| // Copyright (c) 2013, the Dart project authors. Please see the AUTHORS file |
| // for details. All rights reserved. Use of this source code is governed by a |
| // BSD-style license that can be found in the LICENSE file. |
| |
| #include "platform/globals.h" |
| #include "vm/globals.h" // Needed here to get TARGET_ARCH_X64. |
| #if defined(TARGET_ARCH_X64) |
| |
| #include "vm/compiler/backend/il.h" |
| |
| #include "vm/compiler/assembler/assembler.h" |
| #include "vm/compiler/backend/flow_graph.h" |
| #include "vm/compiler/backend/flow_graph_compiler.h" |
| #include "vm/compiler/backend/locations.h" |
| #include "vm/compiler/backend/locations_helpers.h" |
| #include "vm/compiler/backend/range_analysis.h" |
| #include "vm/compiler/ffi/native_calling_convention.h" |
| #include "vm/compiler/jit/compiler.h" |
| #include "vm/dart_entry.h" |
| #include "vm/instructions.h" |
| #include "vm/object_store.h" |
| #include "vm/parser.h" |
| #include "vm/stack_frame.h" |
| #include "vm/stub_code.h" |
| #include "vm/symbols.h" |
| #include "vm/type_testing_stubs.h" |
| |
| #define __ compiler->assembler()-> |
| #define Z (compiler->zone()) |
| |
| namespace dart { |
| |
| // Generic summary for call instructions that have all arguments pushed |
| // on the stack and return the result in a fixed register RAX (or XMM0 if |
| // the return type is double). |
| LocationSummary* Instruction::MakeCallSummary(Zone* zone, |
| const Instruction* instr, |
| LocationSummary* locs) { |
| ASSERT(locs == nullptr || locs->always_calls()); |
| LocationSummary* result = |
| ((locs == nullptr) |
| ? (new (zone) LocationSummary(zone, 0, 0, LocationSummary::kCall)) |
| : locs); |
| const auto representation = instr->representation(); |
| switch (representation) { |
| case kTagged: |
| case kUnboxedInt64: |
| result->set_out( |
| 0, Location::RegisterLocation(CallingConventions::kReturnReg)); |
| break; |
| case kUnboxedDouble: |
| result->set_out( |
| 0, Location::FpuRegisterLocation(CallingConventions::kReturnFpuReg)); |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| return result; |
| } |
| |
| LocationSummary* LoadIndexedUnsafeInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| |
| locs->set_in(0, Location::RequiresRegister()); |
| switch (representation()) { |
| case kTagged: |
| case kUnboxedInt64: |
| locs->set_out(0, Location::RequiresRegister()); |
| break; |
| case kUnboxedDouble: |
| locs->set_out(0, Location::RequiresFpuRegister()); |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| return locs; |
| } |
| |
| void LoadIndexedUnsafeInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(RequiredInputRepresentation(0) == kTagged); // It is a Smi. |
| ASSERT(kSmiTag == 0); |
| ASSERT(kSmiTagSize == 1); |
| |
| const Register index = locs()->in(0).reg(); |
| #if defined(DART_COMPRESSED_POINTERS) |
| // No addressing mode will ignore the upper bits. Cannot use the shorter `orl` |
| // to clear the upper bits as this instructions uses negative indices as part |
| // of FP-relative loads. |
| // TODO(compressed-pointers): Can we guarantee the index is already |
| // sign-extended if always comes for an args-descriptor load? |
| __ movsxd(index, index); |
| #endif |
| |
| switch (representation()) { |
| case kTagged: |
| case kUnboxedInt64: { |
| const auto out = locs()->out(0).reg(); |
| __ movq(out, compiler::Address(base_reg(), index, TIMES_4, offset())); |
| break; |
| } |
| case kUnboxedDouble: { |
| const auto out = locs()->out(0).fpu_reg(); |
| __ movsd(out, compiler::Address(base_reg(), index, TIMES_4, offset())); |
| break; |
| } |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| } |
| |
| DEFINE_BACKEND(StoreIndexedUnsafe, |
| (NoLocation, Register index, Register value)) { |
| ASSERT(instr->RequiredInputRepresentation( |
| StoreIndexedUnsafeInstr::kIndexPos) == kTagged); // It is a Smi. |
| #if defined(DART_COMPRESSED_POINTERS) |
| // No addressing mode will ignore the upper bits. Cannot use the shorter `orl` |
| // to clear the upper bits as this instructions uses negative indices as part |
| // of FP-relative stores. |
| // TODO(compressed-pointers): Can we guarantee the index is already |
| // sign-extended if always comes for an args-descriptor load? |
| __ movsxd(index, index); |
| #endif |
| __ movq(compiler::Address(instr->base_reg(), index, TIMES_4, instr->offset()), |
| value); |
| |
| ASSERT(kSmiTag == 0); |
| ASSERT(kSmiTagSize == 1); |
| } |
| |
| DEFINE_BACKEND(TailCall, (NoLocation, Fixed<Register, ARGS_DESC_REG>)) { |
| compiler->EmitTailCallToStub(instr->code()); |
| |
| // Even though the TailCallInstr will be the last instruction in a basic |
| // block, the flow graph compiler will emit native code for other blocks after |
| // the one containing this instruction and needs to be able to use the pool. |
| // (The `LeaveDartFrame` above disables usages of the pool.) |
| __ set_constant_pool_allowed(true); |
| } |
| |
| LocationSummary* MemoryCopyInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 5; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(kSrcPos, Location::RegisterLocation(RSI)); |
| locs->set_in(kDestPos, Location::RegisterLocation(RDI)); |
| locs->set_in(kSrcStartPos, Location::WritableRegister()); |
| locs->set_in(kDestStartPos, Location::WritableRegister()); |
| locs->set_in(kLengthPos, Location::RegisterLocation(RCX)); |
| return locs; |
| } |
| |
| void MemoryCopyInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| const Register src_start_reg = locs()->in(kSrcStartPos).reg(); |
| const Register dest_start_reg = locs()->in(kDestStartPos).reg(); |
| |
| EmitComputeStartPointer(compiler, src_cid_, src_start(), RSI, src_start_reg); |
| EmitComputeStartPointer(compiler, dest_cid_, dest_start(), RDI, |
| dest_start_reg); |
| if (element_size_ <= 8) { |
| __ SmiUntag(RCX); |
| } else { |
| #if defined(DART_COMPRESSED_POINTERS) |
| __ orl(RCX, RCX); |
| #endif |
| } |
| switch (element_size_) { |
| case 1: |
| __ rep_movsb(); |
| break; |
| case 2: |
| __ rep_movsw(); |
| break; |
| case 4: |
| __ rep_movsd(); |
| break; |
| case 8: |
| case 16: |
| __ rep_movsq(); |
| break; |
| } |
| } |
| |
| void MemoryCopyInstr::EmitComputeStartPointer(FlowGraphCompiler* compiler, |
| classid_t array_cid, |
| Value* start, |
| Register array_reg, |
| Register start_reg) { |
| intptr_t offset; |
| if (IsTypedDataBaseClassId(array_cid)) { |
| __ movq(array_reg, |
| compiler::FieldAddress( |
| array_reg, compiler::target::PointerBase::data_offset())); |
| offset = 0; |
| } else { |
| switch (array_cid) { |
| case kOneByteStringCid: |
| offset = |
| compiler::target::OneByteString::data_offset() - kHeapObjectTag; |
| break; |
| case kTwoByteStringCid: |
| offset = |
| compiler::target::TwoByteString::data_offset() - kHeapObjectTag; |
| break; |
| case kExternalOneByteStringCid: |
| __ movq(array_reg, |
| compiler::FieldAddress(array_reg, |
| compiler::target::ExternalOneByteString:: |
| external_data_offset())); |
| offset = 0; |
| break; |
| case kExternalTwoByteStringCid: |
| __ movq(array_reg, |
| compiler::FieldAddress(array_reg, |
| compiler::target::ExternalTwoByteString:: |
| external_data_offset())); |
| offset = 0; |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| } |
| ScaleFactor scale; |
| switch (element_size_) { |
| case 1: |
| __ SmiUntag(start_reg); |
| scale = TIMES_1; |
| break; |
| case 2: |
| #if defined(DART_COMPRESSED_POINTERS) |
| // Clear garbage upper bits, as no form of lea will ignore them. Assume |
| // start is positive to use the shorter orl over the longer movsxd. |
| __ orl(start_reg, start_reg); |
| #endif |
| scale = TIMES_1; |
| break; |
| case 4: |
| #if defined(DART_COMPRESSED_POINTERS) |
| __ orl(start_reg, start_reg); |
| #endif |
| scale = TIMES_2; |
| break; |
| case 8: |
| #if defined(DART_COMPRESSED_POINTERS) |
| __ orl(start_reg, start_reg); |
| #endif |
| scale = TIMES_4; |
| break; |
| case 16: |
| #if defined(DART_COMPRESSED_POINTERS) |
| __ orl(start_reg, start_reg); |
| #endif |
| scale = TIMES_8; |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| __ leaq(array_reg, compiler::Address(array_reg, start_reg, scale, offset)); |
| } |
| |
| LocationSummary* PushArgumentInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| if (representation() == kUnboxedDouble) { |
| locs->set_in(0, Location::RequiresFpuRegister()); |
| } else if (representation() == kUnboxedInt64) { |
| locs->set_in(0, Location::RequiresRegister()); |
| } else { |
| locs->set_in(0, LocationAnyOrConstant(value())); |
| } |
| return locs; |
| } |
| |
| void PushArgumentInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // In SSA mode, we need an explicit push. Nothing to do in non-SSA mode |
| // where arguments are pushed by their definitions. |
| if (compiler->is_optimizing()) { |
| Location value = locs()->in(0); |
| if (value.IsRegister()) { |
| __ pushq(value.reg()); |
| } else if (value.IsConstant()) { |
| __ PushObject(value.constant()); |
| } else if (value.IsFpuRegister()) { |
| __ AddImmediate(RSP, compiler::Immediate(-kDoubleSize)); |
| __ movsd(compiler::Address(RSP, 0), value.fpu_reg()); |
| } else { |
| ASSERT(value.IsStackSlot()); |
| __ pushq(LocationToStackSlotAddress(value)); |
| } |
| } |
| } |
| |
| LocationSummary* ReturnInstr::MakeLocationSummary(Zone* zone, bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| switch (representation()) { |
| case kTagged: |
| case kUnboxedInt64: |
| locs->set_in(0, |
| Location::RegisterLocation(CallingConventions::kReturnReg)); |
| break; |
| case kUnboxedDouble: |
| locs->set_in( |
| 0, Location::FpuRegisterLocation(CallingConventions::kReturnFpuReg)); |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| return locs; |
| } |
| |
| // Attempt optimized compilation at return instruction instead of at the entry. |
| // The entry needs to be patchable, no inlined objects are allowed in the area |
| // that will be overwritten by the patch instruction: a jump). |
| void ReturnInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| if (locs()->in(0).IsRegister()) { |
| const Register result = locs()->in(0).reg(); |
| ASSERT(result == CallingConventions::kReturnReg); |
| } else { |
| ASSERT(locs()->in(0).IsFpuRegister()); |
| const FpuRegister result = locs()->in(0).fpu_reg(); |
| ASSERT(result == CallingConventions::kReturnFpuReg); |
| } |
| |
| if (compiler->parsed_function().function().IsSuspendableFunction()) { |
| ASSERT(compiler->flow_graph().graph_entry()->NeedsFrame()); |
| const Code& stub = GetReturnStub(compiler); |
| compiler->EmitJumpToStub(stub); |
| return; |
| } |
| |
| if (!compiler->flow_graph().graph_entry()->NeedsFrame()) { |
| __ ret(); |
| return; |
| } |
| |
| #if defined(DEBUG) |
| __ Comment("Stack Check"); |
| compiler::Label done; |
| const intptr_t fp_sp_dist = |
| (compiler::target::frame_layout.first_local_from_fp + 1 - |
| compiler->StackSize()) * |
| kWordSize; |
| ASSERT(fp_sp_dist <= 0); |
| __ movq(RDI, RSP); |
| __ subq(RDI, RBP); |
| __ CompareImmediate(RDI, compiler::Immediate(fp_sp_dist)); |
| __ j(EQUAL, &done, compiler::Assembler::kNearJump); |
| __ int3(); |
| __ Bind(&done); |
| #endif |
| ASSERT(__ constant_pool_allowed()); |
| if (yield_index() != UntaggedPcDescriptors::kInvalidYieldIndex) { |
| compiler->EmitYieldPositionMetadata(source(), yield_index()); |
| } |
| __ LeaveDartFrame(); // Disallows constant pool use. |
| __ ret(); |
| // This ReturnInstr may be emitted out of order by the optimizer. The next |
| // block may be a target expecting a properly set constant pool pointer. |
| __ set_constant_pool_allowed(true); |
| } |
| |
| static const RegisterSet kCalleeSaveRegistersSet( |
| CallingConventions::kCalleeSaveCpuRegisters, |
| CallingConventions::kCalleeSaveXmmRegisters); |
| |
| // Keep in sync with NativeEntryInstr::EmitNativeCode. |
| void NativeReturnInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| EmitReturnMoves(compiler); |
| |
| __ LeaveDartFrame(); |
| |
| // Pop dummy return address. |
| __ popq(TMP); |
| |
| // Anything besides the return register. |
| const Register vm_tag_reg = RBX; |
| const Register old_exit_frame_reg = RCX; |
| const Register old_exit_through_ffi_reg = RDI; |
| |
| __ popq(old_exit_frame_reg); |
| |
| __ popq(old_exit_through_ffi_reg); |
| |
| // Restore top_resource. |
| __ popq(TMP); |
| __ movq( |
| compiler::Address(THR, compiler::target::Thread::top_resource_offset()), |
| TMP); |
| |
| __ popq(vm_tag_reg); |
| |
| // If we were called by a trampoline, it will enter the safepoint on our |
| // behalf. |
| __ TransitionGeneratedToNative( |
| vm_tag_reg, old_exit_frame_reg, old_exit_through_ffi_reg, |
| /*enter_safepoint=*/!NativeCallbackTrampolines::Enabled()); |
| |
| // Restore C++ ABI callee-saved registers. |
| __ PopRegisters(kCalleeSaveRegistersSet); |
| |
| #if defined(DART_TARGET_OS_FUCHSIA) && defined(USING_SHADOW_CALL_STACK) |
| #error Unimplemented |
| #endif |
| |
| // Leave the entry frame. |
| __ LeaveFrame(); |
| |
| // Leave the dummy frame holding the pushed arguments. |
| __ LeaveFrame(); |
| |
| __ ret(); |
| |
| // For following blocks. |
| __ set_constant_pool_allowed(true); |
| } |
| |
| // Detect pattern when one value is zero and another is a power of 2. |
| static bool IsPowerOfTwoKind(intptr_t v1, intptr_t v2) { |
| return (Utils::IsPowerOfTwo(v1) && (v2 == 0)) || |
| (Utils::IsPowerOfTwo(v2) && (v1 == 0)); |
| } |
| |
| LocationSummary* IfThenElseInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| comparison()->InitializeLocationSummary(zone, opt); |
| // TODO(dartbug.com/30952) support convertion of Register to corresponding |
| // least significant byte register (e.g. RAX -> AL, RSI -> SIL, r15 -> r15b). |
| comparison()->locs()->set_out(0, Location::RegisterLocation(RDX)); |
| return comparison()->locs(); |
| } |
| |
| void IfThenElseInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(locs()->out(0).reg() == RDX); |
| |
| // Clear upper part of the out register. We are going to use setcc on it |
| // which is a byte move. |
| __ xorq(RDX, RDX); |
| |
| // Emit comparison code. This must not overwrite the result register. |
| // IfThenElseInstr::Supports() should prevent EmitComparisonCode from using |
| // the labels or returning an invalid condition. |
| BranchLabels labels = {NULL, NULL, NULL}; |
| Condition true_condition = comparison()->EmitComparisonCode(compiler, labels); |
| ASSERT(true_condition != kInvalidCondition); |
| |
| const bool is_power_of_two_kind = IsPowerOfTwoKind(if_true_, if_false_); |
| |
| intptr_t true_value = if_true_; |
| intptr_t false_value = if_false_; |
| |
| if (is_power_of_two_kind) { |
| if (true_value == 0) { |
| // We need to have zero in RDX on true_condition. |
| true_condition = InvertCondition(true_condition); |
| } |
| } else { |
| if (true_value == 0) { |
| // Swap values so that false_value is zero. |
| intptr_t temp = true_value; |
| true_value = false_value; |
| false_value = temp; |
| } else { |
| true_condition = InvertCondition(true_condition); |
| } |
| } |
| |
| __ setcc(true_condition, DL); |
| |
| if (is_power_of_two_kind) { |
| const intptr_t shift = |
| Utils::ShiftForPowerOfTwo(Utils::Maximum(true_value, false_value)); |
| __ shlq(RDX, compiler::Immediate(shift + kSmiTagSize)); |
| } else { |
| __ decq(RDX); |
| __ AndImmediate(RDX, compiler::Immediate(Smi::RawValue(true_value) - |
| Smi::RawValue(false_value))); |
| if (false_value != 0) { |
| __ AddImmediate(RDX, compiler::Immediate(Smi::RawValue(false_value))); |
| } |
| } |
| } |
| |
| LocationSummary* LoadLocalInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 0; |
| const intptr_t stack_index = |
| compiler::target::frame_layout.FrameSlotForVariable(&local()); |
| return LocationSummary::Make(zone, kNumInputs, |
| Location::StackSlot(stack_index, FPREG), |
| LocationSummary::kNoCall); |
| } |
| |
| void LoadLocalInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(!compiler->is_optimizing()); |
| // Nothing to do. |
| } |
| |
| LocationSummary* StoreLocalInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| return LocationSummary::Make(zone, kNumInputs, Location::SameAsFirstInput(), |
| LocationSummary::kNoCall); |
| } |
| |
| void StoreLocalInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| Register value = locs()->in(0).reg(); |
| Register result = locs()->out(0).reg(); |
| ASSERT(result == value); // Assert that register assignment is correct. |
| __ movq(compiler::Address( |
| RBP, compiler::target::FrameOffsetInBytesForVariable(&local())), |
| value); |
| } |
| |
| LocationSummary* ConstantInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 0; |
| return LocationSummary::Make(zone, kNumInputs, |
| compiler::Assembler::IsSafe(value()) |
| ? Location::Constant(this) |
| : Location::RequiresRegister(), |
| LocationSummary::kNoCall); |
| } |
| |
| void ConstantInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // The register allocator drops constant definitions that have no uses. |
| Location out = locs()->out(0); |
| ASSERT(out.IsRegister() || out.IsConstant() || out.IsInvalid()); |
| if (out.IsRegister()) { |
| Register result = out.reg(); |
| __ LoadObject(result, value()); |
| } |
| } |
| |
| void ConstantInstr::EmitMoveToLocation(FlowGraphCompiler* compiler, |
| const Location& destination, |
| Register tmp, |
| intptr_t pair_index) { |
| ASSERT(pair_index == 0); // No pair representation needed on 64-bit. |
| if (destination.IsRegister()) { |
| if (RepresentationUtils::IsUnboxedInteger(representation())) { |
| const int64_t value = Integer::Cast(value_).AsInt64Value(); |
| if (value == 0) { |
| __ xorl(destination.reg(), destination.reg()); |
| } else { |
| __ movq(destination.reg(), compiler::Immediate(value)); |
| } |
| } else { |
| ASSERT(representation() == kTagged); |
| __ LoadObject(destination.reg(), value_); |
| } |
| } else if (destination.IsFpuRegister()) { |
| __ LoadDImmediate(destination.fpu_reg(), Double::Cast(value_).value()); |
| } else if (destination.IsDoubleStackSlot()) { |
| __ LoadDImmediate(FpuTMP, Double::Cast(value_).value()); |
| __ movsd(LocationToStackSlotAddress(destination), FpuTMP); |
| } else { |
| ASSERT(destination.IsStackSlot()); |
| if (RepresentationUtils::IsUnboxedInteger(representation())) { |
| const int64_t value = Integer::Cast(value_).AsInt64Value(); |
| __ movq(LocationToStackSlotAddress(destination), |
| compiler::Immediate(value)); |
| } else { |
| ASSERT(representation() == kTagged); |
| __ StoreObject(LocationToStackSlotAddress(destination), value_); |
| } |
| } |
| } |
| |
| LocationSummary* UnboxedConstantInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const bool is_unboxed_int = |
| RepresentationUtils::IsUnboxedInteger(representation()); |
| ASSERT(!is_unboxed_int || RepresentationUtils::ValueSize(representation()) <= |
| compiler::target::kWordSize); |
| const intptr_t kNumInputs = 0; |
| const intptr_t kNumTemps = is_unboxed_int ? 0 : 1; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| if (is_unboxed_int) { |
| locs->set_out(0, Location::RequiresRegister()); |
| } else { |
| switch (representation()) { |
| case kUnboxedDouble: |
| locs->set_out(0, Location::RequiresFpuRegister()); |
| locs->set_temp(0, Location::RequiresRegister()); |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| } |
| return locs; |
| } |
| |
| void UnboxedConstantInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // The register allocator drops constant definitions that have no uses. |
| if (!locs()->out(0).IsInvalid()) { |
| const Register scratch = |
| RepresentationUtils::IsUnboxedInteger(representation()) |
| ? kNoRegister |
| : locs()->temp(0).reg(); |
| EmitMoveToLocation(compiler, locs()->out(0), scratch); |
| } |
| } |
| |
| LocationSummary* AssertAssignableInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| auto const dst_type_loc = |
| LocationFixedRegisterOrConstant(dst_type(), TypeTestABI::kDstTypeReg); |
| |
| // We want to prevent spilling of the inputs (e.g. function/instantiator tav), |
| // since TTS preserves them. So we make this a `kNoCall` summary, |
| // even though most other registers can be modified by the stub. To tell the |
| // register allocator about it, we reserve all the other registers as |
| // temporary registers. |
| // TODO(http://dartbug.com/32788): Simplify this. |
| |
| const intptr_t kNonChangeableInputRegs = |
| (1 << TypeTestABI::kInstanceReg) | |
| ((dst_type_loc.IsRegister() ? 1 : 0) << TypeTestABI::kDstTypeReg) | |
| (1 << TypeTestABI::kInstantiatorTypeArgumentsReg) | |
| (1 << TypeTestABI::kFunctionTypeArgumentsReg); |
| |
| const intptr_t kNumInputs = 4; |
| |
| // We invoke a stub that can potentially clobber any CPU register |
| // but can only clobber FPU registers on the slow path when |
| // entering runtime. Preserve all FPU registers that are |
| // not guarateed to be preserved by the ABI. |
| const intptr_t kCpuRegistersToPreserve = |
| kDartAvailableCpuRegs & ~kNonChangeableInputRegs; |
| const intptr_t kFpuRegistersToPreserve = |
| CallingConventions::kVolatileXmmRegisters & ~(1 << FpuTMP); |
| |
| const intptr_t kNumTemps = (Utils::CountOneBits64(kCpuRegistersToPreserve) + |
| Utils::CountOneBits64(kFpuRegistersToPreserve)); |
| |
| LocationSummary* summary = new (zone) LocationSummary( |
| zone, kNumInputs, kNumTemps, LocationSummary::kCallCalleeSafe); |
| summary->set_in(kInstancePos, |
| Location::RegisterLocation(TypeTestABI::kInstanceReg)); |
| summary->set_in(kDstTypePos, dst_type_loc); |
| summary->set_in( |
| kInstantiatorTAVPos, |
| Location::RegisterLocation(TypeTestABI::kInstantiatorTypeArgumentsReg)); |
| summary->set_in(kFunctionTAVPos, Location::RegisterLocation( |
| TypeTestABI::kFunctionTypeArgumentsReg)); |
| summary->set_out(0, Location::SameAsFirstInput()); |
| |
| // Let's reserve all registers except for the input ones. |
| intptr_t next_temp = 0; |
| for (intptr_t i = 0; i < kNumberOfCpuRegisters; ++i) { |
| const bool should_preserve = ((1 << i) & kCpuRegistersToPreserve) != 0; |
| if (should_preserve) { |
| summary->set_temp(next_temp++, |
| Location::RegisterLocation(static_cast<Register>(i))); |
| } |
| } |
| |
| for (intptr_t i = 0; i < kNumberOfFpuRegisters; i++) { |
| const bool should_preserve = ((1 << i) & kFpuRegistersToPreserve) != 0; |
| if (should_preserve) { |
| summary->set_temp(next_temp++, Location::FpuRegisterLocation( |
| static_cast<FpuRegister>(i))); |
| } |
| } |
| |
| return summary; |
| } |
| |
| void AssertBooleanInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(locs()->always_calls()); |
| |
| auto object_store = compiler->isolate_group()->object_store(); |
| const auto& assert_boolean_stub = |
| Code::ZoneHandle(compiler->zone(), object_store->assert_boolean_stub()); |
| |
| compiler::Label done; |
| __ testq( |
| AssertBooleanABI::kObjectReg, |
| compiler::Immediate(compiler::target::ObjectAlignment::kBoolVsNullMask)); |
| __ j(NOT_ZERO, &done, compiler::Assembler::kNearJump); |
| compiler->GenerateStubCall(source(), assert_boolean_stub, |
| /*kind=*/UntaggedPcDescriptors::kOther, locs(), |
| deopt_id(), env()); |
| __ Bind(&done); |
| } |
| |
| static Condition TokenKindToIntCondition(Token::Kind kind) { |
| switch (kind) { |
| case Token::kEQ: |
| return EQUAL; |
| case Token::kNE: |
| return NOT_EQUAL; |
| case Token::kLT: |
| return LESS; |
| case Token::kGT: |
| return GREATER; |
| case Token::kLTE: |
| return LESS_EQUAL; |
| case Token::kGTE: |
| return GREATER_EQUAL; |
| default: |
| UNREACHABLE(); |
| return OVERFLOW; |
| } |
| } |
| |
| LocationSummary* EqualityCompareInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 2; |
| if (operation_cid() == kDoubleCid) { |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RequiresFpuRegister()); |
| locs->set_in(1, Location::RequiresFpuRegister()); |
| locs->set_out(0, Location::RequiresRegister()); |
| return locs; |
| } |
| if (operation_cid() == kSmiCid || operation_cid() == kMintCid) { |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| if (is_null_aware()) { |
| locs->set_in(0, Location::RequiresRegister()); |
| locs->set_in(1, Location::RequiresRegister()); |
| } else { |
| locs->set_in(0, LocationRegisterOrConstant(left())); |
| // Only one input can be a constant operand. The case of two constant |
| // operands should be handled by constant propagation. |
| // Only right can be a stack slot. |
| locs->set_in(1, locs->in(0).IsConstant() |
| ? Location::RequiresRegister() |
| : LocationRegisterOrConstant(right())); |
| } |
| locs->set_out(0, Location::RequiresRegister()); |
| return locs; |
| } |
| UNREACHABLE(); |
| return NULL; |
| } |
| |
| static void LoadValueCid(FlowGraphCompiler* compiler, |
| Register value_cid_reg, |
| Register value_reg, |
| compiler::Label* value_is_smi = NULL) { |
| compiler::Label done; |
| if (value_is_smi == NULL) { |
| __ LoadImmediate(value_cid_reg, compiler::Immediate(kSmiCid)); |
| } |
| __ testq(value_reg, compiler::Immediate(kSmiTagMask)); |
| if (value_is_smi == NULL) { |
| __ j(ZERO, &done, compiler::Assembler::kNearJump); |
| } else { |
| __ j(ZERO, value_is_smi); |
| } |
| __ LoadClassId(value_cid_reg, value_reg); |
| __ Bind(&done); |
| } |
| |
| static Condition FlipCondition(Condition condition) { |
| switch (condition) { |
| case EQUAL: |
| return EQUAL; |
| case NOT_EQUAL: |
| return NOT_EQUAL; |
| case LESS: |
| return GREATER; |
| case LESS_EQUAL: |
| return GREATER_EQUAL; |
| case GREATER: |
| return LESS; |
| case GREATER_EQUAL: |
| return LESS_EQUAL; |
| case BELOW: |
| return ABOVE; |
| case BELOW_EQUAL: |
| return ABOVE_EQUAL; |
| case ABOVE: |
| return BELOW; |
| case ABOVE_EQUAL: |
| return BELOW_EQUAL; |
| default: |
| UNIMPLEMENTED(); |
| return EQUAL; |
| } |
| } |
| |
| static void EmitBranchOnCondition( |
| FlowGraphCompiler* compiler, |
| Condition true_condition, |
| BranchLabels labels, |
| compiler::Assembler::JumpDistance jump_distance = |
| compiler::Assembler::kFarJump) { |
| if (labels.fall_through == labels.false_label) { |
| // If the next block is the false successor, fall through to it. |
| __ j(true_condition, labels.true_label, jump_distance); |
| } else { |
| // If the next block is not the false successor, branch to it. |
| Condition false_condition = InvertCondition(true_condition); |
| __ j(false_condition, labels.false_label, jump_distance); |
| |
| // Fall through or jump to the true successor. |
| if (labels.fall_through != labels.true_label) { |
| __ jmp(labels.true_label, jump_distance); |
| } |
| } |
| } |
| |
| static Condition EmitSmiComparisonOp(FlowGraphCompiler* compiler, |
| const LocationSummary& locs, |
| Token::Kind kind) { |
| Location left = locs.in(0); |
| Location right = locs.in(1); |
| ASSERT(!left.IsConstant() || !right.IsConstant()); |
| |
| Condition true_condition = TokenKindToIntCondition(kind); |
| if (left.IsConstant() || right.IsConstant()) { |
| // Ensure constant is on the right. |
| ConstantInstr* constant = NULL; |
| if (left.IsConstant()) { |
| constant = left.constant_instruction(); |
| Location tmp = right; |
| right = left; |
| left = tmp; |
| true_condition = FlipCondition(true_condition); |
| } else { |
| constant = right.constant_instruction(); |
| } |
| |
| if (RepresentationUtils::IsUnboxedInteger(constant->representation())) { |
| int64_t value; |
| const bool ok = compiler::HasIntegerValue(constant->value(), &value); |
| RELEASE_ASSERT(ok); |
| __ OBJ(cmp)(left.reg(), compiler::Immediate(value)); |
| } else { |
| ASSERT(constant->representation() == kTagged); |
| __ CompareObject(left.reg(), right.constant()); |
| } |
| } else if (right.IsStackSlot()) { |
| __ OBJ(cmp)(left.reg(), LocationToStackSlotAddress(right)); |
| } else { |
| __ OBJ(cmp)(left.reg(), right.reg()); |
| } |
| return true_condition; |
| } |
| |
| static Condition EmitInt64ComparisonOp(FlowGraphCompiler* compiler, |
| const LocationSummary& locs, |
| Token::Kind kind) { |
| Location left = locs.in(0); |
| Location right = locs.in(1); |
| ASSERT(!left.IsConstant() || !right.IsConstant()); |
| |
| Condition true_condition = TokenKindToIntCondition(kind); |
| if (left.IsConstant() || right.IsConstant()) { |
| // Ensure constant is on the right. |
| ConstantInstr* constant = NULL; |
| if (left.IsConstant()) { |
| constant = left.constant_instruction(); |
| Location tmp = right; |
| right = left; |
| left = tmp; |
| true_condition = FlipCondition(true_condition); |
| } else { |
| constant = right.constant_instruction(); |
| } |
| |
| if (RepresentationUtils::IsUnboxedInteger(constant->representation())) { |
| int64_t value; |
| const bool ok = compiler::HasIntegerValue(constant->value(), &value); |
| RELEASE_ASSERT(ok); |
| __ cmpq(left.reg(), compiler::Immediate(value)); |
| } else { |
| UNREACHABLE(); |
| } |
| } else if (right.IsStackSlot()) { |
| __ cmpq(left.reg(), LocationToStackSlotAddress(right)); |
| } else { |
| __ cmpq(left.reg(), right.reg()); |
| } |
| return true_condition; |
| } |
| |
| static Condition EmitNullAwareInt64ComparisonOp(FlowGraphCompiler* compiler, |
| const LocationSummary& locs, |
| Token::Kind kind, |
| BranchLabels labels) { |
| ASSERT((kind == Token::kEQ) || (kind == Token::kNE)); |
| const Register left = locs.in(0).reg(); |
| const Register right = locs.in(1).reg(); |
| const Condition true_condition = TokenKindToIntCondition(kind); |
| compiler::Label* equal_result = |
| (true_condition == EQUAL) ? labels.true_label : labels.false_label; |
| compiler::Label* not_equal_result = |
| (true_condition == EQUAL) ? labels.false_label : labels.true_label; |
| |
| // Check if operands have the same value. If they don't, then they could |
| // be equal only if both of them are Mints with the same value. |
| __ OBJ(cmp)(left, right); |
| __ j(EQUAL, equal_result); |
| __ OBJ(mov)(TMP, left); |
| __ OBJ(and)(TMP, right); |
| __ BranchIfSmi(TMP, not_equal_result); |
| __ CompareClassId(left, kMintCid); |
| __ j(NOT_EQUAL, not_equal_result); |
| __ CompareClassId(right, kMintCid); |
| __ j(NOT_EQUAL, not_equal_result); |
| __ movq(TMP, compiler::FieldAddress(left, Mint::value_offset())); |
| __ cmpq(TMP, compiler::FieldAddress(right, Mint::value_offset())); |
| return true_condition; |
| } |
| |
| static Condition TokenKindToDoubleCondition(Token::Kind kind) { |
| switch (kind) { |
| case Token::kEQ: |
| return EQUAL; |
| case Token::kNE: |
| return NOT_EQUAL; |
| case Token::kLT: |
| return BELOW; |
| case Token::kGT: |
| return ABOVE; |
| case Token::kLTE: |
| return BELOW_EQUAL; |
| case Token::kGTE: |
| return ABOVE_EQUAL; |
| default: |
| UNREACHABLE(); |
| return OVERFLOW; |
| } |
| } |
| |
| static Condition EmitDoubleComparisonOp(FlowGraphCompiler* compiler, |
| const LocationSummary& locs, |
| Token::Kind kind, |
| BranchLabels labels) { |
| XmmRegister left = locs.in(0).fpu_reg(); |
| XmmRegister right = locs.in(1).fpu_reg(); |
| |
| __ comisd(left, right); |
| |
| Condition true_condition = TokenKindToDoubleCondition(kind); |
| compiler::Label* nan_result = |
| (true_condition == NOT_EQUAL) ? labels.true_label : labels.false_label; |
| __ j(PARITY_EVEN, nan_result); |
| return true_condition; |
| } |
| |
| Condition EqualityCompareInstr::EmitComparisonCode(FlowGraphCompiler* compiler, |
| BranchLabels labels) { |
| if (is_null_aware()) { |
| ASSERT(operation_cid() == kMintCid); |
| return EmitNullAwareInt64ComparisonOp(compiler, *locs(), kind(), labels); |
| } |
| if (operation_cid() == kSmiCid) { |
| return EmitSmiComparisonOp(compiler, *locs(), kind()); |
| } else if (operation_cid() == kMintCid) { |
| return EmitInt64ComparisonOp(compiler, *locs(), kind()); |
| } else { |
| ASSERT(operation_cid() == kDoubleCid); |
| return EmitDoubleComparisonOp(compiler, *locs(), kind(), labels); |
| } |
| } |
| |
| void ComparisonInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| compiler::Label is_true, is_false; |
| BranchLabels labels = {&is_true, &is_false, &is_false}; |
| Condition true_condition = EmitComparisonCode(compiler, labels); |
| |
| Register result = locs()->out(0).reg(); |
| if (true_condition != kInvalidCondition) { |
| EmitBranchOnCondition(compiler, true_condition, labels, |
| compiler::Assembler::kNearJump); |
| } |
| // Note: We use branches instead of setcc or cmov even when the branch labels |
| // are otherwise unused, as this runs faster for the x86 processors tested on |
| // our benchmarking server. |
| compiler::Label done; |
| __ Bind(&is_false); |
| __ LoadObject(result, Bool::False()); |
| __ jmp(&done, compiler::Assembler::kNearJump); |
| __ Bind(&is_true); |
| __ LoadObject(result, Bool::True()); |
| __ Bind(&done); |
| } |
| |
| void ComparisonInstr::EmitBranchCode(FlowGraphCompiler* compiler, |
| BranchInstr* branch) { |
| BranchLabels labels = compiler->CreateBranchLabels(branch); |
| Condition true_condition = EmitComparisonCode(compiler, labels); |
| if (true_condition != kInvalidCondition) { |
| EmitBranchOnCondition(compiler, true_condition, labels); |
| } |
| } |
| |
| LocationSummary* TestSmiInstr::MakeLocationSummary(Zone* zone, bool opt) const { |
| const intptr_t kNumInputs = 2; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RequiresRegister()); |
| // Only one input can be a constant operand. The case of two constant |
| // operands should be handled by constant propagation. |
| locs->set_in(1, LocationRegisterOrConstant(right())); |
| return locs; |
| } |
| |
| Condition TestSmiInstr::EmitComparisonCode(FlowGraphCompiler* compiler, |
| BranchLabels labels) { |
| Register left_reg = locs()->in(0).reg(); |
| Location right = locs()->in(1); |
| if (right.IsConstant()) { |
| ASSERT(right.constant().IsSmi()); |
| const int64_t imm = Smi::RawValue(Smi::Cast(right.constant()).Value()); |
| __ TestImmediate(left_reg, compiler::Immediate(imm), |
| compiler::kObjectBytes); |
| } else { |
| __ OBJ(test)(left_reg, right.reg()); |
| } |
| Condition true_condition = (kind() == Token::kNE) ? NOT_ZERO : ZERO; |
| return true_condition; |
| } |
| |
| LocationSummary* TestCidsInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 1; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RequiresRegister()); |
| locs->set_temp(0, Location::RequiresRegister()); |
| locs->set_out(0, Location::RequiresRegister()); |
| return locs; |
| } |
| |
| Condition TestCidsInstr::EmitComparisonCode(FlowGraphCompiler* compiler, |
| BranchLabels labels) { |
| ASSERT((kind() == Token::kIS) || (kind() == Token::kISNOT)); |
| Register val_reg = locs()->in(0).reg(); |
| Register cid_reg = locs()->temp(0).reg(); |
| |
| compiler::Label* deopt = |
| CanDeoptimize() |
| ? compiler->AddDeoptStub(deopt_id(), ICData::kDeoptTestCids, |
| licm_hoisted_ ? ICData::kHoisted : 0) |
| : NULL; |
| |
| const intptr_t true_result = (kind() == Token::kIS) ? 1 : 0; |
| const ZoneGrowableArray<intptr_t>& data = cid_results(); |
| ASSERT(data[0] == kSmiCid); |
| bool result = data[1] == true_result; |
| __ testq(val_reg, compiler::Immediate(kSmiTagMask)); |
| __ j(ZERO, result ? labels.true_label : labels.false_label); |
| __ LoadClassId(cid_reg, val_reg); |
| for (intptr_t i = 2; i < data.length(); i += 2) { |
| const intptr_t test_cid = data[i]; |
| ASSERT(test_cid != kSmiCid); |
| result = data[i + 1] == true_result; |
| __ cmpq(cid_reg, compiler::Immediate(test_cid)); |
| __ j(EQUAL, result ? labels.true_label : labels.false_label); |
| } |
| // No match found, deoptimize or default action. |
| if (deopt == NULL) { |
| // If the cid is not in the list, jump to the opposite label from the cids |
| // that are in the list. These must be all the same (see asserts in the |
| // constructor). |
| compiler::Label* target = result ? labels.false_label : labels.true_label; |
| if (target != labels.fall_through) { |
| __ jmp(target); |
| } |
| } else { |
| __ jmp(deopt); |
| } |
| // Dummy result as this method already did the jump, there's no need |
| // for the caller to branch on a condition. |
| return kInvalidCondition; |
| } |
| |
| LocationSummary* RelationalOpInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 2; |
| const intptr_t kNumTemps = 0; |
| if (operation_cid() == kDoubleCid) { |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresFpuRegister()); |
| summary->set_in(1, Location::RequiresFpuRegister()); |
| summary->set_out(0, Location::RequiresRegister()); |
| return summary; |
| } |
| if (operation_cid() == kSmiCid || operation_cid() == kMintCid) { |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, LocationRegisterOrConstant(left())); |
| // Only one input can be a constant operand. The case of two constant |
| // operands should be handled by constant propagation. |
| summary->set_in(1, summary->in(0).IsConstant() |
| ? Location::RequiresRegister() |
| : LocationRegisterOrConstant(right())); |
| summary->set_out(0, Location::RequiresRegister()); |
| return summary; |
| } |
| UNREACHABLE(); |
| return NULL; |
| } |
| |
| Condition RelationalOpInstr::EmitComparisonCode(FlowGraphCompiler* compiler, |
| BranchLabels labels) { |
| if (operation_cid() == kSmiCid) { |
| return EmitSmiComparisonOp(compiler, *locs(), kind()); |
| } else if (operation_cid() == kMintCid) { |
| return EmitInt64ComparisonOp(compiler, *locs(), kind()); |
| } else { |
| ASSERT(operation_cid() == kDoubleCid); |
| return EmitDoubleComparisonOp(compiler, *locs(), kind(), labels); |
| } |
| } |
| |
| void NativeCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| SetupNative(); |
| Register result = locs()->out(0).reg(); |
| const intptr_t argc_tag = NativeArguments::ComputeArgcTag(function()); |
| |
| // All arguments are already @RSP due to preceding PushArgument()s. |
| ASSERT(ArgumentCount() == |
| function().NumParameters() + (function().IsGeneric() ? 1 : 0)); |
| |
| // Push the result place holder initialized to NULL. |
| __ PushObject(Object::null_object()); |
| |
| // Pass a pointer to the first argument in R13 (we avoid using RAX here to |
| // simplify the stub code that will call native code). |
| __ leaq(R13, compiler::Address(RSP, ArgumentCount() * kWordSize)); |
| |
| __ LoadImmediate(R10, compiler::Immediate(argc_tag)); |
| const Code* stub; |
| if (link_lazily()) { |
| stub = &StubCode::CallBootstrapNative(); |
| compiler::ExternalLabel label(NativeEntry::LinkNativeCallEntry()); |
| __ LoadNativeEntry(RBX, &label, |
| compiler::ObjectPoolBuilderEntry::kPatchable); |
| compiler->GeneratePatchableCall(source(), *stub, |
| UntaggedPcDescriptors::kOther, locs()); |
| } else { |
| if (is_bootstrap_native()) { |
| stub = &StubCode::CallBootstrapNative(); |
| } else if (is_auto_scope()) { |
| stub = &StubCode::CallAutoScopeNative(); |
| } else { |
| stub = &StubCode::CallNoScopeNative(); |
| } |
| const compiler::ExternalLabel label( |
| reinterpret_cast<uword>(native_c_function())); |
| __ LoadNativeEntry(RBX, &label, |
| compiler::ObjectPoolBuilderEntry::kNotPatchable); |
| // We can never lazy-deopt here because natives are never optimized. |
| ASSERT(!compiler->is_optimizing()); |
| compiler->GenerateNonLazyDeoptableStubCall( |
| source(), *stub, UntaggedPcDescriptors::kOther, locs()); |
| } |
| __ popq(result); |
| |
| __ Drop(ArgumentCount()); // Drop the arguments. |
| } |
| |
| #define R(r) (1 << r) |
| |
| LocationSummary* FfiCallInstr::MakeLocationSummary(Zone* zone, |
| bool is_optimizing) const { |
| // Use R10 as a temp. register. We can't use RDI, RSI, RDX, R8, R9 as they are |
| // argument registers, and R11 is TMP. |
| return MakeLocationSummaryInternal( |
| zone, is_optimizing, |
| (R(CallingConventions::kSecondNonArgumentRegister) | R(R10) | |
| R(CallingConventions::kFfiAnyNonAbiRegister))); |
| } |
| |
| #undef R |
| |
| void FfiCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| const Register target_address = locs()->in(TargetAddressIndex()).reg(); |
| |
| // The temps are indexed according to their register number. |
| // For regular calls, this holds the FP for rebasing the original locations |
| // during EmitParamMoves. |
| const Register saved_fp = locs()->temp(0).reg(); |
| const Register temp = locs()->temp(1).reg(); |
| // For leaf calls, this holds the SP used to restore the pre-aligned SP after |
| // the call. |
| // Note: R12 doubles as CODE_REG, which gets clobbered during frame setup in |
| // regular calls. |
| const Register saved_sp = locs()->temp(2).reg(); |
| |
| // Ensure these are callee-saved register and are preserved across the call. |
| ASSERT(IsCalleeSavedRegister(saved_sp)); |
| ASSERT(IsCalleeSavedRegister(saved_fp)); |
| // Other temps don't need to be preserved. |
| |
| if (is_leaf_) { |
| __ movq(saved_sp, SPREG); |
| } else { |
| __ movq(saved_fp, FPREG); |
| // Make a space to put the return address. |
| __ pushq(compiler::Immediate(0)); |
| |
| // We need to create a dummy "exit frame". It will share the same pool |
| // pointer but have a null code object. |
| __ LoadObject(CODE_REG, Code::null_object()); |
| __ set_constant_pool_allowed(false); |
| __ EnterDartFrame(0, PP); |
| } |
| |
| // Reserve space for the arguments that go on the stack (if any), then align. |
| __ ReserveAlignedFrameSpace(marshaller_.RequiredStackSpaceInBytes()); |
| |
| if (is_leaf_) { |
| EmitParamMoves(compiler, FPREG, saved_fp, TMP); |
| } else { |
| EmitParamMoves(compiler, saved_fp, saved_sp, TMP); |
| } |
| |
| if (compiler::Assembler::EmittingComments()) { |
| __ Comment(is_leaf_ ? "Leaf Call" : "Call"); |
| } |
| |
| if (is_leaf_) { |
| #if !defined(PRODUCT) |
| // Set the thread object's top_exit_frame_info and VMTag to enable the |
| // profiler to determine that thread is no longer executing Dart code. |
| __ movq(compiler::Address( |
| THR, compiler::target::Thread::top_exit_frame_info_offset()), |
| FPREG); |
| __ movq(compiler::Assembler::VMTagAddress(), target_address); |
| #endif |
| |
| __ CallCFunction(target_address, /*restore_rsp=*/true); |
| |
| #if !defined(PRODUCT) |
| __ movq(compiler::Assembler::VMTagAddress(), |
| compiler::Immediate(compiler::target::Thread::vm_tag_dart_id())); |
| __ movq(compiler::Address( |
| THR, compiler::target::Thread::top_exit_frame_info_offset()), |
| compiler::Immediate(0)); |
| #endif |
| } else { |
| // We need to copy a dummy return address up into the dummy stack frame so |
| // the stack walker will know which safepoint to use. RIP points to the |
| // *next* instruction, so 'AddressRIPRelative' loads the address of the |
| // following 'movq'. |
| __ leaq(temp, compiler::Address::AddressRIPRelative(0)); |
| compiler->EmitCallsiteMetadata(InstructionSource(), deopt_id(), |
| UntaggedPcDescriptors::Kind::kOther, locs(), |
| env()); |
| __ movq(compiler::Address(FPREG, kSavedCallerPcSlotFromFp * kWordSize), |
| temp); |
| |
| if (CanExecuteGeneratedCodeInSafepoint()) { |
| // Update information in the thread object and enter a safepoint. |
| __ movq(temp, compiler::Immediate( |
| compiler::target::Thread::exit_through_ffi())); |
| |
| __ TransitionGeneratedToNative(target_address, FPREG, temp, |
| /*enter_safepoint=*/true); |
| |
| __ CallCFunction(target_address, /*restore_rsp=*/true); |
| |
| // Update information in the thread object and leave the safepoint. |
| __ TransitionNativeToGenerated(/*leave_safepoint=*/true); |
| } else { |
| // We cannot trust that this code will be executable within a safepoint. |
| // Therefore we delegate the responsibility of entering/exiting the |
| // safepoint to a stub which is in the VM isolate's heap, which will never |
| // lose execute permission. |
| __ movq(temp, |
| compiler::Address( |
| THR, compiler::target::Thread:: |
| call_native_through_safepoint_entry_point_offset())); |
| |
| // Calls RBX within a safepoint. RBX and R12 are clobbered. |
| __ movq(RBX, target_address); |
| __ call(temp); |
| } |
| } |
| |
| // Pass the `saved_fp` reg. as a temp to clobber since we're done with it. |
| EmitReturnMoves(compiler, temp, saved_fp); |
| |
| if (is_leaf_) { |
| // Restore the pre-aligned SP. |
| __ movq(SPREG, saved_sp); |
| } else { |
| __ LeaveDartFrame(); |
| // Restore the global object pool after returning from runtime (old space is |
| // moving, so the GOP could have been relocated). |
| if (FLAG_precompiled_mode) { |
| __ movq(PP, compiler::Address(THR, Thread::global_object_pool_offset())); |
| } |
| __ set_constant_pool_allowed(true); |
| |
| // Instead of returning to the "fake" return address, we just pop it. |
| __ popq(temp); |
| } |
| } |
| |
| // Keep in sync with NativeReturnInstr::EmitNativeCode. |
| void NativeEntryInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| __ Bind(compiler->GetJumpLabel(this)); |
| |
| // Create a dummy frame holding the pushed arguments. This simplifies |
| // NativeReturnInstr::EmitNativeCode. |
| __ EnterFrame(0); |
| |
| #if defined(DART_TARGET_OS_FUCHSIA) && defined(USING_SHADOW_CALL_STACK) |
| #error Unimplemented |
| #endif |
| |
| // Save the argument registers, in reverse order. |
| SaveArguments(compiler); |
| |
| // Enter the entry frame. Push a dummy return address for consistency with |
| // EnterFrame on ARM(64). NativeParameterInstr expects this frame has size |
| // -exit_link_slot_from_entry_fp, verified below. |
| __ PushImmediate(compiler::Immediate(0)); |
| __ EnterFrame(0); |
| |
| // Save a space for the code object. |
| __ PushImmediate(compiler::Immediate(0)); |
| |
| // InvokeDartCodeStub saves the arguments descriptor here. We don't have one, |
| // but we need to follow the same frame layout for the stack walker. |
| __ PushImmediate(compiler::Immediate(0)); |
| |
| // Save ABI callee-saved registers. |
| __ PushRegisters(kCalleeSaveRegistersSet); |
| |
| // Load the address of DLRT_GetThreadForNativeCallback without using Thread. |
| if (FLAG_precompiled_mode) { |
| compiler->LoadBSSEntry(BSS::Relocation::DRT_GetThreadForNativeCallback, RAX, |
| RCX); |
| } else if (!NativeCallbackTrampolines::Enabled()) { |
| // In JIT mode, we can just paste the address of the runtime entry into the |
| // generated code directly. This is not a problem since we don't save |
| // callbacks into JIT snapshots. |
| __ movq(RAX, compiler::Immediate(reinterpret_cast<intptr_t>( |
| DLRT_GetThreadForNativeCallback))); |
| } |
| |
| // Create another frame to align the frame before continuing in "native" code. |
| // If we were called by a trampoline, it has already loaded the thread. |
| if (!NativeCallbackTrampolines::Enabled()) { |
| __ EnterFrame(0); |
| __ ReserveAlignedFrameSpace(0); |
| |
| COMPILE_ASSERT(RAX != CallingConventions::kArg1Reg); |
| __ movq(CallingConventions::kArg1Reg, compiler::Immediate(callback_id_)); |
| __ CallCFunction(RAX); |
| __ movq(THR, RAX); |
| |
| __ LeaveFrame(); |
| } |
| |
| // Save the current VMTag on the stack. |
| __ movq(RAX, compiler::Assembler::VMTagAddress()); |
| __ pushq(RAX); |
| |
| // Save top resource. |
| __ pushq( |
| compiler::Address(THR, compiler::target::Thread::top_resource_offset())); |
| __ movq( |
| compiler::Address(THR, compiler::target::Thread::top_resource_offset()), |
| compiler::Immediate(0)); |
| |
| __ pushq(compiler::Address( |
| THR, compiler::target::Thread::exit_through_ffi_offset())); |
| |
| // Save top exit frame info. Stack walker expects it to be here. |
| __ pushq(compiler::Address( |
| THR, compiler::target::Thread::top_exit_frame_info_offset())); |
| |
| // In debug mode, verify that we've pushed the top exit frame info at the |
| // correct offset from FP. |
| __ EmitEntryFrameVerification(); |
| |
| // Either DLRT_GetThreadForNativeCallback or the callback trampoline (caller) |
| // will leave the safepoint for us. |
| __ TransitionNativeToGenerated(/*exit_safepoint=*/false); |
| |
| // Load the code object. |
| __ movq(RAX, compiler::Address( |
| THR, compiler::target::Thread::callback_code_offset())); |
| __ LoadCompressed( |
| RAX, compiler::FieldAddress( |
| RAX, compiler::target::GrowableObjectArray::data_offset())); |
| __ LoadCompressed( |
| CODE_REG, |
| compiler::FieldAddress( |
| RAX, compiler::target::Array::data_offset() + |
| callback_id_ * compiler::target::kCompressedWordSize)); |
| |
| // Put the code object in the reserved slot. |
| __ movq(compiler::Address(FPREG, |
| kPcMarkerSlotFromFp * compiler::target::kWordSize), |
| CODE_REG); |
| |
| if (FLAG_precompiled_mode) { |
| __ movq(PP, |
| compiler::Address( |
| THR, compiler::target::Thread::global_object_pool_offset())); |
| } else { |
| __ xorq(PP, PP); // GC-safe value into PP. |
| } |
| |
| // Load a GC-safe value for arguments descriptor (unused but tagged). |
| __ xorq(ARGS_DESC_REG, ARGS_DESC_REG); |
| |
| // Push a dummy return address which suggests that we are inside of |
| // InvokeDartCodeStub. This is how the stack walker detects an entry frame. |
| __ movq(RAX, |
| compiler::Address( |
| THR, compiler::target::Thread::invoke_dart_code_stub_offset())); |
| __ pushq(compiler::FieldAddress( |
| RAX, compiler::target::Code::entry_point_offset())); |
| |
| // Continue with Dart frame setup. |
| FunctionEntryInstr::EmitNativeCode(compiler); |
| } |
| |
| #define R(r) (1 << r) |
| |
| LocationSummary* CCallInstr::MakeLocationSummary(Zone* zone, |
| bool is_optimizing) const { |
| constexpr Register saved_fp = CallingConventions::kSecondNonArgumentRegister; |
| return MakeLocationSummaryInternal(zone, (R(saved_fp))); |
| } |
| |
| #undef R |
| |
| void CCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| const Register saved_fp = locs()->temp(0).reg(); |
| const Register temp0 = TMP; |
| |
| // TODO(http://dartbug.com/47778): If we knew whether the stack was aligned |
| // at this point, we could omit having a frame. |
| __ MoveRegister(saved_fp, FPREG); |
| |
| const intptr_t frame_space = native_calling_convention_.StackTopInBytes(); |
| __ EnterCFrame(frame_space); |
| |
| EmitParamMoves(compiler, saved_fp, temp0); |
| const Register target_address = locs()->in(TargetAddressIndex()).reg(); |
| __ CallCFunction(target_address); |
| |
| __ LeaveCFrame(); |
| } |
| |
| static bool CanBeImmediateIndex(Value* index, intptr_t cid) { |
| if (!index->definition()->IsConstant()) return false; |
| const Object& constant = index->definition()->AsConstant()->value(); |
| if (!constant.IsSmi()) return false; |
| const Smi& smi_const = Smi::Cast(constant); |
| const intptr_t scale = Instance::ElementSizeFor(cid); |
| const intptr_t data_offset = Instance::DataOffsetFor(cid); |
| const int64_t disp = smi_const.AsInt64Value() * scale + data_offset; |
| return Utils::IsInt(32, disp); |
| } |
| |
| LocationSummary* OneByteStringFromCharCodeInstr::MakeLocationSummary( |
| Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| // TODO(fschneider): Allow immediate operands for the char code. |
| return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(), |
| LocationSummary::kNoCall); |
| } |
| |
| void OneByteStringFromCharCodeInstr::EmitNativeCode( |
| FlowGraphCompiler* compiler) { |
| ASSERT(compiler->is_optimizing()); |
| Register char_code = locs()->in(0).reg(); |
| Register result = locs()->out(0).reg(); |
| |
| #if defined(DART_COMPRESSED_POINTERS) |
| // The upper half of a compressed Smi contains undefined bits, but no x64 |
| // addressing mode will ignore these bits. Assume that the index is |
| // non-negative and clear the upper bits, which is shorter than |
| // sign-extension (movsxd). Note: we don't bother to ensure index is a |
| // writable input because any other instructions using it must also not |
| // rely on the upper bits. |
| __ orl(char_code, char_code); |
| #endif |
| __ movq(result, |
| compiler::Address(THR, Thread::predefined_symbols_address_offset())); |
| __ movq(result, |
| compiler::Address(result, char_code, |
| TIMES_HALF_WORD_SIZE, // Char code is a smi. |
| Symbols::kNullCharCodeSymbolOffset * kWordSize)); |
| } |
| |
| LocationSummary* StringToCharCodeInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(), |
| LocationSummary::kNoCall); |
| } |
| |
| void StringToCharCodeInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(cid_ == kOneByteStringCid); |
| Register str = locs()->in(0).reg(); |
| Register result = locs()->out(0).reg(); |
| compiler::Label is_one, done; |
| __ LoadCompressedSmi(result, |
| compiler::FieldAddress(str, String::length_offset())); |
| __ cmpq(result, compiler::Immediate(Smi::RawValue(1))); |
| __ j(EQUAL, &is_one, compiler::Assembler::kNearJump); |
| __ movq(result, compiler::Immediate(Smi::RawValue(-1))); |
| __ jmp(&done); |
| __ Bind(&is_one); |
| __ movzxb(result, compiler::FieldAddress(str, OneByteString::data_offset())); |
| __ SmiTag(result); |
| __ Bind(&done); |
| } |
| |
| LocationSummary* Utf8ScanInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 5; |
| const intptr_t kNumTemps = 1; |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::Any()); // decoder |
| summary->set_in(1, Location::WritableRegister()); // bytes |
| summary->set_in(2, Location::WritableRegister()); // start |
| summary->set_in(3, Location::WritableRegister()); // end |
| summary->set_in(4, Location::RequiresRegister()); // table |
| summary->set_temp(0, Location::RequiresRegister()); |
| summary->set_out(0, Location::RequiresRegister()); |
| return summary; |
| } |
| |
| void Utf8ScanInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| const Register bytes_reg = locs()->in(1).reg(); |
| const Register start_reg = locs()->in(2).reg(); |
| const Register end_reg = locs()->in(3).reg(); |
| const Register table_reg = locs()->in(4).reg(); |
| const Register size_reg = locs()->out(0).reg(); |
| |
| const Register bytes_ptr_reg = start_reg; |
| const Register bytes_end_reg = end_reg; |
| const Register bytes_end_minus_16_reg = bytes_reg; |
| const Register flags_reg = locs()->temp(0).reg(); |
| const Register temp_reg = TMP; |
| const XmmRegister vector_reg = FpuTMP; |
| |
| static const intptr_t kSizeMask = 0x03; |
| static const intptr_t kFlagsMask = 0x3C; |
| |
| compiler::Label scan_ascii, ascii_loop, ascii_loop_in, nonascii_loop; |
| compiler::Label rest, rest_loop, rest_loop_in, done; |
| |
| // Address of input bytes. |
| __ movq(bytes_reg, |
| compiler::FieldAddress(bytes_reg, |
| compiler::target::PointerBase::data_offset())); |
| |
| // Pointers to start, end and end-16. |
| __ leaq(bytes_ptr_reg, compiler::Address(bytes_reg, start_reg, TIMES_1, 0)); |
| __ leaq(bytes_end_reg, compiler::Address(bytes_reg, end_reg, TIMES_1, 0)); |
| __ leaq(bytes_end_minus_16_reg, compiler::Address(bytes_end_reg, -16)); |
| |
| // Initialize size and flags. |
| __ xorq(size_reg, size_reg); |
| __ xorq(flags_reg, flags_reg); |
| |
| __ jmp(&scan_ascii, compiler::Assembler::kNearJump); |
| |
| // Loop scanning through ASCII bytes one 16-byte vector at a time. |
| // While scanning, the size register contains the size as it was at the start |
| // of the current block of ASCII bytes, minus the address of the start of the |
| // block. After the block, the end address of the block is added to update the |
| // size to include the bytes in the block. |
| __ Bind(&ascii_loop); |
| __ addq(bytes_ptr_reg, compiler::Immediate(16)); |
| __ Bind(&ascii_loop_in); |
| |
| // Exit vectorized loop when there are less than 16 bytes left. |
| __ cmpq(bytes_ptr_reg, bytes_end_minus_16_reg); |
| __ j(UNSIGNED_GREATER, &rest, compiler::Assembler::kNearJump); |
| |
| // Find next non-ASCII byte within the next 16 bytes. |
| // Note: In principle, we should use MOVDQU here, since the loaded value is |
| // used as input to an integer instruction. In practice, according to Agner |
| // Fog, there is no penalty for using the wrong kind of load. |
| __ movups(vector_reg, compiler::Address(bytes_ptr_reg, 0)); |
| __ pmovmskb(temp_reg, vector_reg); |
| __ bsfq(temp_reg, temp_reg); |
| __ j(EQUAL, &ascii_loop, compiler::Assembler::kNearJump); |
| |
| // Point to non-ASCII byte and update size. |
| __ addq(bytes_ptr_reg, temp_reg); |
| __ addq(size_reg, bytes_ptr_reg); |
| |
| // Read first non-ASCII byte. |
| __ movzxb(temp_reg, compiler::Address(bytes_ptr_reg, 0)); |
| |
| // Loop over block of non-ASCII bytes. |
| __ Bind(&nonascii_loop); |
| __ addq(bytes_ptr_reg, compiler::Immediate(1)); |
| |
| // Update size and flags based on byte value. |
| __ movzxb(temp_reg, compiler::FieldAddress( |
| table_reg, temp_reg, TIMES_1, |
| compiler::target::OneByteString::data_offset())); |
| __ orq(flags_reg, temp_reg); |
| __ andq(temp_reg, compiler::Immediate(kSizeMask)); |
| __ addq(size_reg, temp_reg); |
| |
| // Stop if end is reached. |
| __ cmpq(bytes_ptr_reg, bytes_end_reg); |
| __ j(UNSIGNED_GREATER_EQUAL, &done, compiler::Assembler::kNearJump); |
| |
| // Go to ASCII scan if next byte is ASCII, otherwise loop. |
| __ movzxb(temp_reg, compiler::Address(bytes_ptr_reg, 0)); |
| __ testq(temp_reg, compiler::Immediate(0x80)); |
| __ j(NOT_EQUAL, &nonascii_loop, compiler::Assembler::kNearJump); |
| |
| // Enter the ASCII scanning loop. |
| __ Bind(&scan_ascii); |
| __ subq(size_reg, bytes_ptr_reg); |
| __ jmp(&ascii_loop_in); |
| |
| // Less than 16 bytes left. Process the remaining bytes individually. |
| __ Bind(&rest); |
| |
| // Update size after ASCII scanning loop. |
| __ addq(size_reg, bytes_ptr_reg); |
| __ jmp(&rest_loop_in, compiler::Assembler::kNearJump); |
| |
| __ Bind(&rest_loop); |
| |
| // Read byte and increment pointer. |
| __ movzxb(temp_reg, compiler::Address(bytes_ptr_reg, 0)); |
| __ addq(bytes_ptr_reg, compiler::Immediate(1)); |
| |
| // Update size and flags based on byte value. |
| __ movzxb(temp_reg, compiler::FieldAddress( |
| table_reg, temp_reg, TIMES_1, |
| compiler::target::OneByteString::data_offset())); |
| __ orq(flags_reg, temp_reg); |
| __ andq(temp_reg, compiler::Immediate(kSizeMask)); |
| __ addq(size_reg, temp_reg); |
| |
| // Stop if end is reached. |
| __ Bind(&rest_loop_in); |
| __ cmpq(bytes_ptr_reg, bytes_end_reg); |
| __ j(UNSIGNED_LESS, &rest_loop, compiler::Assembler::kNearJump); |
| __ Bind(&done); |
| |
| // Write flags to field. |
| __ andq(flags_reg, compiler::Immediate(kFlagsMask)); |
| if (!IsScanFlagsUnboxed()) { |
| __ SmiTag(flags_reg); |
| } |
| Register decoder_reg; |
| const Location decoder_location = locs()->in(0); |
| if (decoder_location.IsStackSlot()) { |
| __ movq(temp_reg, LocationToStackSlotAddress(decoder_location)); |
| decoder_reg = temp_reg; |
| } else { |
| decoder_reg = decoder_location.reg(); |
| } |
| const auto scan_flags_field_offset = scan_flags_field_.offset_in_bytes(); |
| if (scan_flags_field_.is_compressed() && !IsScanFlagsUnboxed()) { |
| __ OBJ(or)(compiler::FieldAddress(decoder_reg, scan_flags_field_offset), |
| flags_reg); |
| } else { |
| __ orq(compiler::FieldAddress(decoder_reg, scan_flags_field_offset), |
| flags_reg); |
| } |
| } |
| |
| LocationSummary* LoadUntaggedInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(), |
| LocationSummary::kNoCall); |
| } |
| |
| void LoadUntaggedInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| Register obj = locs()->in(0).reg(); |
| Register result = locs()->out(0).reg(); |
| if (object()->definition()->representation() == kUntagged) { |
| __ movq(result, compiler::Address(obj, offset())); |
| } else { |
| ASSERT(object()->definition()->representation() == kTagged); |
| __ movq(result, compiler::FieldAddress(obj, offset())); |
| } |
| } |
| |
| LocationSummary* LoadIndexedInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 2; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RequiresRegister()); |
| // For tagged index with index_scale=1 as well as untagged index with |
| // index_scale=16 we need a writable register due to addressing mode |
| // restrictions on X64. |
| const bool need_writable_index_register = |
| (index_scale() == 1 && !index_unboxed_) || |
| (index_scale() == 16 && index_unboxed_); |
| locs->set_in( |
| 1, CanBeImmediateIndex(index(), class_id()) |
| ? Location::Constant(index()->definition()->AsConstant()) |
| : (need_writable_index_register ? Location::WritableRegister() |
| : Location::RequiresRegister())); |
| if ((representation() == kUnboxedDouble) || |
| (representation() == kUnboxedFloat32x4) || |
| (representation() == kUnboxedInt32x4) || |
| (representation() == kUnboxedFloat64x2)) { |
| locs->set_out(0, Location::RequiresFpuRegister()); |
| } else { |
| locs->set_out(0, Location::RequiresRegister()); |
| } |
| return locs; |
| } |
| |
| void LoadIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // The array register points to the backing store for external arrays. |
| const Register array = locs()->in(0).reg(); |
| const Location index = locs()->in(1); |
| |
| intptr_t index_scale = index_scale_; |
| if (index.IsRegister()) { |
| if (index_scale == 1 && !index_unboxed_) { |
| __ SmiUntag(index.reg()); |
| } else if (index_scale == 16 && index_unboxed_) { |
| // X64 does not support addressing mode using TIMES_16. |
| __ SmiTag(index.reg()); |
| index_scale >>= 1; |
| } else if (!index_unboxed_) { |
| #if defined(DART_COMPRESSED_POINTERS) |
| // The upper half of a compressed Smi contains undefined bits, but no x64 |
| // addressing mode will ignore these bits. Assume that the index is |
| // non-negative and clear the upper bits, which is shorter than |
| // sign-extension (movsxd). Note: we don't bother to ensure index is a |
| // writable input because any other instructions using it must also not |
| // rely on the upper bits. |
| __ orl(index.reg(), index.reg()); |
| #endif |
| } |
| } else { |
| ASSERT(index.IsConstant()); |
| } |
| |
| compiler::Address element_address = |
| index.IsRegister() ? compiler::Assembler::ElementAddressForRegIndex( |
| IsExternal(), class_id(), index_scale, |
| index_unboxed_, array, index.reg()) |
| : compiler::Assembler::ElementAddressForIntIndex( |
| IsExternal(), class_id(), index_scale, array, |
| Smi::Cast(index.constant()).Value()); |
| |
| if (representation() == kUnboxedDouble || |
| representation() == kUnboxedFloat32x4 || |
| representation() == kUnboxedInt32x4 || |
| representation() == kUnboxedFloat64x2) { |
| XmmRegister result = locs()->out(0).fpu_reg(); |
| if (class_id() == kTypedDataFloat32ArrayCid) { |
| // Load single precision float. |
| __ movss(result, element_address); |
| } else if (class_id() == kTypedDataFloat64ArrayCid) { |
| __ movsd(result, element_address); |
| } else { |
| ASSERT((class_id() == kTypedDataInt32x4ArrayCid) || |
| (class_id() == kTypedDataFloat32x4ArrayCid) || |
| (class_id() == kTypedDataFloat64x2ArrayCid)); |
| __ movups(result, element_address); |
| } |
| return; |
| } |
| |
| Register result = locs()->out(0).reg(); |
| switch (class_id()) { |
| case kTypedDataInt32ArrayCid: |
| ASSERT(representation() == kUnboxedInt32); |
| __ movsxd(result, element_address); |
| break; |
| case kTypedDataUint32ArrayCid: |
| ASSERT(representation() == kUnboxedUint32); |
| __ movl(result, element_address); |
| break; |
| case kTypedDataInt64ArrayCid: |
| case kTypedDataUint64ArrayCid: |
| ASSERT(representation() == kUnboxedInt64); |
| __ movq(result, element_address); |
| break; |
| case kTypedDataInt8ArrayCid: |
| ASSERT(representation() == kUnboxedIntPtr); |
| __ movsxb(result, element_address); |
| break; |
| case kTypedDataUint8ArrayCid: |
| case kTypedDataUint8ClampedArrayCid: |
| case kExternalTypedDataUint8ArrayCid: |
| case kExternalTypedDataUint8ClampedArrayCid: |
| case kOneByteStringCid: |
| case kExternalOneByteStringCid: |
| ASSERT(representation() == kUnboxedIntPtr); |
| __ movzxb(result, element_address); |
| break; |
| case kTypedDataInt16ArrayCid: |
| ASSERT(representation() == kUnboxedIntPtr); |
| __ movsxw(result, element_address); |
| break; |
| case kTypedDataUint16ArrayCid: |
| case kTwoByteStringCid: |
| case kExternalTwoByteStringCid: |
| ASSERT(representation() == kUnboxedIntPtr); |
| __ movzxw(result, element_address); |
| break; |
| default: |
| ASSERT(representation() == kTagged); |
| ASSERT((class_id() == kArrayCid) || (class_id() == kImmutableArrayCid) || |
| (class_id() == kTypeArgumentsCid)); |
| __ LoadCompressed(result, element_address); |
| break; |
| } |
| } |
| |
| LocationSummary* LoadCodeUnitsInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 2; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresRegister()); |
| // The smi index is either untagged (element size == 1), or it is left smi |
| // tagged (for all element sizes > 1). |
| summary->set_in(1, index_scale() == 1 ? Location::WritableRegister() |
| : Location::RequiresRegister()); |
| summary->set_out(0, Location::RequiresRegister()); |
| return summary; |
| } |
| |
| void LoadCodeUnitsInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // The string register points to the backing store for external strings. |
| const Register str = locs()->in(0).reg(); |
| const Location index = locs()->in(1); |
| |
| compiler::Address element_address = |
| compiler::Assembler::ElementAddressForRegIndex( |
| IsExternal(), class_id(), index_scale(), /*index_unboxed=*/false, str, |
| index.reg()); |
| |
| if ((index_scale() == 1)) { |
| __ SmiUntag(index.reg()); |
| } else { |
| #if defined(DART_COMPRESSED_POINTERS) |
| // The upper half of a compressed Smi contains undefined bits, but no x64 |
| // addressing mode will ignore these bits. Assume that the index is |
| // non-negative and clear the upper bits, which is shorter than |
| // sign-extension (movsxd). Note: we don't bother to ensure index is a |
| // writable input because any other instructions using it must also not |
| // rely on the upper bits. |
| __ orl(index.reg(), index.reg()); |
| #endif |
| } |
| Register result = locs()->out(0).reg(); |
| switch (class_id()) { |
| case kOneByteStringCid: |
| case kExternalOneByteStringCid: |
| switch (element_count()) { |
| case 1: |
| __ movzxb(result, element_address); |
| break; |
| case 2: |
| __ movzxw(result, element_address); |
| break; |
| case 4: |
| __ movl(result, element_address); |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| ASSERT(can_pack_into_smi()); |
| __ SmiTag(result); |
| break; |
| case kTwoByteStringCid: |
| case kExternalTwoByteStringCid: |
| switch (element_count()) { |
| case 1: |
| __ movzxw(result, element_address); |
| break; |
| case 2: |
| __ movl(result, element_address); |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| ASSERT(can_pack_into_smi()); |
| __ SmiTag(result); |
| break; |
| default: |
| UNREACHABLE(); |
| break; |
| } |
| } |
| |
| LocationSummary* StoreIndexedInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 3; |
| const intptr_t kNumTemps = |
| class_id() == kArrayCid && ShouldEmitStoreBarrier() ? 1 : 0; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RequiresRegister()); |
| // For tagged index with index_scale=1 as well as untagged index with |
| // index_scale=16 we need a writable register due to addressing mode |
| // restrictions on X64. |
| const bool need_writable_index_register = |
| (index_scale() == 1 && !index_unboxed_) || |
| (index_scale() == 16 && index_unboxed_); |
| locs->set_in( |
| 1, CanBeImmediateIndex(index(), class_id()) |
| ? Location::Constant(index()->definition()->AsConstant()) |
| : (need_writable_index_register ? Location::WritableRegister() |
| : Location::RequiresRegister())); |
| switch (class_id()) { |
| case kArrayCid: |
| locs->set_in(2, ShouldEmitStoreBarrier() |
| ? Location::RegisterLocation(kWriteBarrierValueReg) |
| : LocationRegisterOrConstant(value())); |
| if (ShouldEmitStoreBarrier()) { |
| locs->set_in(0, Location::RegisterLocation(kWriteBarrierObjectReg)); |
| locs->set_temp(0, Location::RegisterLocation(kWriteBarrierSlotReg)); |
| } |
| break; |
| case kExternalTypedDataUint8ArrayCid: |
| case kExternalTypedDataUint8ClampedArrayCid: |
| case kTypedDataInt8ArrayCid: |
| case kTypedDataUint8ArrayCid: |
| case kTypedDataUint8ClampedArrayCid: |
| case kOneByteStringCid: |
| case kTwoByteStringCid: |
| // TODO(fschneider): Add location constraint for byte registers (RAX, |
| // RBX, RCX, RDX) instead of using a fixed register. |
| locs->set_in(2, LocationFixedRegisterOrSmiConstant(value(), RAX)); |
| break; |
| case kTypedDataInt16ArrayCid: |
| case kTypedDataUint16ArrayCid: |
| // Writable register because the value must be untagged before storing. |
| locs->set_in(2, Location::WritableRegister()); |
| break; |
| case kTypedDataInt32ArrayCid: |
| case kTypedDataUint32ArrayCid: |
| case kTypedDataInt64ArrayCid: |
| case kTypedDataUint64ArrayCid: |
| locs->set_in(2, Location::RequiresRegister()); |
| break; |
| case kTypedDataFloat32ArrayCid: |
| case kTypedDataFloat64ArrayCid: |
| // TODO(srdjan): Support Float64 constants. |
| locs->set_in(2, Location::RequiresFpuRegister()); |
| break; |
| case kTypedDataInt32x4ArrayCid: |
| case kTypedDataFloat64x2ArrayCid: |
| case kTypedDataFloat32x4ArrayCid: |
| locs->set_in(2, Location::RequiresFpuRegister()); |
| break; |
| default: |
| UNREACHABLE(); |
| return NULL; |
| } |
| return locs; |
| } |
| |
| void StoreIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // The array register points to the backing store for external arrays. |
| const Register array = locs()->in(0).reg(); |
| const Location index = locs()->in(1); |
| |
| intptr_t index_scale = index_scale_; |
| if (index.IsRegister()) { |
| if (index_scale == 1 && !index_unboxed_) { |
| __ SmiUntag(index.reg()); |
| } else if (index_scale == 16 && index_unboxed_) { |
| // X64 does not support addressing mode using TIMES_16. |
| __ SmiTag(index.reg()); |
| index_scale >>= 1; |
| } else if (!index_unboxed_) { |
| #if defined(DART_COMPRESSED_POINTERS) |
| // The upper half of a compressed Smi contains undefined bits, but no x64 |
| // addressing mode will ignore these bits. Assume that the index is |
| // non-negative and clear the upper bits, which is shorter than |
| // sign-extension (movsxd). Note: we don't bother to ensure index is a |
| // writable input because any other instructions using it must also not |
| // rely on the upper bits. |
| __ orl(index.reg(), index.reg()); |
| #endif |
| } |
| } else { |
| ASSERT(index.IsConstant()); |
| } |
| |
| compiler::Address element_address = |
| index.IsRegister() ? compiler::Assembler::ElementAddressForRegIndex( |
| IsExternal(), class_id(), index_scale, |
| index_unboxed_, array, index.reg()) |
| : compiler::Assembler::ElementAddressForIntIndex( |
| IsExternal(), class_id(), index_scale, array, |
| Smi::Cast(index.constant()).Value()); |
| |
| switch (class_id()) { |
| case kArrayCid: |
| if (ShouldEmitStoreBarrier()) { |
| Register value = locs()->in(2).reg(); |
| Register slot = locs()->temp(0).reg(); |
| __ leaq(slot, element_address); |
| __ StoreCompressedIntoArray(array, slot, value, CanValueBeSmi()); |
| } else if (locs()->in(2).IsConstant()) { |
| const Object& constant = locs()->in(2).constant(); |
| __ StoreCompressedIntoObjectNoBarrier(array, element_address, constant); |
| } else { |
| Register value = locs()->in(2).reg(); |
| __ StoreCompressedIntoObjectNoBarrier(array, element_address, value); |
| } |
| break; |
| case kOneByteStringCid: |
| case kTypedDataInt8ArrayCid: |
| case kTypedDataUint8ArrayCid: |
| case kExternalTypedDataUint8ArrayCid: |
| ASSERT(RequiredInputRepresentation(2) == kUnboxedIntPtr); |
| if (locs()->in(2).IsConstant()) { |
| const Smi& constant = Smi::Cast(locs()->in(2).constant()); |
| __ movb(element_address, |
| compiler::Immediate(static_cast<int8_t>(constant.Value()))); |
| } else { |
| __ movb(element_address, ByteRegisterOf(locs()->in(2).reg())); |
| } |
| break; |
| case kTypedDataUint8ClampedArrayCid: |
| case kExternalTypedDataUint8ClampedArrayCid: { |
| ASSERT(RequiredInputRepresentation(2) == kUnboxedIntPtr); |
| if (locs()->in(2).IsConstant()) { |
| const Smi& constant = Smi::Cast(locs()->in(2).constant()); |
| intptr_t value = constant.Value(); |
| // Clamp to 0x0 or 0xFF respectively. |
| if (value > 0xFF) { |
| value = 0xFF; |
| } else if (value < 0) { |
| value = 0; |
| } |
| __ movb(element_address, |
| compiler::Immediate(static_cast<int8_t>(value))); |
| } else { |
| const Register storedValueReg = locs()->in(2).reg(); |
| compiler::Label store_value, store_0xff; |
| __ CompareImmediate(storedValueReg, compiler::Immediate(0xFF)); |
| __ j(BELOW_EQUAL, &store_value, compiler::Assembler::kNearJump); |
| // Clamp to 0x0 or 0xFF respectively. |
| __ j(GREATER, &store_0xff); |
| __ xorq(storedValueReg, storedValueReg); |
| __ jmp(&store_value, compiler::Assembler::kNearJump); |
| __ Bind(&store_0xff); |
| __ LoadImmediate(storedValueReg, compiler::Immediate(0xFF)); |
| __ Bind(&store_value); |
| __ movb(element_address, ByteRegisterOf(storedValueReg)); |
| } |
| break; |
| } |
| case kTwoByteStringCid: |
| case kTypedDataInt16ArrayCid: |
| case kTypedDataUint16ArrayCid: { |
| ASSERT(RequiredInputRepresentation(2) == kUnboxedIntPtr); |
| Register value = locs()->in(2).reg(); |
| __ movw(element_address, value); |
| break; |
| } |
| case kTypedDataInt32ArrayCid: |
| case kTypedDataUint32ArrayCid: { |
| Register value = locs()->in(2).reg(); |
| __ movl(element_address, value); |
| break; |
| } |
| case kTypedDataInt64ArrayCid: |
| case kTypedDataUint64ArrayCid: { |
| Register value = locs()->in(2).reg(); |
| __ movq(element_address, value); |
| break; |
| } |
| case kTypedDataFloat32ArrayCid: |
| __ movss(element_address, locs()->in(2).fpu_reg()); |
| break; |
| case kTypedDataFloat64ArrayCid: |
| __ movsd(element_address, locs()->in(2).fpu_reg()); |
| break; |
| case kTypedDataInt32x4ArrayCid: |
| case kTypedDataFloat64x2ArrayCid: |
| case kTypedDataFloat32x4ArrayCid: |
| __ movups(element_address, locs()->in(2).fpu_reg()); |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| } |
| |
| LocationSummary* GuardFieldClassInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| |
| const intptr_t value_cid = value()->Type()->ToCid(); |
| const intptr_t field_cid = field().guarded_cid(); |
| |
| const bool emit_full_guard = !opt || (field_cid == kIllegalCid); |
| const bool needs_value_cid_temp_reg = |
| (value_cid == kDynamicCid) && (emit_full_guard || (field_cid != kSmiCid)); |
| const bool needs_field_temp_reg = emit_full_guard; |
| |
| intptr_t num_temps = 0; |
| if (needs_value_cid_temp_reg) { |
| num_temps++; |
| } |
| if (needs_field_temp_reg) { |
| num_temps++; |
| } |
| |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, num_temps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresRegister()); |
| |
| for (intptr_t i = 0; i < num_temps; i++) { |
| summary->set_temp(i, Location::RequiresRegister()); |
| } |
| |
| return summary; |
| } |
| |
| void GuardFieldClassInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(compiler::target::UntaggedObject::kClassIdTagSize == 16); |
| ASSERT(sizeof(UntaggedField::guarded_cid_) == 2); |
| ASSERT(sizeof(UntaggedField::is_nullable_) == 2); |
| |
| const intptr_t value_cid = value()->Type()->ToCid(); |
| const intptr_t field_cid = field().guarded_cid(); |
| const intptr_t nullability = field().is_nullable() ? kNullCid : kIllegalCid; |
| |
| if (field_cid == kDynamicCid) { |
| return; // Nothing to emit. |
| } |
| |
| const bool emit_full_guard = |
| !compiler->is_optimizing() || (field_cid == kIllegalCid); |
| |
| const bool needs_value_cid_temp_reg = |
| (value_cid == kDynamicCid) && (emit_full_guard || (field_cid != kSmiCid)); |
| |
| const bool needs_field_temp_reg = emit_full_guard; |
| |
| const Register value_reg = locs()->in(0).reg(); |
| |
| const Register value_cid_reg = |
| needs_value_cid_temp_reg ? locs()->temp(0).reg() : kNoRegister; |
| |
| const Register field_reg = needs_field_temp_reg |
| ? locs()->temp(locs()->temp_count() - 1).reg() |
| : kNoRegister; |
| |
| compiler::Label ok, fail_label; |
| |
| compiler::Label* deopt = NULL; |
| if (compiler->is_optimizing()) { |
| deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptGuardField); |
| } |
| |
| compiler::Label* fail = (deopt != NULL) ? deopt : &fail_label; |
| |
| if (emit_full_guard) { |
| __ LoadObject(field_reg, Field::ZoneHandle(field().Original())); |
| |
| compiler::FieldAddress field_cid_operand(field_reg, |
| Field::guarded_cid_offset()); |
| compiler::FieldAddress field_nullability_operand( |
| field_reg, Field::is_nullable_offset()); |
| |
| if (value_cid == kDynamicCid) { |
| LoadValueCid(compiler, value_cid_reg, value_reg); |
| |
| __ cmpw(value_cid_reg, field_cid_operand); |
| __ j(EQUAL, &ok); |
| __ cmpw(value_cid_reg, field_nullability_operand); |
| } else if (value_cid == kNullCid) { |
| __ cmpw(field_nullability_operand, compiler::Immediate(value_cid)); |
| } else { |
| __ cmpw(field_cid_operand, compiler::Immediate(value_cid)); |
| } |
| __ j(EQUAL, &ok); |
| |
| // Check if the tracked state of the guarded field can be initialized |
| // inline. If the field needs length check or requires type arguments and |
| // class hierarchy processing for exactness tracking then we fall through |
| // into runtime which is responsible for computing offset of the length |
| // field based on the class id. |
| const bool is_complicated_field = |
| field().needs_length_check() || |
| field().static_type_exactness_state().IsUninitialized(); |
| if (!is_complicated_field) { |
| // Uninitialized field can be handled inline. Check if the |
| // field is still unitialized. |
| __ cmpw(field_cid_operand, compiler::Immediate(kIllegalCid)); |
| __ j(NOT_EQUAL, fail); |
| |
| if (value_cid == kDynamicCid) { |
| __ movw(field_cid_operand, value_cid_reg); |
| __ movw(field_nullability_operand, value_cid_reg); |
| } else { |
| ASSERT(field_reg != kNoRegister); |
| __ movw(field_cid_operand, compiler::Immediate(value_cid)); |
| __ movw(field_nullability_operand, compiler::Immediate(value_cid)); |
| } |
| |
| __ jmp(&ok); |
| } |
| |
| if (deopt == NULL) { |
| __ Bind(fail); |
| |
| __ cmpw(compiler::FieldAddress(field_reg, Field::guarded_cid_offset()), |
| compiler::Immediate(kDynamicCid)); |
| __ j(EQUAL, &ok); |
| |
| __ pushq(field_reg); |
| __ pushq(value_reg); |
| ASSERT(!compiler->is_optimizing()); // No deopt info needed. |
| __ CallRuntime(kUpdateFieldCidRuntimeEntry, 2); |
| __ Drop(2); // Drop the field and the value. |
| } else { |
| __ jmp(fail); |
| } |
| } else { |
| ASSERT(compiler->is_optimizing()); |
| ASSERT(deopt != NULL); |
| |
| // Field guard class has been initialized and is known. |
| if (value_cid == kDynamicCid) { |
| // Value's class id is not known. |
| __ testq(value_reg, compiler::Immediate(kSmiTagMask)); |
| |
| if (field_cid != kSmiCid) { |
| __ j(ZERO, fail); |
| __ LoadClassId(value_cid_reg, value_reg); |
| __ CompareImmediate(value_cid_reg, compiler::Immediate(field_cid)); |
| } |
| |
| if (field().is_nullable() && (field_cid != kNullCid)) { |
| __ j(EQUAL, &ok); |
| __ CompareObject(value_reg, Object::null_object()); |
| } |
| |
| __ j(NOT_EQUAL, fail); |
| } else if (value_cid == field_cid) { |
| // This would normaly be caught by Canonicalize, but RemoveRedefinitions |
| // may sometimes produce the situation after the last Canonicalize pass. |
| } else { |
| // Both value's and field's class id is known. |
| ASSERT(value_cid != nullability); |
| __ jmp(fail); |
| } |
| } |
| __ Bind(&ok); |
| } |
| |
| LocationSummary* GuardFieldLengthInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| if (!opt || (field().guarded_list_length() == Field::kUnknownFixedLength)) { |
| const intptr_t kNumTemps = 3; |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresRegister()); |
| // We need temporaries for field object, length offset and expected length. |
| summary->set_temp(0, Location::RequiresRegister()); |
| summary->set_temp(1, Location::RequiresRegister()); |
| summary->set_temp(2, Location::RequiresRegister()); |
| return summary; |
| } else { |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, 0, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresRegister()); |
| return summary; |
| } |
| UNREACHABLE(); |
| } |
| |
| void GuardFieldLengthInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| if (field().guarded_list_length() == Field::kNoFixedLength) { |
| return; // Nothing to emit. |
| } |
| |
| compiler::Label* deopt = |
| compiler->is_optimizing() |
| ? compiler->AddDeoptStub(deopt_id(), ICData::kDeoptGuardField) |
| : NULL; |
| |
| const Register value_reg = locs()->in(0).reg(); |
| |
| if (!compiler->is_optimizing() || |
| (field().guarded_list_length() == Field::kUnknownFixedLength)) { |
| const Register field_reg = locs()->temp(0).reg(); |
| const Register offset_reg = locs()->temp(1).reg(); |
| const Register length_reg = locs()->temp(2).reg(); |
| |
| compiler::Label ok; |
| |
| __ LoadObject(field_reg, Field::ZoneHandle(field().Original())); |
| |
| __ movsxb( |
| offset_reg, |
| compiler::FieldAddress( |
| field_reg, Field::guarded_list_length_in_object_offset_offset())); |
| __ LoadCompressedSmi( |
| length_reg, |
| compiler::FieldAddress(field_reg, Field::guarded_list_length_offset())); |
| |
| __ cmpq(offset_reg, compiler::Immediate(0)); |
| __ j(NEGATIVE, &ok); |
| |
| // Load the length from the value. GuardFieldClass already verified that |
| // value's class matches guarded class id of the field. |
| // offset_reg contains offset already corrected by -kHeapObjectTag that is |
| // why we use Address instead of FieldAddress. |
| __ OBJ(cmp)(length_reg, |
| compiler::Address(value_reg, offset_reg, TIMES_1, 0)); |
| |
| if (deopt == NULL) { |
| __ j(EQUAL, &ok); |
| |
| __ pushq(field_reg); |
| __ pushq(value_reg); |
| ASSERT(!compiler->is_optimizing()); // No deopt info needed. |
| __ CallRuntime(kUpdateFieldCidRuntimeEntry, 2); |
| __ Drop(2); // Drop the field and the value. |
| } else { |
| __ j(NOT_EQUAL, deopt); |
| } |
| |
| __ Bind(&ok); |
| } else { |
| ASSERT(compiler->is_optimizing()); |
| ASSERT(field().guarded_list_length() >= 0); |
| ASSERT(field().guarded_list_length_in_object_offset() != |
| Field::kUnknownLengthOffset); |
| |
| __ CompareImmediate( |
| compiler::FieldAddress(value_reg, |
| field().guarded_list_length_in_object_offset()), |
| compiler::Immediate(Smi::RawValue(field().guarded_list_length()))); |
| __ j(NOT_EQUAL, deopt); |
| } |
| } |
| |
| LocationSummary* GuardFieldTypeInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 1; |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| summary->set_in(0, Location::RequiresRegister()); |
| summary->set_temp(0, Location::RequiresRegister()); |
| return summary; |
| } |
| |
| void GuardFieldTypeInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| // Should never emit GuardFieldType for fields that are marked as NotTracking. |
| ASSERT(field().static_type_exactness_state().IsTracking()); |
| if (!field().static_type_exactness_state().NeedsFieldGuard()) { |
| // Nothing to do: we only need to perform checks for trivially invariant |
| // fields. If optimizing Canonicalize pass should have removed |
| // this instruction. |
| return; |
| } |
| |
| compiler::Label* deopt = |
| compiler->is_optimizing() |
| ? compiler->AddDeoptStub(deopt_id(), ICData::kDeoptGuardField) |
| : NULL; |
| |
| compiler::Label ok; |
| |
| const Register value_reg = locs()->in(0).reg(); |
| const Register temp = locs()->temp(0).reg(); |
| |
| // Skip null values for nullable fields. |
| if (!compiler->is_optimizing() || field().is_nullable()) { |
| __ CompareObject(value_reg, Object::Handle()); |
| __ j(EQUAL, &ok); |
| } |
| |
| // Get the state. |
| const Field& original = |
| Field::ZoneHandle(compiler->zone(), field().Original()); |
| __ LoadObject(temp, original); |
| __ movsxb(temp, compiler::FieldAddress( |
| temp, Field::static_type_exactness_state_offset())); |
| |
| if (!compiler->is_optimizing()) { |
| // Check if field requires checking (it is in unitialized or trivially |
| // exact state). |
| __ cmpq(temp, |
| compiler::Immediate(StaticTypeExactnessState::kUninitialized)); |
| __ j(LESS, &ok); |
| } |
| |
| compiler::Label call_runtime; |
| if (field().static_type_exactness_state().IsUninitialized()) { |
| // Can't initialize the field state inline in optimized code. |
| __ cmpq(temp, |
| compiler::Immediate(StaticTypeExactnessState::kUninitialized)); |
| __ j(EQUAL, compiler->is_optimizing() ? deopt : &call_runtime); |
| } |
| |
| // At this point temp is known to be type arguments offset in words. |
| __ movq(temp, compiler::FieldAddress(value_reg, temp, |
| TIMES_COMPRESSED_WORD_SIZE, 0)); |
| __ CompareObject(temp, TypeArguments::ZoneHandle( |
| compiler->zone(), |
| AbstractType::Handle(field().type()).arguments())); |
| if (deopt != nullptr) { |
| __ j(NOT_EQUAL, deopt); |
| } else { |
| __ j(EQUAL, &ok); |
| |
| __ Bind(&call_runtime); |
| __ PushObject(original); |
| __ pushq(value_reg); |
| ASSERT(!compiler->is_optimizing()); // No deopt info needed. |
| __ CallRuntime(kUpdateFieldCidRuntimeEntry, 2); |
| __ Drop(2); |
| } |
| |
| __ Bind(&ok); |
| } |
| |
| LocationSummary* StoreInstanceFieldInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 2; |
| const intptr_t kNumTemps = (IsUnboxedDartFieldStore() && opt) |
| ? (FLAG_precompiled_mode ? 0 : 2) |
| : (IsPotentialUnboxedDartFieldStore() ? 3 : 0); |
| LocationSummary* summary = new (zone) LocationSummary( |
| zone, kNumInputs, kNumTemps, |
| (!FLAG_precompiled_mode && |
| ((IsUnboxedDartFieldStore() && opt && is_initialization()) || |
| IsPotentialUnboxedDartFieldStore())) |
| ? LocationSummary::kCallOnSlowPath |
| : LocationSummary::kNoCall); |
| |
| summary->set_in(kInstancePos, Location::RequiresRegister()); |
| if (slot().representation() != kTagged) { |
| ASSERT(RepresentationUtils::IsUnboxedInteger(slot().representation())); |
| ASSERT(RepresentationUtils::ValueSize(slot().representation()) <= |
| compiler::target::kWordSize); |
| summary->set_in(kValuePos, Location::RequiresRegister()); |
| } else if (IsUnboxedDartFieldStore() && opt) { |
| summary->set_in(kValuePos, Location::RequiresFpuRegister()); |
| if (!FLAG_precompiled_mode) { |
| summary->set_temp(0, Location::RequiresRegister()); |
| summary->set_temp(1, Location::RequiresRegister()); |
| } |
| } else if (IsPotentialUnboxedDartFieldStore()) { |
| summary->set_in(kValuePos, ShouldEmitStoreBarrier() |
| ? Location::WritableRegister() |
| : Location::RequiresRegister()); |
| summary->set_temp(0, Location::RequiresRegister()); |
| summary->set_temp(1, Location::RequiresRegister()); |
| summary->set_temp(2, opt ? Location::RequiresFpuRegister() |
| : Location::FpuRegisterLocation(XMM1)); |
| } else { |
| summary->set_in(kValuePos, |
| ShouldEmitStoreBarrier() |
| ? Location::RegisterLocation(kWriteBarrierValueReg) |
| : LocationRegisterOrConstant(value())); |
| } |
| return summary; |
| } |
| |
| static void EnsureMutableBox(FlowGraphCompiler* compiler, |
| StoreInstanceFieldInstr* instruction, |
| Register box_reg, |
| const Class& cls, |
| Register instance_reg, |
| intptr_t offset, |
| Register temp) { |
| compiler::Label done; |
| __ LoadCompressed(box_reg, compiler::FieldAddress(instance_reg, offset)); |
| __ CompareObject(box_reg, Object::null_object()); |
| __ j(NOT_EQUAL, &done); |
| BoxAllocationSlowPath::Allocate(compiler, instruction, cls, box_reg, temp); |
| __ movq(temp, box_reg); |
| __ StoreCompressedIntoObject(instance_reg, |
| compiler::FieldAddress(instance_reg, offset), |
| temp, compiler::Assembler::kValueIsNotSmi); |
| |
| __ Bind(&done); |
| } |
| |
| void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| ASSERT(compiler::target::UntaggedObject::kClassIdTagSize == 16); |
| ASSERT(sizeof(UntaggedField::guarded_cid_) == 2); |
| ASSERT(sizeof(UntaggedField::is_nullable_) == 2); |
| |
| compiler::Label skip_store; |
| |
| const Register instance_reg = locs()->in(kInstancePos).reg(); |
| const intptr_t offset_in_bytes = OffsetInBytes(); |
| ASSERT(offset_in_bytes > 0); // Field is finalized and points after header. |
| |
| if (slot().representation() != kTagged) { |
| ASSERT(memory_order_ != compiler::AssemblerBase::kRelease); |
| ASSERT(RepresentationUtils::IsUnboxedInteger(slot().representation())); |
| const Register value = locs()->in(kValuePos).reg(); |
| __ Comment("NativeUnboxedStoreInstanceFieldInstr"); |
| __ StoreFieldToOffset( |
| value, instance_reg, offset_in_bytes, |
| RepresentationUtils::OperandSize(slot().representation())); |
| return; |
| } |
| |
| if (IsUnboxedDartFieldStore() && compiler->is_optimizing()) { |
| ASSERT(memory_order_ != compiler::AssemblerBase::kRelease); |
| XmmRegister value = locs()->in(kValuePos).fpu_reg(); |
| const intptr_t cid = slot().field().UnboxedFieldCid(); |
| |
| // Real unboxed field |
| if (FLAG_precompiled_mode) { |
| switch (cid) { |
| case kDoubleCid: |
| __ Comment("UnboxedDoubleStoreInstanceFieldInstr"); |
| __ movsd(compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value); |
| return; |
| case kFloat32x4Cid: |
| __ Comment("UnboxedFloat32x4StoreInstanceFieldInstr"); |
| __ movups(compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value); |
| return; |
| case kFloat64x2Cid: |
| __ Comment("UnboxedFloat64x2StoreInstanceFieldInstr"); |
| __ movups(compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value); |
| return; |
| default: |
| UNREACHABLE(); |
| } |
| } |
| |
| Register temp = locs()->temp(0).reg(); |
| Register temp2 = locs()->temp(1).reg(); |
| |
| if (is_initialization()) { |
| const Class* cls = NULL; |
| switch (cid) { |
| case kDoubleCid: |
| cls = &compiler->double_class(); |
| break; |
| case kFloat32x4Cid: |
| cls = &compiler->float32x4_class(); |
| break; |
| case kFloat64x2Cid: |
| cls = &compiler->float64x2_class(); |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| |
| BoxAllocationSlowPath::Allocate(compiler, this, *cls, temp, temp2); |
| __ movq(temp2, temp); |
| __ StoreCompressedIntoObject( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| temp2, compiler::Assembler::kValueIsNotSmi); |
| } else { |
| __ LoadCompressed(temp, |
| compiler::FieldAddress(instance_reg, offset_in_bytes)); |
| } |
| switch (cid) { |
| case kDoubleCid: |
| __ Comment("UnboxedDoubleStoreInstanceFieldInstr"); |
| __ movsd(compiler::FieldAddress(temp, Double::value_offset()), value); |
| break; |
| case kFloat32x4Cid: |
| __ Comment("UnboxedFloat32x4StoreInstanceFieldInstr"); |
| __ movups(compiler::FieldAddress(temp, Float32x4::value_offset()), |
| value); |
| break; |
| case kFloat64x2Cid: |
| __ Comment("UnboxedFloat64x2StoreInstanceFieldInstr"); |
| __ movups(compiler::FieldAddress(temp, Float64x2::value_offset()), |
| value); |
| break; |
| default: |
| UNREACHABLE(); |
| } |
| return; |
| } |
| |
| if (IsPotentialUnboxedDartFieldStore()) { |
| ASSERT(memory_order_ != compiler::AssemblerBase::kRelease); |
| Register value_reg = locs()->in(kValuePos).reg(); |
| Register temp = locs()->temp(0).reg(); |
| Register temp2 = locs()->temp(1).reg(); |
| FpuRegister fpu_temp = locs()->temp(2).fpu_reg(); |
| |
| if (ShouldEmitStoreBarrier()) { |
| // Value input is a writable register and should be manually preserved |
| // across allocation slow-path. |
| locs()->live_registers()->Add(locs()->in(kValuePos), kTagged); |
| } |
| |
| compiler::Label store_pointer; |
| compiler::Label store_double; |
| compiler::Label store_float32x4; |
| compiler::Label store_float64x2; |
| |
| __ LoadObject(temp, Field::ZoneHandle(Z, slot().field().Original())); |
| |
| __ cmpw(compiler::FieldAddress(temp, Field::is_nullable_offset()), |
| compiler::Immediate(kNullCid)); |
| __ j(EQUAL, &store_pointer); |
| |
| __ movzxb(temp2, compiler::FieldAddress(temp, Field::kind_bits_offset())); |
| __ testq(temp2, compiler::Immediate(1 << Field::kUnboxingCandidateBit)); |
| __ j(ZERO, &store_pointer); |
| |
| __ cmpw(compiler::FieldAddress(temp, Field::guarded_cid_offset()), |
| compiler::Immediate(kDoubleCid)); |
| __ j(EQUAL, &store_double); |
| |
| __ cmpw(compiler::FieldAddress(temp, Field::guarded_cid_offset()), |
| compiler::Immediate(kFloat32x4Cid)); |
| __ j(EQUAL, &store_float32x4); |
| |
| __ cmpw(compiler::FieldAddress(temp, Field::guarded_cid_offset()), |
| compiler::Immediate(kFloat64x2Cid)); |
| __ j(EQUAL, &store_float64x2); |
| |
| // Fall through. |
| __ jmp(&store_pointer); |
| |
| if (!compiler->is_optimizing()) { |
| locs()->live_registers()->Add(locs()->in(kInstancePos)); |
| locs()->live_registers()->Add(locs()->in(kValuePos)); |
| } |
| |
| { |
| __ Bind(&store_double); |
| EnsureMutableBox(compiler, this, temp, compiler->double_class(), |
| instance_reg, offset_in_bytes, temp2); |
| __ movsd(fpu_temp, |
| compiler::FieldAddress(value_reg, Double::value_offset())); |
| __ movsd(compiler::FieldAddress(temp, Double::value_offset()), fpu_temp); |
| __ jmp(&skip_store); |
| } |
| |
| { |
| __ Bind(&store_float32x4); |
| EnsureMutableBox(compiler, this, temp, compiler->float32x4_class(), |
| instance_reg, offset_in_bytes, temp2); |
| __ movups(fpu_temp, |
| compiler::FieldAddress(value_reg, Float32x4::value_offset())); |
| __ movups(compiler::FieldAddress(temp, Float32x4::value_offset()), |
| fpu_temp); |
| __ jmp(&skip_store); |
| } |
| |
| { |
| __ Bind(&store_float64x2); |
| EnsureMutableBox(compiler, this, temp, compiler->float64x2_class(), |
| instance_reg, offset_in_bytes, temp2); |
| __ movups(fpu_temp, |
| compiler::FieldAddress(value_reg, Float64x2::value_offset())); |
| __ movups(compiler::FieldAddress(temp, Float64x2::value_offset()), |
| fpu_temp); |
| __ jmp(&skip_store); |
| } |
| |
| __ Bind(&store_pointer); |
| } |
| |
| const bool compressed = slot().is_compressed(); |
| if (ShouldEmitStoreBarrier()) { |
| Register value_reg = locs()->in(kValuePos).reg(); |
| if (!compressed) { |
| __ StoreIntoObject(instance_reg, |
| compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value_reg, CanValueBeSmi(), memory_order_); |
| } else { |
| __ StoreCompressedIntoObject( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value_reg, CanValueBeSmi(), memory_order_); |
| } |
| } else { |
| if (locs()->in(kValuePos).IsConstant()) { |
| const auto& value = locs()->in(kValuePos).constant(); |
| if (!compressed) { |
| __ StoreIntoObjectNoBarrier( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value, memory_order_); |
| } else { |
| __ StoreCompressedIntoObjectNoBarrier( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value, memory_order_); |
| } |
| } else { |
| Register value_reg = locs()->in(kValuePos).reg(); |
| if (!compressed) { |
| __ StoreIntoObjectNoBarrier( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value_reg, memory_order_); |
| } else { |
| __ StoreCompressedIntoObjectNoBarrier( |
| instance_reg, compiler::FieldAddress(instance_reg, offset_in_bytes), |
| value_reg, memory_order_); |
| } |
| } |
| } |
| __ Bind(&skip_store); |
| } |
| |
| LocationSummary* StoreStaticFieldInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 1; |
| const intptr_t kNumTemps = 1; |
| LocationSummary* locs = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall); |
| locs->set_in(0, Location::RegisterLocation(kWriteBarrierValueReg)); |
| locs->set_temp(0, Location::RequiresRegister()); |
| return locs; |
| } |
| |
| void StoreStaticFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) { |
| Register value = locs()->in(0).reg(); |
| Register temp = locs()->temp(0).reg(); |
| |
| compiler->used_static_fields().Add(&field()); |
| |
| __ movq(temp, |
| compiler::Address( |
| THR, compiler::target::Thread::field_table_values_offset())); |
| // Note: static fields ids won't be changed by hot-reload. |
| __ movq( |
| compiler::Address(temp, compiler::target::FieldTable::OffsetOf(field())), |
| value); |
| } |
| |
| LocationSummary* InstanceOfInstr::MakeLocationSummary(Zone* zone, |
| bool opt) const { |
| const intptr_t kNumInputs = 3; |
| const intptr_t kNumTemps = 0; |
| LocationSummary* summary = new (zone) |
| LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall); |
| summary->set_in(0, Location::RegisterLocation(TypeTestABI::kInstanceReg)); |
| summary->set_in(1, Location::RegisterLocation( |
| TypeTestABI:: |