValue(node) >= 0; |
| } |
| return false; |
| default: |
| return false; |
| } |
| } |
| |
| void InstructionSelector::VisitChangeUint32ToUint64(Node* node) { |
| X64OperandGenerator g(this); |
| Node* value = node->InputAt(0); |
| if (ZeroExtendsWord32ToWord64(value)) { |
| // These 32-bit operations implicitly zero-extend to 64-bit on x64, so the |
| // zero-extension is a no-op. |
| return EmitIdentity(node); |
| } |
| Emit(kX64Movl, g.DefineAsRegister(node), g.Use(value)); |
| } |
| |
| namespace { |
| |
| void VisitRO(InstructionSelector* selector, Node* node, |
| InstructionCode opcode) { |
| X64OperandGenerator g(selector); |
| selector->Emit(opcode, g.DefineAsRegister(node), g.Use(node->InputAt(0))); |
| } |
| |
| void VisitRR(InstructionSelector* selector, Node* node, |
| InstructionCode opcode) { |
| X64OperandGenerator g(selector); |
| selector->Emit(opcode, g.DefineAsRegister(node), |
| g.UseRegister(node->InputAt(0))); |
| } |
| |
| void VisitRRO(InstructionSelector* selector, Node* node, |
| InstructionCode opcode) { |
| X64OperandGenerator g(selector); |
| selector->Emit(opcode, g.DefineSameAsFirst(node), |
| g.UseRegister(node->InputAt(0)), g.Use(node->InputAt(1))); |
| } |
| |
| void VisitFloatBinop(InstructionSelector* selector, Node* node, |
| InstructionCode avx_opcode, InstructionCode sse_opcode) { |
| X64OperandGenerator g(selector); |
| Node* left = node->InputAt(0); |
| Node* right = node->InputAt(1); |
| InstructionOperand inputs[8]; |
| size_t input_count = 0; |
| InstructionOperand outputs[1]; |
| size_t output_count = 0; |
| |
| if (left == right) { |
| // If both inputs refer to the same operand, enforce allocating a register |
| // for both of them to ensure that we don't end up generating code like |
| // this: |
| // |
| // movss rax, [rbp-0x10] |
| // addss rax, [rbp-0x10] |
| // jo label |
| InstructionOperand const input = g.UseRegister(left); |
| inputs[input_count++] = input; |
| inputs[input_count++] = input; |
| } else { |
| int effect_level = selector->GetEffectLevel(node); |
| if (node->op()->HasProperty(Operator::kCommutative) && |
| (g.CanBeBetterLeftOperand(right) || |
| g.CanBeMemoryOperand(avx_opcode, node, left, effect_level)) && |
| (!g.CanBeBetterLeftOperand(left) || |
| !g.CanBeMemoryOperand(avx_opcode, node, right, effect_level))) { |
| std::swap(left, right); |
| } |
| if (g.CanBeMemoryOperand(avx_opcode, node, right, effect_level)) { |
| inputs[input_count++] = g.UseRegister(left); |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(right, inputs, &input_count); |
| avx_opcode |= AddressingModeField::encode(addressing_mode); |
| sse_opcode |= AddressingModeField::encode(addressing_mode); |
| } else { |
| inputs[input_count++] = g.UseRegister(left); |
| inputs[input_count++] = g.Use(right); |
| } |
| } |
| |
| DCHECK_NE(0u, input_count); |
| DCHECK_GE(arraysize(inputs), input_count); |
| |
| if (selector->IsSupported(AVX)) { |
| outputs[output_count++] = g.DefineAsRegister(node); |
| DCHECK_EQ(1u, output_count); |
| DCHECK_GE(arraysize(outputs), output_count); |
| selector->Emit(avx_opcode, output_count, outputs, input_count, inputs); |
| } else { |
| outputs[output_count++] = g.DefineSameAsFirst(node); |
| DCHECK_EQ(1u, output_count); |
| DCHECK_GE(arraysize(outputs), output_count); |
| selector->Emit(sse_opcode, output_count, outputs, input_count, inputs); |
| } |
| } |
| |
| void VisitFloatUnop(InstructionSelector* selector, Node* node, Node* input, |
| ArchOpcode opcode) { |
| X64OperandGenerator g(selector); |
| if (selector->IsSupported(AVX)) { |
| selector->Emit(opcode, g.DefineAsRegister(node), g.UseRegister(input)); |
| } else { |
| selector->Emit(opcode, g.DefineSameAsFirst(node), g.UseRegister(input)); |
| } |
| } |
| |
| } // namespace |
| |
| #define RO_OP_LIST(V) \ |
| V(Word64Clz, kX64Lzcnt) \ |
| V(Word32Clz, kX64Lzcnt32) \ |
| V(Word64Ctz, kX64Tzcnt) \ |
| V(Word32Ctz, kX64Tzcnt32) \ |
| V(Word64Popcnt, kX64Popcnt) \ |
| V(Word32Popcnt, kX64Popcnt32) \ |
| V(Float64Sqrt, kSSEFloat64Sqrt) \ |
| V(Float32Sqrt, kSSEFloat32Sqrt) \ |
| V(ChangeFloat64ToInt32, kSSEFloat64ToInt32) \ |
| V(ChangeFloat64ToInt64, kSSEFloat64ToInt64) \ |
| V(ChangeFloat64ToUint32, kSSEFloat64ToUint32 | MiscField::encode(1)) \ |
| V(TruncateFloat64ToInt64, kSSEFloat64ToInt64) \ |
| V(TruncateFloat64ToUint32, kSSEFloat64ToUint32 | MiscField::encode(0)) \ |
| V(ChangeFloat64ToUint64, kSSEFloat64ToUint64) \ |
| V(TruncateFloat64ToFloat32, kSSEFloat64ToFloat32) \ |
| V(ChangeFloat32ToFloat64, kSSEFloat32ToFloat64) \ |
| V(TruncateFloat32ToInt32, kSSEFloat32ToInt32) \ |
| V(TruncateFloat32ToUint32, kSSEFloat32ToUint32) \ |
| V(ChangeInt32ToFloat64, kSSEInt32ToFloat64) \ |
| V(ChangeInt64ToFloat64, kSSEInt64ToFloat64) \ |
| V(ChangeUint32ToFloat64, kSSEUint32ToFloat64) \ |
| V(RoundFloat64ToInt32, kSSEFloat64ToInt32) \ |
| V(RoundInt32ToFloat32, kSSEInt32ToFloat32) \ |
| V(RoundInt64ToFloat32, kSSEInt64ToFloat32) \ |
| V(RoundUint64ToFloat32, kSSEUint64ToFloat32) \ |
| V(RoundInt64ToFloat64, kSSEInt64ToFloat64) \ |
| V(RoundUint64ToFloat64, kSSEUint64ToFloat64) \ |
| V(RoundUint32ToFloat32, kSSEUint32ToFloat32) \ |
| V(BitcastFloat32ToInt32, kX64BitcastFI) \ |
| V(BitcastFloat64ToInt64, kX64BitcastDL) \ |
| V(BitcastInt32ToFloat32, kX64BitcastIF) \ |
| V(BitcastInt64ToFloat64, kX64BitcastLD) \ |
| V(Float64ExtractLowWord32, kSSEFloat64ExtractLowWord32) \ |
| V(Float64ExtractHighWord32, kSSEFloat64ExtractHighWord32) \ |
| V(SignExtendWord8ToInt32, kX64Movsxbl) \ |
| V(SignExtendWord16ToInt32, kX64Movsxwl) \ |
| V(SignExtendWord8ToInt64, kX64Movsxbq) \ |
| V(SignExtendWord16ToInt64, kX64Movsxwq) \ |
| V(SignExtendWord32ToInt64, kX64Movsxlq) |
| |
| #define RR_OP_LIST(V) \ |
| V(Float32RoundDown, kSSEFloat32Round | MiscField::encode(kRoundDown)) \ |
| V(Float64RoundDown, kSSEFloat64Round | MiscField::encode(kRoundDown)) \ |
| V(Float32RoundUp, kSSEFloat32Round | MiscField::encode(kRoundUp)) \ |
| V(Float64RoundUp, kSSEFloat64Round | MiscField::encode(kRoundUp)) \ |
| V(Float32RoundTruncate, kSSEFloat32Round | MiscField::encode(kRoundToZero)) \ |
| V(Float64RoundTruncate, kSSEFloat64Round | MiscField::encode(kRoundToZero)) \ |
| V(Float32RoundTiesEven, \ |
| kSSEFloat32Round | MiscField::encode(kRoundToNearest)) \ |
| V(Float64RoundTiesEven, \ |
| kSSEFloat64Round | MiscField::encode(kRoundToNearest)) \ |
| V(F32x4Ceil, kX64F32x4Round | MiscField::encode(kRoundUp)) \ |
| V(F32x4Floor, kX64F32x4Round | MiscField::encode(kRoundDown)) \ |
| V(F32x4Trunc, kX64F32x4Round | MiscField::encode(kRoundToZero)) \ |
| V(F32x4NearestInt, kX64F32x4Round | MiscField::encode(kRoundToNearest)) \ |
| V(F64x2Ceil, kX64F64x2Round | MiscField::encode(kRoundUp)) \ |
| V(F64x2Floor, kX64F64x2Round | MiscField::encode(kRoundDown)) \ |
| V(F64x2Trunc, kX64F64x2Round | MiscField::encode(kRoundToZero)) \ |
| V(F64x2NearestInt, kX64F64x2Round | MiscField::encode(kRoundToNearest)) |
| |
| #define RO_VISITOR(Name, opcode) \ |
| void InstructionSelector::Visit##Name(Node* node) { \ |
| VisitRO(this, node, opcode); \ |
| } |
| RO_OP_LIST(RO_VISITOR) |
| #undef RO_VISITOR |
| #undef RO_OP_LIST |
| |
| #define RR_VISITOR(Name, opcode) \ |
| void InstructionSelector::Visit##Name(Node* node) { \ |
| VisitRR(this, node, opcode); \ |
| } |
| RR_OP_LIST(RR_VISITOR) |
| #undef RR_VISITOR |
| #undef RR_OP_LIST |
| |
| void InstructionSelector::VisitTruncateFloat64ToWord32(Node* node) { |
| VisitRR(this, node, kArchTruncateDoubleToI); |
| } |
| |
| void InstructionSelector::VisitTruncateInt64ToInt32(Node* node) { |
| // We rely on the fact that TruncateInt64ToInt32 zero extends the |
| // value (see ZeroExtendsWord32ToWord64). So all code paths here |
| // have to satisfy that condition. |
| X64OperandGenerator g(this); |
| Node* value = node->InputAt(0); |
| if (CanCover(node, value)) { |
| switch (value->opcode()) { |
| case IrOpcode::kWord64Sar: |
| case IrOpcode::kWord64Shr: { |
| Int64BinopMatcher m(value); |
| if (m.right().Is(32)) { |
| if (CanCoverTransitively(node, value, value->InputAt(0)) && |
| TryMatchLoadWord64AndShiftRight(this, value, kX64Movl)) { |
| return EmitIdentity(node); |
| } |
| Emit(kX64Shr, g.DefineSameAsFirst(node), |
| g.UseRegister(m.left().node()), g.TempImmediate(32)); |
| return; |
| } |
| break; |
| } |
| case IrOpcode::kLoad: |
| case IrOpcode::kLoadImmutable: { |
| if (TryMergeTruncateInt64ToInt32IntoLoad(this, node, value)) { |
| return; |
| } |
| break; |
| } |
| default: |
| break; |
| } |
| } |
| Emit(kX64Movl, g.DefineAsRegister(node), g.Use(value)); |
| } |
| |
| void InstructionSelector::VisitFloat32Add(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat32Add, kSSEFloat32Add); |
| } |
| |
| void InstructionSelector::VisitFloat32Sub(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat32Sub, kSSEFloat32Sub); |
| } |
| |
| void InstructionSelector::VisitFloat32Mul(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat32Mul, kSSEFloat32Mul); |
| } |
| |
| void InstructionSelector::VisitFloat32Div(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat32Div, kSSEFloat32Div); |
| } |
| |
| void InstructionSelector::VisitFloat32Abs(Node* node) { |
| VisitFloatUnop(this, node, node->InputAt(0), kX64Float32Abs); |
| } |
| |
| void InstructionSelector::VisitFloat32Max(Node* node) { |
| VisitRRO(this, node, kSSEFloat32Max); |
| } |
| |
| void InstructionSelector::VisitFloat32Min(Node* node) { |
| VisitRRO(this, node, kSSEFloat32Min); |
| } |
| |
| void InstructionSelector::VisitFloat64Add(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat64Add, kSSEFloat64Add); |
| } |
| |
| void InstructionSelector::VisitFloat64Sub(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat64Sub, kSSEFloat64Sub); |
| } |
| |
| void InstructionSelector::VisitFloat64Mul(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat64Mul, kSSEFloat64Mul); |
| } |
| |
| void InstructionSelector::VisitFloat64Div(Node* node) { |
| VisitFloatBinop(this, node, kAVXFloat64Div, kSSEFloat64Div); |
| } |
| |
| void InstructionSelector::VisitFloat64Mod(Node* node) { |
| X64OperandGenerator g(this); |
| InstructionOperand temps[] = {g.TempRegister(rax)}; |
| Emit(kSSEFloat64Mod, g.DefineSameAsFirst(node), |
| g.UseRegister(node->InputAt(0)), g.UseRegister(node->InputAt(1)), 1, |
| temps); |
| } |
| |
| void InstructionSelector::VisitFloat64Max(Node* node) { |
| VisitRRO(this, node, kSSEFloat64Max); |
| } |
| |
| void InstructionSelector::VisitFloat64Min(Node* node) { |
| VisitRRO(this, node, kSSEFloat64Min); |
| } |
| |
| void InstructionSelector::VisitFloat64Abs(Node* node) { |
| VisitFloatUnop(this, node, node->InputAt(0), kX64Float64Abs); |
| } |
| |
| void InstructionSelector::VisitFloat64RoundTiesAway(Node* node) { |
| UNREACHABLE(); |
| } |
| |
| void InstructionSelector::VisitFloat32Neg(Node* node) { |
| VisitFloatUnop(this, node, node->InputAt(0), kX64Float32Neg); |
| } |
| |
| void InstructionSelector::VisitFloat64Neg(Node* node) { |
| VisitFloatUnop(this, node, node->InputAt(0), kX64Float64Neg); |
| } |
| |
| void InstructionSelector::VisitFloat64Ieee754Binop(Node* node, |
| InstructionCode opcode) { |
| X64OperandGenerator g(this); |
| Emit(opcode, g.DefineAsFixed(node, xmm0), g.UseFixed(node->InputAt(0), xmm0), |
| g.UseFixed(node->InputAt(1), xmm1)) |
| ->MarkAsCall(); |
| } |
| |
| void InstructionSelector::VisitFloat64Ieee754Unop(Node* node, |
| InstructionCode opcode) { |
| X64OperandGenerator g(this); |
| Emit(opcode, g.DefineAsFixed(node, xmm0), g.UseFixed(node->InputAt(0), xmm0)) |
| ->MarkAsCall(); |
| } |
| |
| void InstructionSelector::EmitPrepareArguments( |
| ZoneVector<PushParameter>* arguments, const CallDescriptor* call_descriptor, |
| Node* node) { |
| X64OperandGenerator g(this); |
| |
| // Prepare for C function call. |
| if (call_descriptor->IsCFunctionCall()) { |
| Emit(kArchPrepareCallCFunction | MiscField::encode(static_cast<int>( |
| call_descriptor->ParameterCount())), |
| 0, nullptr, 0, nullptr); |
| |
| // Poke any stack arguments. |
| for (size_t n = 0; n < arguments->size(); ++n) { |
| PushParameter input = (*arguments)[n]; |
| if (input.node) { |
| int slot = static_cast<int>(n); |
| InstructionOperand value = g.CanBeImmediate(input.node) |
| ? g.UseImmediate(input.node) |
| : g.UseRegister(input.node); |
| Emit(kX64Poke | MiscField::encode(slot), g.NoOutput(), value); |
| } |
| } |
| } else { |
| // Push any stack arguments. |
| int effect_level = GetEffectLevel(node); |
| int stack_decrement = 0; |
| for (PushParameter input : base::Reversed(*arguments)) { |
| stack_decrement += kSystemPointerSize; |
| // Skip holes in the param array. These represent both extra slots for |
| // multi-slot values and padding slots for alignment. |
| if (input.node == nullptr) continue; |
| InstructionOperand decrement = g.UseImmediate(stack_decrement); |
| stack_decrement = 0; |
| if (g.CanBeImmediate(input.node)) { |
| Emit(kX64Push, g.NoOutput(), decrement, g.UseImmediate(input.node)); |
| } else if (IsSupported(INTEL_ATOM) || |
| sequence()->IsFP(GetVirtualRegister(input.node))) { |
| // TODO(titzer): X64Push cannot handle stack->stack double moves |
| // because there is no way to encode fixed double slots. |
| Emit(kX64Push, g.NoOutput(), decrement, g.UseRegister(input.node)); |
| } else if (g.CanBeMemoryOperand(kX64Push, node, input.node, |
| effect_level)) { |
| InstructionOperand outputs[1]; |
| InstructionOperand inputs[5]; |
| size_t input_count = 0; |
| inputs[input_count++] = decrement; |
| AddressingMode mode = g.GetEffectiveAddressMemoryOperand( |
| input.node, inputs, &input_count); |
| InstructionCode opcode = kX64Push | AddressingModeField::encode(mode); |
| Emit(opcode, 0, outputs, input_count, inputs); |
| } else { |
| Emit(kX64Push, g.NoOutput(), decrement, g.UseAny(input.node)); |
| } |
| } |
| } |
| } |
| |
| void InstructionSelector::EmitPrepareResults( |
| ZoneVector<PushParameter>* results, const CallDescriptor* call_descriptor, |
| Node* node) { |
| X64OperandGenerator g(this); |
| for (PushParameter output : *results) { |
| if (!output.location.IsCallerFrameSlot()) continue; |
| // Skip any alignment holes in nodes. |
| if (output.node != nullptr) { |
| DCHECK(!call_descriptor->IsCFunctionCall()); |
| if (output.location.GetType() == MachineType::Float32()) { |
| MarkAsFloat32(output.node); |
| } else if (output.location.GetType() == MachineType::Float64()) { |
| MarkAsFloat64(output.node); |
| } else if (output.location.GetType() == MachineType::Simd128()) { |
| MarkAsSimd128(output.node); |
| } |
| InstructionOperand result = g.DefineAsRegister(output.node); |
| int offset = call_descriptor->GetOffsetToReturns(); |
| int reverse_slot = -output.location.GetLocation() - offset; |
| InstructionOperand slot = g.UseImmediate(reverse_slot); |
| Emit(kX64Peek, 1, &result, 1, &slot); |
| } |
| } |
| } |
| |
| bool InstructionSelector::IsTailCallAddressImmediate() { return true; } |
| |
| namespace { |
| |
| void VisitCompareWithMemoryOperand(InstructionSelector* selector, |
| InstructionCode opcode, Node* left, |
| InstructionOperand right, |
| FlagsContinuation* cont) { |
| DCHECK(IrOpcode::kLoad == left->opcode() || |
| IrOpcode::kLoadImmutable == left->opcode()); |
| X64OperandGenerator g(selector); |
| size_t input_count = 0; |
| InstructionOperand inputs[6]; |
| AddressingMode addressing_mode = |
| g.GetEffectiveAddressMemoryOperand(left, inputs, &input_count); |
| opcode |= AddressingModeField::encode(addressing_mode); |
| inputs[input_count++] = right; |
| if (cont->IsSelect()) { |
| if (opcode == kUnorderedEqual) { |
| cont->Negate(); |
| inputs[input_count++] = g.UseRegister(cont->true_value()); |
| inputs[input_count++] = g.Use(cont->false_value()); |
| } else { |
| inputs[input_count++] = g.UseRegister(cont->false_value()); |
| inputs[input_count++] = g.Use(cont->true_value()); |
| } |
| } |
| |
| selector->EmitWithContinuation(opcode, 0, nullptr, input_count, inputs, cont); |
| } |
| |
| // Shared routine for multiple compare operations. |
| void VisitCompare(InstructionSelector* selector, InstructionCode opcode, |
| InstructionOperand left, InstructionOperand right, |
| FlagsContinuation* cont) { |
| if (cont->IsSelect()) { |
| X64OperandGenerator g(selector); |
| InstructionOperand inputs[4] = {left, right}; |
| if (cont->condition() == kUnorderedEqual) { |
| cont->Negate(); |
| inputs[2] = g.UseRegister(cont->true_value()); |
| inputs[3] = g.Use(cont->false_value()); |
| } else { |
| inputs[2] = g.UseRegister(cont->false_value()); |
| inputs[3] = g.Use(cont->true_value()); |
| } |
| selector->EmitWithContinuation(opcode, 0, nullptr, 4, inputs, cont); |
| return; |
| } |
| selector->EmitWithContinuation(opcode, left, right, cont); |
| } |
| |
| // Shared routine for multiple compare operations. |
| void VisitCompare(InstructionSelector* selector, InstructionCode opcode, |
| Node* left, Node* right, FlagsContinuation* cont, |
| bool commutative) { |
| X64OperandGenerator g(selector); |
| if (commutative && g.CanBeBetterLeftOperand(right)) { |
| std::swap(left, right); |
| } |
| VisitCompare(selector, opcode, g.UseRegister(left), g.Use(right), cont); |
| } |
| |
| MachineType MachineTypeForNarrow(Node* node, Node* hint_node) { |
| if (hint_node->opcode() == IrOpcode::kLoad || |
| hint_node->opcode() == IrOpcode::kLoadImmutable) { |
| MachineType hint = LoadRepresentationOf(hint_node->op()); |
| if (node->opcode() == IrOpcode::kInt32Constant || |
| node->opcode() == IrOpcode::kInt64Constant) { |
| int64_t constant = node->opcode() == IrOpcode::kInt32Constant |
| ? OpParameter<int32_t>(node->op()) |
| : OpParameter<int64_t>(node->op()); |
| if (hint == MachineType::Int8()) { |
| if (constant >= std::numeric_limits<int8_t>::min() && |
| constant <= std::numeric_limits<int8_t>::max()) { |
| return hint; |
| } |
| } else if (hint == MachineType::Uint8()) { |
| if (constant >= std::numeric_limits<uint8_t>::min() && |
| constant <= std::numeric_limits<uint8_t>::max()) { |
| return hint; |
| } |
| } else if (hint == MachineType::Int16()) { |
| if (constant >= std::numeric_limits<int16_t>::min() && |
| constant <= std::numeric_limits<int16_t>::max()) { |
| return hint; |
| } |
| } else if (hint == MachineType::Uint16()) { |
| if (constant >= std::numeric_limits<uint16_t>::min() && |
| constant <= std::numeric_limits<uint16_t>::max()) { |
| return hint; |
| } |
| } else if (hint == MachineType::Int32()) { |
| if (constant >= std::numeric_limits<int32_t>::min() && |
| constant <= std::numeric_limits<int32_t>::max()) { |
| return hint; |
| } |
| } else if (hint == MachineType::Uint32()) { |
| if (constant >= std::numeric_limits<uint32_t>::min() && |
| constant <= std::numeric_limits<uint32_t>::max()) |
| return hint; |
| } |
| } |
| } |
| return node->opcode() == IrOpcode::kLoad || |
| node->opcode() == IrOpcode::kLoadImmutable |
| ? LoadRepresentationOf(node->op()) |
| : MachineType::None(); |
| } |
| |
| bool IsIntConstant(Node* node) { |
| return node->opcode() == IrOpcode::kInt32Constant || |
| node->opcode() == IrOpcode::kInt64Constant; |
| } |
| |
| bool IsWordAnd(Node* node) { |
| return node->opcode() == IrOpcode::kWord32And || |
| node->opcode() == IrOpcode::kWord64And; |
| } |
| |
| // The result of WordAnd with a positive interger constant in X64 is known to |
| // be sign(zero)-extended. Comparing this result with another positive interger |
| // constant can have narrowed operand. |
| MachineType MachineTypeForNarrowWordAnd(Node* and_node, Node* constant_node) { |
| Node* and_left = and_node->InputAt(0); |
| Node* and_right = and_node->InputAt(1); |
| Node* and_constant_node = IsIntConstant(and_right) |
| ? and_right |
| : IsIntConstant(and_left) ? and_left : nullptr; |
| |
| if (and_constant_node != nullptr) { |
| int64_t and_constant = |
| and_constant_node->opcode() == IrOpcode::kInt32Constant |
| ? OpParameter<int32_t>(and_constant_node->op()) |
| : OpParameter<int64_t>(and_constant_node->op()); |
| int64_t cmp_constant = constant_node->opcode() == IrOpcode::kInt32Constant |
| ? OpParameter<int32_t>(constant_node->op()) |
| : OpParameter<int64_t>(constant_node->op()); |
| if (and_constant >= 0 && cmp_constant >= 0) { |
| int64_t constant = |
| and_constant > cmp_constant ? and_constant : cmp_constant; |
| if (constant <= std::numeric_limits<int8_t>::max()) { |
| return MachineType::Int8(); |
| } else if (constant <= std::numeric_limits<uint8_t>::max()) { |
| return MachineType::Uint8(); |
| } else if (constant <= std::numeric_limits<int16_t>::max()) { |
| return MachineType::Int16(); |
| } else if (constant <= std::numeric_limits<uint16_t>::max()) { |
| return MachineType::Uint16(); |
| } else if (constant <= std::numeric_limits<int32_t>::max()) { |
| return MachineType::Int32(); |
| } else if (constant <= std::numeric_limits<uint32_t>::max()) { |
| return MachineType::Uint32(); |
| } |
| } |
| } |
| |
| return MachineType::None(); |
| } |
| |
| // Tries to match the size of the given opcode to that of the operands, if |
| // possible. |
| InstructionCode TryNarrowOpcodeSize(InstructionCode opcode, Node* left, |
| Node* right, FlagsContinuation* cont) { |
| MachineType left_type = MachineType::None(); |
| MachineType right_type = MachineType::None(); |
| if (IsWordAnd(left) && IsIntConstant(right)) { |
| left_type = MachineTypeForNarrowWordAnd(left, right); |
| right_type = left_type; |
| } else if (IsWordAnd(right) && IsIntConstant(left)) { |
| right_type = MachineTypeForNarrowWordAnd(right, left); |
| left_type = right_type; |
| } else { |
| // TODO(epertoso): we can probably get some size information out phi nodes. |
| // If the load representations don't match, both operands will be |
| // zero/sign-extended to 32bit. |
| left_type = MachineTypeForNarrow(left, right); |
| right_type = MachineTypeForNarrow(right, left); |
| } |
| if (left_type == right_type) { |
| switch (left_type.representation()) { |
| case MachineRepresentation::kBit: |
| case MachineRepresentation::kWord8: { |
| if (opcode == kX64Test || opcode == kX64Test32) return kX64Test8; |
| if (opcode == kX64Cmp || opcode == kX64Cmp32) { |
| if (left_type.semantic() == MachineSemantic::kUint32) { |
| cont->OverwriteUnsignedIfSigned(); |
| } else { |
| CHECK_EQ(MachineSemantic::kInt32, left_type.semantic()); |
| } |
| return kX64Cmp8; |
| } |
| break; |
| } |
| case MachineRepresentation::kWord16: |
| if (opcode == kX64Test || opcode == kX64Test32) return kX64Test16; |
| if (opcode == kX64Cmp || opcode == kX64Cmp32) { |
| if (left_type.semantic() == MachineSemantic::kUint32) { |
| cont->OverwriteUnsignedIfSigned(); |
| } else { |
| CHECK_EQ(MachineSemantic::kInt32, left_type.semantic()); |
| } |
| return kX64Cmp16; |
| } |
| break; |
| case MachineRepresentation::kWord32: |
| if (opcode == kX64Test) return kX64Test32; |
| if (opcode == kX64Cmp) { |
| if (left_type.semantic() == MachineSemantic::kUint32) { |
| cont->OverwriteUnsignedIfSigned(); |
| } else { |
| CHECK_EQ(MachineSemantic::kInt32, left_type.semantic()); |
| } |
| return kX64Cmp32; |
| } |
| break; |
| #ifdef V8_COMPRESS_POINTERS |
| case MachineRepresentation::kTaggedSigned: |
| case MachineRepresentation::kTaggedPointer: |
| case MachineRepresentation::kTagged: |
| // When pointer compression is enabled the lower 32-bits uniquely |
| // identify tagged value. |
| if (opcode == kX64Cmp) return kX64Cmp32; |
| break; |
| #endif |
| default: |
| break; |
| } |
| } |
| return opcode; |
| } |
| |
| // Shared routine for multiple word compare operations. |
| void VisitWordCompare(InstructionSelector* selector, Node* node, |
| InstructionCode opcode, FlagsContinuation* cont) { |
| X64OperandGenerator g(selector); |
| Node* left = node->InputAt(0); |
| Node* right = node->InputAt(1); |
| |
| // The 32-bit comparisons automatically truncate Word64 |
| // values to Word32 range, no need to do that explicitly. |
| if (opcode == kX64Cmp32 || opcode == kX64Test32) { |
| if (left->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| left = left->InputAt(0); |
| } |
| |
| if (right->opcode() == IrOpcode::kTruncateInt64ToInt32) { |
| right = right->InputAt(0); |
| } |
| } |
| |
| opcode = TryNarrowOpcodeSize(opcode, left, right, cont); |
| |
| // If one of the two inputs is an immediate, make sure it's on the right, or |
| // if one of the two inputs is a memory operand, make sure it's on the left. |
| int effect_level = selector->GetEffectLevel(node, cont); |
| |
| if ((!g.CanBeImmediate(right) && g.CanBeImmediate(left)) || |
| (g.CanBeMemoryOperand(opcode, node, right, effect_level) && |
| !g.CanBeMemoryOperand(opcode, node, left, effect_level))) { |
| if (!node->op()->HasProperty(Operator::kCommutative)) cont->Commute(); |
| std::swap(left, right); |
| } |
| |
| // Match immediates on right side of comparison. |
| if (g.CanBeImmediate(right)) { |
| if (g.CanBeMemoryOperand(opcode, node, left, effect_level)) { |
| return VisitCompareWithMemoryOperand(selector, opcode, left, |
| g.UseImmediate(right), cont); |
| } |
| return VisitCompare(selector, opcode, g.Use(left), g.UseImmediate(right), |
| cont); |
| } |
| |
| // Match memory operands on left side of comparison. |
| if (g.CanBeMemoryOperand(opcode, node, left, effect_level)) { |
| return VisitCompareWithMemoryOperand(selector, opcode, left, |
| g.UseRegister(right), cont); |
| } |
| |
| return VisitCompare(selector, opcode, left, right, cont, |
| node->op()->HasProperty(Operator::kCommutative)); |
| } |
| |
| void VisitWord64EqualImpl(InstructionSelector* selector, Node* node, |
| FlagsContinuation* cont) { |
| if (selector->CanUseRootsRegister()) { |
| X64OperandGenerator g(selector); |
| const RootsTable& roots_table = selector->isolate()->roots_table(); |
| RootIndex root_index; |
| HeapObjectBinopMatcher m(node); |
| if (m.right().HasResolvedValue() && |
| roots_table.IsRootHandle(m.right().ResolvedValue(), &root_index)) { |
| InstructionCode opcode = |
| kX64Cmp | AddressingModeField::encode(kMode_Root); |
| return VisitCompare( |
| selector, opcode, |
| g.TempImmediate( |
| TurboAssemblerBase::RootRegisterOffsetForRootIndex(root_index)), |
| g.UseRegister(m.left().node()), cont); |
| } |
| } |
| VisitWordCompare(selector, node, kX64Cmp, cont); |
| } |
| |
| void VisitWord32EqualImpl(InstructionSelector* selector, Node* node, |
| FlagsContinuation* cont) { |
| if (COMPRESS_POINTERS_BOOL && selector->CanUseRootsRegister()) { |
| X64OperandGenerator g(selector); |
| const RootsTable& roots_table = selector->isolate()->roots_table(); |
| RootIndex root_index; |
| Node* left = nullptr; |
| Handle<HeapObject> right; |
| // HeapConstants and CompressedHeapConstants can be treated the same when |
| // using them as an input to a 32-bit comparison. Check whether either is |
| // present. |
| { |
| CompressedHeapObjectBinopMatcher m(node); |
| if (m.right().HasResolvedValue()) { |
| left = m.left().node(); |
| right = m.right().ResolvedValue(); |
| } else { |
| HeapObjectBinopMatcher m2(node); |
| if (m2.right().HasResolvedValue()) { |
| left = m2.left().node(); |
| right = m2.right().ResolvedValue(); |
| } |
| } |
| } |
| if (!right.is_null() && roots_table.IsRootHandle(right, &root_index)) { |
| DCHECK_NE(left, nullptr); |
| InstructionCode opcode = |
| kX64Cmp32 | AddressingModeField::encode(kMode_Root); |
| return VisitCompare( |
| selector, opcode, |
| g.TempImmediate( |
| TurboAssemblerBase::RootRegisterOffsetForRootIndex(root_index)), |
| g.UseRegister(left), cont); |
| } |
| } |
| VisitWordCompare(selector, node, kX64Cmp32, cont); |
| } |
| |
| // Shared routine for comparison with zero. |
| void VisitCompareZero(InstructionSelector* selector, Node* user, Node* node, |
| InstructionCode opcode, FlagsContinuation* cont) { |
| X64OperandGenerator g(selector); |
| if (cont->IsBranch() && |
| (cont->condition() == kNotEqual || cont->condition() == kEqual)) { |
| switch (node->opcode()) { |
| #define FLAGS_SET_BINOP_LIST(V) \ |
| V(kInt32Add, VisitBinop, kX64Add32) \ |
| V(kInt32Sub, VisitBinop, kX64Sub32) \ |
| V(kWord32And, VisitBinop, kX64And32) \ |
| V(kWord32Or, VisitBinop, kX64Or32) \ |
| V(kInt64Add, VisitBinop, kX64Add) \ |
| V(kInt64Sub, VisitBinop, kX64Sub) \ |
| V(kWord64And, VisitBinop, kX64And) \ |
| V(kWord64Or, VisitBinop, kX64Or) |
| #define FLAGS_SET_BINOP(opcode, Visit, archOpcode) \ |
| case IrOpcode::opcode: \ |
| if (selector->IsOnlyUserOfNodeInSameBlock(user, node)) { \ |
| return Visit(selector, node, archOpcode, cont); \ |
| } \ |
| break; |
| FLAGS_SET_BINOP_LIST(FLAGS_SET_BINOP) |
| #undef FLAGS_SET_BINOP_LIST |
| #undef FLAGS_SET_BINOP |
| |
| #define TRY_VISIT_WORD32_SHIFT TryVisitWordShift<Int32BinopMatcher, 32> |
| #define TRY_VISIT_WORD64_SHIFT TryVisitWordShift<Int64BinopMatcher, 64> |
| // Skip Word64Sar/Word32Sar since no instruction reduction in most cases. |
| #define FLAGS_SET_SHIFT_LIST(V) \ |
| V(kWord32Shl, TRY_VISIT_WORD32_SHIFT, kX64Shl32) \ |
| V(kWord32Shr, TRY_VISIT_WORD32_SHIFT, kX64Shr32) \ |
| V(kWord64Shl, TRY_VISIT_WORD64_SHIFT, kX64Shl) \ |
| V(kWord64Shr, TRY_VISIT_WORD64_SHIFT, kX64Shr) |
| #define FLAGS_SET_SHIFT(opcode, TryVisit, archOpcode) \ |
| case IrOpcode::opcode: \ |
| if (selector->IsOnlyUserOfNodeInSameBlock(user, node)) { \ |
| if (TryVisit(selector, node, archOpcode, cont)) return; \ |
| } \ |
| break; |
| FLAGS_SET_SHIFT_LIST(FLAGS_SET_SHIFT) |
| #undef TRY_VISIT_WORD32_SHIFT |
| #undef TRY_VISIT_WORD64_SHIFT |
| #undef FLAGS_SET_SHIFT_LIST |
| #undef FLAGS_SET_SHIFT |
| default: |
| break; |
| } |
| } |
| int effect_level = selector->GetEffectLevel(node, cont); |
| if (node->opcode() == IrOpcode::kLoad || |
| node->opcode() == IrOpcode::kLoadImmutable) { |
| switch (LoadRepresentationOf(node->op()).representation()) { |
| case MachineRepresentation::kWord8: |
| if (opcode == kX64Cmp32) { |
| opcode = kX64Cmp8; |
| } else if (opcode == kX64Test32) { |
| opcode = kX64Test8; |
| } |
| break; |
| case MachineRepresentation::kWord16: |
| if (opcode == kX64Cmp32) { |
| opcode = kX64Cmp16; |
| } else if (opcode == kX64Test32) { |
| opcode = kX64Test16; |
| } |
| break; |
| default: |
| break; |
| } |
| } |
| if (g.CanBeMemoryOperand(opcode, user, node, effect_level)) { |
| VisitCompareWithMemoryOperand(selector, opcode, node, g.TempImmediate(0), |
| cont); |
| } else { |
| VisitCompare(selector, opcode, g.Use(node), g.TempImmediate(0), cont); |
| } |
| } |
| |
| // Shared routine for multiple float32 compare operations (inputs commuted). |
| void VisitFloat32Compare(InstructionSelector* selector, Node* node, |
| FlagsContinuation* cont) { |
| Node* const left = node->InputAt(0); |
| Node* const right = node->InputAt(1); |
| InstructionCode const opcode = |
| selector->IsSupported(AVX) ? kAVXFloat32Cmp : kSSEFloat32Cmp; |
| VisitCompare(selector, opcode, right, left, cont, false); |
| } |
| |
| // Shared routine for multiple float64 compare operations (inputs commuted). |
| void VisitFloat64Compare(InstructionSelector* selector, Node* node, |
| FlagsContinuation* cont) { |
| Node* const left = node->InputAt(0); |
| Node* const right = node->InputAt(1); |
| InstructionCode const opcode = |
| selector->IsSupported(AVX) ? kAVXFloat64Cmp : kSSEFloat64Cmp; |
| VisitCompare(selector, opcode, right, left, cont, false); |
| } |
| |
| // Shared routine for Word32/Word64 Atomic Binops |
| void VisitAtomicBinop(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode, AtomicWidth width) { |
| X64OperandGenerator g(selector); |
| Node* base = node->InputAt(0); |
| Node* index = node->InputAt(1); |
| Node* value = node->InputAt(2); |
| AddressingMode addressing_mode; |
| InstructionOperand inputs[] = { |
| g.UseUniqueRegister(value), g.UseUniqueRegister(base), |
| g.GetEffectiveIndexOperand(index, &addressing_mode)}; |
| InstructionOperand outputs[] = {g.DefineAsFixed(node, rax)}; |
| InstructionOperand temps[] = {g.TempRegister()}; |
| InstructionCode code = opcode | AddressingModeField::encode(addressing_mode) | |
| AtomicWidthField::encode(width); |
| selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs, |
| arraysize(temps), temps); |
| } |
| |
| // Shared routine for Word32/Word64 Atomic CmpExchg |
| void VisitAtomicCompareExchange(InstructionSelector* selector, Node* node, |
| ArchOpcode opcode, AtomicWidth width) { |
| X64OperandGenerator g(selector); |
| Node* base = node->InputAt(0); |
| Node* index = node->InputAt(1); |
| Node* old_value = node->InputAt(2); |
| Node* new_value = node->InputAt(3); |
| AddressingMode addressing_mode; |
| InstructionOperand inputs[] = { |
| g.UseFixed(old_value, rax), g.UseUniqueRegister(new_value), |
| g.UseUniqueRegister(base), |
| g.GetEffectiveIndexOperand(index, &addressing_mode)}; |
| InstructionOperand outputs[] = {g.DefineAsFixed(node, rax)}; |
| InstructionCode code = opcode | AddressingModeField::encode(addressing_mode) | |
| AtomicWidthField::encode(width); |
| selector->Emit(code, arraysize(outputs), outputs, arraysize(inputs), inputs); |
| } |
| |
| // Used instead of CanCover in VisitWordCompareZero: even if CanCover(user, |
| // node) returns false, if |node| is a comparison, then it does not require any |
| // registers, and can thus be covered by |user|. |
| bool CanCoverForCompareZero(InstructionSelector* selector, Node* user, |
| Node* node) { |
| if (selector->CanCover(user, node)) { |
| return true; |
| } |
| // Checking if |node| is a comparison. If so, it doesn't required any |
| // registers, and, as such, it can always be covered by |user|. |
| switch (node->opcode()) { |
| #define CHECK_CMP_OP(op) \ |
| case IrOpcode::k##op: \ |
| return true; |
| MACHINE_COMPARE_BINOP_LIST(CHECK_CMP_OP) |
| #undef CHECK_CMP_OP |
| default: |
| break; |
| } |
| return false; |
| } |
| |
| } // namespace |
| |
| // Shared routine for word comparison against zero. |
| void InstructionSelector::VisitWordCompareZero(Node* user, Node* value, |
| FlagsContinuation* cont) { |
| // Try to combine with comparisons against 0 by simply inverting the branch. |
| while (value->opcode() == IrOpcode::kWord32Equal && CanCover(user, value)) { |
| Int32BinopMatcher m(value); |
| if (!m.right().Is(0)) break; |
| |
| user = value; |
| value = m.left().node(); |
| cont->Negate(); |
| } |
| |
| if (CanCoverForCompareZero(this, user, value)) { |
| switch (value->opcode()) { |
| case IrOpcode::kWord32Equal: |
| cont->OverwriteAndNegateIfEqual(kEqual); |
| return VisitWord32EqualImpl(this, value, cont); |
| case IrOpcode::kInt32LessThan: |
| cont->OverwriteAndNegateIfEqual(kSignedLessThan); |
| return VisitWordCompare(this, value, kX64Cmp32, cont); |
| case IrOpcode::kInt32LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kSignedLessThanOrEqual); |
| return VisitWordCompare(this, value, kX64Cmp32, cont); |
| case IrOpcode::kUint32LessThan: |
| cont->OverwriteAndNegateIfEqual(kUnsignedLessThan); |
| return VisitWordCompare(this, value, kX64Cmp32, cont); |
| case IrOpcode::kUint32LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kUnsignedLessThanOrEqual); |
| return VisitWordCompare(this, value, kX64Cmp32, cont); |
| case IrOpcode::kWord64Equal: { |
| cont->OverwriteAndNegateIfEqual(kEqual); |
| Int64BinopMatcher m(value); |
| if (m.right().Is(0) && CanCover(user, value)) { |
| // Try to combine the branch with a comparison. |
| Node* const eq_user = m.node(); |
| Node* const eq_value = m.left().node(); |
| if (CanCover(eq_user, eq_value)) { |
| switch (eq_value->opcode()) { |
| case IrOpcode::kInt64Sub: |
| return VisitWordCompare(this, eq_value, kX64Cmp, cont); |
| case IrOpcode::kWord64And: |
| return VisitWordCompare(this, eq_value, kX64Test, cont); |
| default: |
| break; |
| } |
| } |
| return VisitCompareZero(this, eq_user, eq_value, kX64Cmp, cont); |
| } |
| return VisitWord64EqualImpl(this, value, cont); |
| } |
| case IrOpcode::kInt64LessThan: |
| cont->OverwriteAndNegateIfEqual(kSignedLessThan); |
| return VisitWordCompare(this, value, kX64Cmp, cont); |
| case IrOpcode::kInt64LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kSignedLessThanOrEqual); |
| return VisitWordCompare(this, value, kX64Cmp, cont); |
| case IrOpcode::kUint64LessThan: |
| cont->OverwriteAndNegateIfEqual(kUnsignedLessThan); |
| return VisitWordCompare(this, value, kX64Cmp, cont); |
| case IrOpcode::kUint64LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kUnsignedLessThanOrEqual); |
| return VisitWordCompare(this, value, kX64Cmp, cont); |
| case IrOpcode::kFloat32Equal: |
| cont->OverwriteAndNegateIfEqual(kUnorderedEqual); |
| return VisitFloat32Compare(this, value, cont); |
| case IrOpcode::kFloat32LessThan: |
| cont->OverwriteAndNegateIfEqual(kUnsignedGreaterThan); |
| return VisitFloat32Compare(this, value, cont); |
| case IrOpcode::kFloat32LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kUnsignedGreaterThanOrEqual); |
| return VisitFloat32Compare(this, value, cont); |
| case IrOpcode::kFloat64Equal: |
| cont->OverwriteAndNegateIfEqual(kUnorderedEqual); |
| return VisitFloat64Compare(this, value, cont); |
| case IrOpcode::kFloat64LessThan: { |
| Float64BinopMatcher m(value); |
| if (m.left().Is(0.0) && m.right().IsFloat64Abs()) { |
| // This matches the pattern |
| // |
| // Float64LessThan(#0.0, Float64Abs(x)) |
| // |
| // which TurboFan generates for NumberToBoolean in the general case, |
| // and which evaluates to false if x is 0, -0 or NaN. We can compile |
| // this to a simple (v)ucomisd using not_equal flags condition, which |
| // avoids the costly Float64Abs. |
| cont->OverwriteAndNegateIfEqual(kNotEqual); |
| InstructionCode const opcode = |
| IsSupported(AVX) ? kAVXFloat64Cmp : kSSEFloat64Cmp; |
| return VisitCompare(this, opcode, m.left().node(), |
| m.right().InputAt(0), cont, false); |
| } |
| cont->OverwriteAndNegateIfEqual(kUnsignedGreaterThan); |
| return VisitFloat64Compare(this, value, cont); |
| } |
| case IrOpcode::kFloat64LessThanOrEqual: |
| cont->OverwriteAndNegateIfEqual(kUnsignedGreaterThanOrEqual); |
| return VisitFloat64Compare(this, value, cont); |
| case IrOpcode::kProjection: |
| // Check if this is the overflow output projection of an |
| // <Operation>WithOverflow node. |
| if (ProjectionIndexOf(value->op()) == 1u) { |
| // We cannot combine the <Operation>WithOverflow with this branch |
| |