diff options
Diffstat (limited to 'compiler')
25 files changed, 1145 insertions, 801 deletions
diff --git a/compiler/Android.mk b/compiler/Android.mk index 61ebf1d618..16a158cf6f 100644 --- a/compiler/Android.mk +++ b/compiler/Android.mk @@ -153,6 +153,7 @@ LIBART_COMPILER_SRC_FILES_x86_64 := \ optimizing/intrinsics_x86_64.cc \ optimizing/code_generator_x86_64.cc \ utils/x86_64/assembler_x86_64.cc \ + utils/x86_64/jni_macro_assembler_x86_64.cc \ utils/x86_64/managed_register_x86_64.cc \ diff --git a/compiler/driver/compiler_options.cc b/compiler/driver/compiler_options.cc index f20dba34a6..30ba8c9e74 100644 --- a/compiler/driver/compiler_options.cc +++ b/compiler/driver/compiler_options.cc @@ -44,7 +44,9 @@ CompilerOptions::CompilerOptions() init_failure_output_(nullptr), dump_cfg_file_name_(""), dump_cfg_append_(false), - force_determinism_(false) { + force_determinism_(false), + register_allocation_strategy_(RegisterAllocator::kRegisterAllocatorDefault), + passes_to_run_(nullptr) { } CompilerOptions::~CompilerOptions() { @@ -74,7 +76,9 @@ CompilerOptions::CompilerOptions(CompilerFilter::Filter compiler_filter, bool abort_on_hard_verifier_failure, const std::string& dump_cfg_file_name, bool dump_cfg_append, - bool force_determinism + bool force_determinism, + RegisterAllocator::Strategy regalloc_strategy, + const std::vector<std::string>* passes_to_run ) : // NOLINT(whitespace/parens) compiler_filter_(compiler_filter), huge_method_threshold_(huge_method_threshold), @@ -99,7 +103,9 @@ CompilerOptions::CompilerOptions(CompilerFilter::Filter compiler_filter, init_failure_output_(init_failure_output), dump_cfg_file_name_(dump_cfg_file_name), dump_cfg_append_(dump_cfg_append), - force_determinism_(force_determinism) { + force_determinism_(force_determinism), + register_allocation_strategy_(regalloc_strategy), + passes_to_run_(passes_to_run) { } void CompilerOptions::ParseHugeMethodMax(const StringPiece& option, UsageFn Usage) { @@ -144,6 +150,19 @@ void CompilerOptions::ParseDumpInitFailures(const StringPiece& option, } } +void CompilerOptions::ParseRegisterAllocationStrategy(const StringPiece& option, + UsageFn Usage) { + DCHECK(option.starts_with("--register-allocation-strategy=")); + StringPiece choice = option.substr(strlen("--register-allocation-strategy=")).data(); + if (choice == "linear-scan") { + register_allocation_strategy_ = RegisterAllocator::Strategy::kRegisterAllocatorLinearScan; + } else if (choice == "graph-color") { + register_allocation_strategy_ = RegisterAllocator::Strategy::kRegisterAllocatorGraphColor; + } else { + Usage("Unrecognized register allocation strategy. Try linear-scan, or graph-color."); + } +} + bool CompilerOptions::ParseCompilerOption(const StringPiece& option, UsageFn Usage) { if (option.starts_with("--compiler-filter=")) { const char* compiler_filter_string = option.substr(strlen("--compiler-filter=")).data(); @@ -190,6 +209,8 @@ bool CompilerOptions::ParseCompilerOption(const StringPiece& option, UsageFn Usa dump_cfg_file_name_ = option.substr(strlen("--dump-cfg=")).data(); } else if (option.starts_with("--dump-cfg-append")) { dump_cfg_append_ = true; + } else if (option.starts_with("--register-allocation-strategy=")) { + ParseRegisterAllocationStrategy(option, Usage); } else { // Option not recognized. return false; diff --git a/compiler/driver/compiler_options.h b/compiler/driver/compiler_options.h index 60b700ad91..abc58d7dda 100644 --- a/compiler/driver/compiler_options.h +++ b/compiler/driver/compiler_options.h @@ -24,6 +24,7 @@ #include "base/macros.h" #include "compiler_filter.h" #include "globals.h" +#include "optimizing/register_allocator.h" #include "utils.h" namespace art { @@ -74,7 +75,9 @@ class CompilerOptions FINAL { bool abort_on_hard_verifier_failure, const std::string& dump_cfg_file_name, bool dump_cfg_append, - bool force_determinism); + bool force_determinism, + RegisterAllocator::Strategy regalloc_strategy, + const std::vector<std::string>* passes_to_run); CompilerFilter::Filter GetCompilerFilter() const { return compiler_filter_; @@ -244,6 +247,14 @@ class CompilerOptions FINAL { return force_determinism_; } + RegisterAllocator::Strategy GetRegisterAllocationStrategy() const { + return register_allocation_strategy_; + } + + const std::vector<std::string>* GetPassesToRun() const { + return passes_to_run_; + } + private: void ParseDumpInitFailures(const StringPiece& option, UsageFn Usage); void ParseDumpCfgPasses(const StringPiece& option, UsageFn Usage); @@ -254,6 +265,7 @@ class CompilerOptions FINAL { void ParseSmallMethodMax(const StringPiece& option, UsageFn Usage); void ParseLargeMethodMax(const StringPiece& option, UsageFn Usage); void ParseHugeMethodMax(const StringPiece& option, UsageFn Usage); + void ParseRegisterAllocationStrategy(const StringPiece& option, UsageFn Usage); CompilerFilter::Filter compiler_filter_; size_t huge_method_threshold_; @@ -297,6 +309,16 @@ class CompilerOptions FINAL { // outcomes. bool force_determinism_; + RegisterAllocator::Strategy register_allocation_strategy_; + + // If not null, specifies optimization passes which will be run instead of defaults. + // Note that passes_to_run_ is not checked for correctness and providing an incorrect + // list of passes can lead to unexpected compiler behaviour. This is caused by dependencies + // between passes. Failing to satisfy them can for example lead to compiler crashes. + // Passing pass names which are not recognized by the compiler will result in + // compiler-dependant behavior. + const std::vector<std::string>* passes_to_run_; + friend class Dex2Oat; DISALLOW_COPY_AND_ASSIGN(CompilerOptions); diff --git a/compiler/jit/jit_compiler.cc b/compiler/jit/jit_compiler.cc index 178533849b..6f6a8f5a3b 100644 --- a/compiler/jit/jit_compiler.cc +++ b/compiler/jit/jit_compiler.cc @@ -32,6 +32,7 @@ #include "oat_file-inl.h" #include "oat_quick_method_header.h" #include "object_lock.h" +#include "optimizing/register_allocator.h" #include "thread_list.h" namespace art { @@ -110,7 +111,9 @@ JitCompiler::JitCompiler() { /* abort_on_hard_verifier_failure */ false, /* dump_cfg_file_name */ "", /* dump_cfg_append */ false, - /* force_determinism */ false)); + /* force_determinism */ false, + RegisterAllocator::kRegisterAllocatorDefault, + /* passes_to_run */ nullptr)); for (const std::string& argument : Runtime::Current()->GetCompilerOptions()) { compiler_options_->ParseCompilerOption(argument, Usage); } diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc index c105940f28..5eaf11e9fb 100644 --- a/compiler/optimizing/code_generator_arm.cc +++ b/compiler/optimizing/code_generator_arm.cc @@ -432,6 +432,11 @@ class ReadBarrierMarkSlowPathARM : public SlowPathCode { (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) << "Unexpected instruction in read barrier marking slow path: " << instruction_->DebugName(); + // The read barrier instrumentation of object ArrayGet + // instructions does not support the HIntermediateAddress + // instruction. + DCHECK(!(instruction_->IsArrayGet() && + instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); __ Bind(GetEntryLabel()); // No need to save live registers; it's taken care of by the @@ -512,6 +517,11 @@ class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCode { (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) << "Unexpected instruction in read barrier for heap reference slow path: " << instruction_->DebugName(); + // The read barrier instrumentation of object ArrayGet + // instructions does not support the HIntermediateAddress + // instruction. + DCHECK(!(instruction_->IsArrayGet() && + instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); __ Bind(GetEntryLabel()); SaveLiveRegisters(codegen, locations); @@ -4467,8 +4477,6 @@ void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { Primitive::Type type = instruction->GetType(); HInstruction* array_instr = instruction->GetArray(); bool has_intermediate_address = array_instr->IsIntermediateAddress(); - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); switch (type) { case Primitive::kPrimBoolean: @@ -4503,6 +4511,11 @@ void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { } case Primitive::kPrimNot: { + // The read barrier instrumentation of object ArrayGet + // instructions does not support the HIntermediateAddress + // instruction. + DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); + static_assert( sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); @@ -4645,8 +4658,6 @@ void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { Location value_loc = locations->InAt(2); HInstruction* array_instr = instruction->GetArray(); bool has_intermediate_address = array_instr->IsIntermediateAddress(); - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); switch (value_type) { case Primitive::kPrimBoolean: @@ -4911,8 +4922,6 @@ void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { } void LocationsBuilderARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); @@ -4927,9 +4936,6 @@ void InstructionCodeGeneratorARM::VisitIntermediateAddress(HIntermediateAddress* Location first = locations->InAt(0); Location second = locations->InAt(1); - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); - if (second.IsRegister()) { __ add(out.AsRegister<Register>(), first.AsRegister<Register>(), diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc index 54c9efcafe..9ceb3109cd 100644 --- a/compiler/optimizing/code_generator_arm64.cc +++ b/compiler/optimizing/code_generator_arm64.cc @@ -598,6 +598,11 @@ class ReadBarrierMarkSlowPathARM64 : public SlowPathCodeARM64 { (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) << "Unexpected instruction in read barrier marking slow path: " << instruction_->DebugName(); + // The read barrier instrumentation of object ArrayGet + // instructions does not support the HIntermediateAddress + // instruction. + DCHECK(!(instruction_->IsArrayGet() && + instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); __ Bind(GetEntryLabel()); // No need to save live registers; it's taken care of by the @@ -680,7 +685,9 @@ class ReadBarrierForHeapReferenceSlowPathARM64 : public SlowPathCodeARM64 { (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) << "Unexpected instruction in read barrier for heap reference slow path: " << instruction_->DebugName(); - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. + // The read barrier instrumentation of object ArrayGet + // instructions does not support the HIntermediateAddress + // instruction. DCHECK(!(instruction_->IsArrayGet() && instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); @@ -1983,8 +1990,6 @@ void InstructionCodeGeneratorARM64::VisitArm64DataProcWithShifterOp( } void LocationsBuilderARM64::VisitIntermediateAddress(HIntermediateAddress* instruction) { - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); locations->SetInAt(0, Location::RequiresRegister()); @@ -1992,10 +1997,7 @@ void LocationsBuilderARM64::VisitIntermediateAddress(HIntermediateAddress* instr locations->SetOut(Location::RequiresRegister()); } -void InstructionCodeGeneratorARM64::VisitIntermediateAddress( - HIntermediateAddress* instruction) { - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); +void InstructionCodeGeneratorARM64::VisitIntermediateAddress(HIntermediateAddress* instruction) { __ Add(OutputRegister(instruction), InputRegisterAt(instruction, 0), Operand(InputOperandAt(instruction, 1))); @@ -2091,11 +2093,15 @@ void InstructionCodeGeneratorARM64::VisitArrayGet(HArrayGet* instruction) { // Block pools between `Load` and `MaybeRecordImplicitNullCheck`. BlockPoolsScope block_pools(masm); + // The read barrier instrumentation of object ArrayGet instructions + // does not support the HIntermediateAddress instruction. + DCHECK(!((type == Primitive::kPrimNot) && + instruction->GetArray()->IsIntermediateAddress() && + kEmitCompilerReadBarrier)); + if (type == Primitive::kPrimNot && kEmitCompilerReadBarrier && kUseBakerReadBarrier) { // Object ArrayGet with Baker's read barrier case. Register temp = temps.AcquireW(); - // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. - DCHECK(!instruction->GetArray()->IsIntermediateAddress()); // Note that a potential implicit null check is handled in the // CodeGeneratorARM64::GenerateArrayLoadWithBakerReadBarrier call. codegen_->GenerateArrayLoadWithBakerReadBarrier( @@ -2109,9 +2115,6 @@ void InstructionCodeGeneratorARM64::VisitArrayGet(HArrayGet* instruction) { } else { Register temp = temps.AcquireSameSizeAs(obj); if (instruction->GetArray()->IsIntermediateAddress()) { - // The read barrier instrumentation does not support the - // HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); // We do not need to compute the intermediate address from the array: the // input instruction has done it already. See the comment in // `TryExtractArrayAccessAddress()`. @@ -2201,9 +2204,6 @@ void InstructionCodeGeneratorARM64::VisitArraySet(HArraySet* instruction) { UseScratchRegisterScope temps(masm); Register temp = temps.AcquireSameSizeAs(array); if (instruction->GetArray()->IsIntermediateAddress()) { - // The read barrier instrumentation does not support the - // HIntermediateAddress instruction yet. - DCHECK(!kEmitCompilerReadBarrier); // We do not need to compute the intermediate address from the array: the // input instruction has done it already. See the comment in // `TryExtractArrayAccessAddress()`. @@ -2223,7 +2223,6 @@ void InstructionCodeGeneratorARM64::VisitArraySet(HArraySet* instruction) { codegen_->Store(value_type, value, destination); codegen_->MaybeRecordImplicitNullCheck(instruction); } else { - DCHECK(needs_write_barrier); DCHECK(!instruction->GetArray()->IsIntermediateAddress()); vixl::aarch64::Label done; SlowPathCodeARM64* slow_path = nullptr; diff --git a/compiler/optimizing/dead_code_elimination.h b/compiler/optimizing/dead_code_elimination.h index 0ce0ec1402..58e700deba 100644 --- a/compiler/optimizing/dead_code_elimination.h +++ b/compiler/optimizing/dead_code_elimination.h @@ -31,13 +31,11 @@ class HDeadCodeElimination : public HOptimization { public: HDeadCodeElimination(HGraph* graph, OptimizingCompilerStats* stats = nullptr, - const char* name = kInitialDeadCodeEliminationPassName) + const char* name = kDeadCodeEliminationPassName) : HOptimization(graph, name, stats) {} void Run() OVERRIDE; - - static constexpr const char* kInitialDeadCodeEliminationPassName = "dead_code_elimination"; - static constexpr const char* kFinalDeadCodeEliminationPassName = "dead_code_elimination_final"; + static constexpr const char* kDeadCodeEliminationPassName = "dead_code_elimination"; private: void MaybeRecordDeadBlock(HBasicBlock* block); diff --git a/compiler/optimizing/dex_cache_array_fixups_arm.h b/compiler/optimizing/dex_cache_array_fixups_arm.h index 015f910328..9142e29eff 100644 --- a/compiler/optimizing/dex_cache_array_fixups_arm.h +++ b/compiler/optimizing/dex_cache_array_fixups_arm.h @@ -26,7 +26,9 @@ namespace arm { class DexCacheArrayFixups : public HOptimization { public: DexCacheArrayFixups(HGraph* graph, OptimizingCompilerStats* stats) - : HOptimization(graph, "dex_cache_array_fixups_arm", stats) {} + : HOptimization(graph, kDexCacheArrayFixupsArmPassName, stats) {} + + static constexpr const char* kDexCacheArrayFixupsArmPassName = "dex_cache_array_fixups_arm"; void Run() OVERRIDE; }; diff --git a/compiler/optimizing/dex_cache_array_fixups_mips.h b/compiler/optimizing/dex_cache_array_fixups_mips.h index 21056e130a..861a199d6c 100644 --- a/compiler/optimizing/dex_cache_array_fixups_mips.h +++ b/compiler/optimizing/dex_cache_array_fixups_mips.h @@ -29,9 +29,11 @@ namespace mips { class DexCacheArrayFixups : public HOptimization { public: DexCacheArrayFixups(HGraph* graph, CodeGenerator* codegen, OptimizingCompilerStats* stats) - : HOptimization(graph, "dex_cache_array_fixups_mips", stats), + : HOptimization(graph, kDexCacheArrayFixupsMipsPassName, stats), codegen_(codegen) {} + static constexpr const char* kDexCacheArrayFixupsMipsPassName = "dex_cache_array_fixups_mips"; + void Run() OVERRIDE; private: diff --git a/compiler/optimizing/induction_var_analysis.h b/compiler/optimizing/induction_var_analysis.h index 7c74816c26..cd4c830645 100644 --- a/compiler/optimizing/induction_var_analysis.h +++ b/compiler/optimizing/induction_var_analysis.h @@ -39,9 +39,9 @@ class HInductionVarAnalysis : public HOptimization { void Run() OVERRIDE; - private: static constexpr const char* kInductionPassName = "induction_var_analysis"; + private: struct NodeInfo { explicit NodeInfo(uint32_t d) : depth(d), done(false) {} uint32_t depth; diff --git a/compiler/optimizing/instruction_simplifier_arm.h b/compiler/optimizing/instruction_simplifier_arm.h index 3d297dacc0..782110c40a 100644 --- a/compiler/optimizing/instruction_simplifier_arm.h +++ b/compiler/optimizing/instruction_simplifier_arm.h @@ -48,7 +48,9 @@ class InstructionSimplifierArmVisitor : public HGraphVisitor { class InstructionSimplifierArm : public HOptimization { public: InstructionSimplifierArm(HGraph* graph, OptimizingCompilerStats* stats) - : HOptimization(graph, "instruction_simplifier_arm", stats) {} + : HOptimization(graph, kInstructionSimplifierArmPassName, stats) {} + + static constexpr const char* kInstructionSimplifierArmPassName = "instruction_simplifier_arm"; void Run() OVERRIDE { InstructionSimplifierArmVisitor visitor(graph_, stats_); diff --git a/compiler/optimizing/instruction_simplifier_arm64.h b/compiler/optimizing/instruction_simplifier_arm64.h index 28648b3bea..f71684efe9 100644 --- a/compiler/optimizing/instruction_simplifier_arm64.h +++ b/compiler/optimizing/instruction_simplifier_arm64.h @@ -82,8 +82,9 @@ class InstructionSimplifierArm64Visitor : public HGraphVisitor { class InstructionSimplifierArm64 : public HOptimization { public: InstructionSimplifierArm64(HGraph* graph, OptimizingCompilerStats* stats) - : HOptimization(graph, "instruction_simplifier_arm64", stats) {} - + : HOptimization(graph, kInstructionSimplifierArm64PassName, stats) {} + static constexpr const char* kInstructionSimplifierArm64PassName + = "instruction_simplifier_arm64"; void Run() OVERRIDE { InstructionSimplifierArm64Visitor visitor(graph_, stats_); visitor.VisitReversePostOrder(); diff --git a/compiler/optimizing/instruction_simplifier_shared.cc b/compiler/optimizing/instruction_simplifier_shared.cc index 8f7778fe68..6632cd9969 100644 --- a/compiler/optimizing/instruction_simplifier_shared.cc +++ b/compiler/optimizing/instruction_simplifier_shared.cc @@ -231,15 +231,6 @@ bool TryExtractArrayAccessAddress(HInstruction* access, HInstruction* array, HInstruction* index, size_t data_offset) { - if (kEmitCompilerReadBarrier) { - // The read barrier instrumentation does not support the - // HIntermediateAddress instruction yet. - // - // TODO: Handle this case properly in the ARM64 and ARM code generator and - // re-enable this optimization; otherwise, remove this TODO. - // b/26601270 - return false; - } if (index->IsConstant() || (index->IsBoundsCheck() && index->AsBoundsCheck()->GetIndex()->IsConstant())) { // When the index is a constant all the addressing can be fitted in the @@ -251,6 +242,13 @@ bool TryExtractArrayAccessAddress(HInstruction* access, // The access may require a runtime call or the original array pointer. return false; } + if (kEmitCompilerReadBarrier && + access->IsArrayGet() && + access->AsArrayGet()->GetType() == Primitive::kPrimNot) { + // For object arrays, the read barrier instrumentation requires + // the original array pointer. + return false; + } // Proceed to extract the base address computation. HGraph* graph = access->GetBlock()->GetGraph(); diff --git a/compiler/optimizing/optimization.h b/compiler/optimizing/optimization.h index 2f59d4cd5b..0819fb01ac 100644 --- a/compiler/optimizing/optimization.h +++ b/compiler/optimizing/optimization.h @@ -37,7 +37,10 @@ class HOptimization : public ArenaObject<kArenaAllocOptimization> { virtual ~HOptimization() {} - // Return the name of the pass. + // Return the name of the pass. Pass names for a single HOptimization should be of form + // <optimization_name> or <optimization_name>$<pass_name> for common <optimization_name> prefix. + // Example: 'instruction_simplifier', 'instruction_simplifier$after_bce', + // 'instruction_simplifier$before_codegen'. const char* GetPassName() const { return pass_name_; } // Perform the analysis itself. diff --git a/compiler/optimizing/optimizing_compiler.cc b/compiler/optimizing/optimizing_compiler.cc index d5b0d77fe5..f7c82d1987 100644 --- a/compiler/optimizing/optimizing_compiler.cc +++ b/compiler/optimizing/optimizing_compiler.cc @@ -95,6 +95,8 @@ namespace art { static constexpr size_t kArenaAllocatorMemoryReportThreshold = 8 * MB; +static constexpr const char* kPassNameSeparator = "$"; + /** * Used by the code generator, to allocate the code in a vector. */ @@ -266,7 +268,7 @@ class PassScope : public ValueObject { class OptimizingCompiler FINAL : public Compiler { public: explicit OptimizingCompiler(CompilerDriver* driver); - ~OptimizingCompiler(); + ~OptimizingCompiler() OVERRIDE; bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file) const OVERRIDE; @@ -305,17 +307,17 @@ class OptimizingCompiler FINAL : public Compiler { OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_); - protected: - virtual void RunOptimizations(HGraph* graph, - CodeGenerator* codegen, - CompilerDriver* driver, - const DexCompilationUnit& dex_compilation_unit, - PassObserver* pass_observer, - StackHandleScopeCollection* handles) const; + private: + void RunOptimizations(HGraph* graph, + CodeGenerator* codegen, + CompilerDriver* driver, + const DexCompilationUnit& dex_compilation_unit, + PassObserver* pass_observer, + StackHandleScopeCollection* handles) const; - virtual void RunOptimizations(HOptimization* optimizations[], - size_t length, - PassObserver* pass_observer) const; + void RunOptimizations(HOptimization* optimizations[], + size_t length, + PassObserver* pass_observer) const; private: // Create a 'CompiledMethod' for an optimized graph. @@ -420,6 +422,117 @@ static bool InstructionSetSupportsReadBarrier(InstructionSet instruction_set) { || instruction_set == kX86_64; } +static HOptimization* BuildOptimization( + const std::string& opt_name, + ArenaAllocator* arena, + HGraph* graph, + OptimizingCompilerStats* stats, + CodeGenerator* codegen, + CompilerDriver* driver, + const DexCompilationUnit& dex_compilation_unit, + StackHandleScopeCollection* handles, + SideEffectsAnalysis* most_recent_side_effects, + HInductionVarAnalysis* most_recent_induction) { + if (opt_name == arm::InstructionSimplifierArm::kInstructionSimplifierArmPassName) { + return new (arena) arm::InstructionSimplifierArm(graph, stats); + } else if (opt_name == arm64::InstructionSimplifierArm64::kInstructionSimplifierArm64PassName) { + return new (arena) arm64::InstructionSimplifierArm64(graph, stats); + } else if (opt_name == BoundsCheckElimination::kBoundsCheckEliminationPassName) { + CHECK(most_recent_side_effects != nullptr && most_recent_induction != nullptr); + return new (arena) BoundsCheckElimination(graph, + *most_recent_side_effects, + most_recent_induction); + } else if (opt_name == GVNOptimization::kGlobalValueNumberingPassName) { + CHECK(most_recent_side_effects != nullptr); + return new (arena) GVNOptimization(graph, *most_recent_side_effects); + } else if (opt_name == HConstantFolding::kConstantFoldingPassName) { + return new (arena) HConstantFolding(graph); + } else if (opt_name == HDeadCodeElimination::kDeadCodeEliminationPassName) { + return new (arena) HDeadCodeElimination(graph, stats); + } else if (opt_name == HInliner::kInlinerPassName) { + size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_; + return new (arena) HInliner(graph, // outer_graph + graph, // outermost_graph + codegen, + dex_compilation_unit, // outer_compilation_unit + dex_compilation_unit, // outermost_compilation_unit + driver, + handles, + stats, + number_of_dex_registers, + /* depth */ 0); + } else if (opt_name == HSharpening::kSharpeningPassName) { + return new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver); + } else if (opt_name == HSelectGenerator::kSelectGeneratorPassName) { + return new (arena) HSelectGenerator(graph, stats); + } else if (opt_name == HInductionVarAnalysis::kInductionPassName) { + return new (arena) HInductionVarAnalysis(graph); + } else if (opt_name == InstructionSimplifier::kInstructionSimplifierPassName) { + return new (arena) InstructionSimplifier(graph, stats); + } else if (opt_name == IntrinsicsRecognizer::kIntrinsicsRecognizerPassName) { + return new (arena) IntrinsicsRecognizer(graph, driver, stats); + } else if (opt_name == LICM::kLoopInvariantCodeMotionPassName) { + CHECK(most_recent_side_effects != nullptr); + return new (arena) LICM(graph, *most_recent_side_effects, stats); + } else if (opt_name == LoadStoreElimination::kLoadStoreEliminationPassName) { + CHECK(most_recent_side_effects != nullptr); + return new (arena) LoadStoreElimination(graph, *most_recent_side_effects); + } else if (opt_name == mips::DexCacheArrayFixups::kDexCacheArrayFixupsMipsPassName) { + return new (arena) mips::DexCacheArrayFixups(graph, codegen, stats); + } else if (opt_name == mips::PcRelativeFixups::kPcRelativeFixupsMipsPassName) { + return new (arena) mips::PcRelativeFixups(graph, codegen, stats); + } else if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) { + return new (arena) SideEffectsAnalysis(graph); + } else if (opt_name == x86::PcRelativeFixups::kPcRelativeFixupsX86PassName) { + return new (arena) x86::PcRelativeFixups(graph, codegen, stats); + } else if (opt_name == x86::X86MemoryOperandGeneration::kX86MemoryOperandGenerationPassName) { + return new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats); + } + return nullptr; +} + +static ArenaVector<HOptimization*> BuildOptimizations( + const std::vector<std::string>& pass_names, + ArenaAllocator* arena, + HGraph* graph, + OptimizingCompilerStats* stats, + CodeGenerator* codegen, + CompilerDriver* driver, + const DexCompilationUnit& dex_compilation_unit, + StackHandleScopeCollection* handles) { + // Few HOptimizations constructors require SideEffectsAnalysis or HInductionVarAnalysis + // instances. This method assumes that each of them expects the nearest instance preceeding it + // in the pass name list. + SideEffectsAnalysis* most_recent_side_effects = nullptr; + HInductionVarAnalysis* most_recent_induction = nullptr; + ArenaVector<HOptimization*> ret(arena->Adapter()); + for (std::string pass_name : pass_names) { + size_t pos = pass_name.find(kPassNameSeparator); // Strip suffix to get base pass name. + std::string opt_name = pos == std::string::npos ? pass_name : pass_name.substr(0, pos); + + HOptimization* opt = BuildOptimization( + opt_name, + arena, + graph, + stats, + codegen, + driver, + dex_compilation_unit, + handles, + most_recent_side_effects, + most_recent_induction); + CHECK(opt != nullptr) << "Couldn't build optimization: \"" << pass_name << "\""; + ret.push_back(opt); + + if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) { + most_recent_side_effects = down_cast<SideEffectsAnalysis*>(opt); + } else if (opt_name == HInductionVarAnalysis::kInductionPassName) { + most_recent_induction = down_cast<HInductionVarAnalysis*>(opt); + } + } + return ret; +} + void OptimizingCompiler::RunOptimizations(HOptimization* optimizations[], size_t length, PassObserver* pass_observer) const { @@ -444,11 +557,11 @@ void OptimizingCompiler::MaybeRunInliner(HGraph* graph, } size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_; HInliner* inliner = new (graph->GetArena()) HInliner( - graph, - graph, + graph, // outer_graph + graph, // outermost_graph codegen, - dex_compilation_unit, - dex_compilation_unit, + dex_compilation_unit, // outer_compilation_unit + dex_compilation_unit, // outermost_compilation_unit driver, handles, stats, @@ -473,7 +586,7 @@ void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set, arm::InstructionSimplifierArm* simplifier = new (arena) arm::InstructionSimplifierArm(graph, stats); SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); - GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN_after_arch"); + GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch"); HOptimization* arm_optimizations[] = { simplifier, side_effects, @@ -489,7 +602,7 @@ void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set, arm64::InstructionSimplifierArm64* simplifier = new (arena) arm64::InstructionSimplifierArm64(graph, stats); SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); - GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN_after_arch"); + GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch"); HOptimization* arm64_optimizations[] = { simplifier, side_effects, @@ -518,7 +631,7 @@ void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set, x86::PcRelativeFixups* pc_relative_fixups = new (arena) x86::PcRelativeFixups(graph, codegen, stats); x86::X86MemoryOperandGeneration* memory_gen = - new(arena) x86::X86MemoryOperandGeneration(graph, stats, codegen); + new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats); HOptimization* x86_optimizations[] = { pc_relative_fixups, memory_gen @@ -530,7 +643,7 @@ void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set, #ifdef ART_ENABLE_CODEGEN_x86_64 case kX86_64: { x86::X86MemoryOperandGeneration* memory_gen = - new(arena) x86::X86MemoryOperandGeneration(graph, stats, codegen); + new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats); HOptimization* x86_64_optimizations[] = { memory_gen }; @@ -546,7 +659,8 @@ void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set, NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects. static void AllocateRegisters(HGraph* graph, CodeGenerator* codegen, - PassObserver* pass_observer) { + PassObserver* pass_observer, + RegisterAllocator::Strategy strategy) { { PassScope scope(PrepareForRegisterAllocation::kPrepareForRegisterAllocationPassName, pass_observer); @@ -559,7 +673,7 @@ static void AllocateRegisters(HGraph* graph, } { PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer); - RegisterAllocator::Create(graph->GetArena(), codegen, liveness)->AllocateRegisters(); + RegisterAllocator::Create(graph->GetArena(), codegen, liveness, strategy)->AllocateRegisters(); } } @@ -571,15 +685,30 @@ void OptimizingCompiler::RunOptimizations(HGraph* graph, StackHandleScopeCollection* handles) const { OptimizingCompilerStats* stats = compilation_stats_.get(); ArenaAllocator* arena = graph->GetArena(); + if (driver->GetCompilerOptions().GetPassesToRun() != nullptr) { + ArenaVector<HOptimization*> optimizations = BuildOptimizations( + *driver->GetCompilerOptions().GetPassesToRun(), + arena, + graph, + stats, + codegen, + driver, + dex_compilation_unit, + handles); + RunOptimizations(&optimizations[0], optimizations.size(), pass_observer); + return; + } + HDeadCodeElimination* dce1 = new (arena) HDeadCodeElimination( - graph, stats, HDeadCodeElimination::kInitialDeadCodeEliminationPassName); + graph, stats, "dead_code_elimination$initial"); HDeadCodeElimination* dce2 = new (arena) HDeadCodeElimination( - graph, stats, HDeadCodeElimination::kFinalDeadCodeEliminationPassName); + graph, stats, "dead_code_elimination$final"); HConstantFolding* fold1 = new (arena) HConstantFolding(graph); InstructionSimplifier* simplify1 = new (arena) InstructionSimplifier(graph, stats); HSelectGenerator* select_generator = new (arena) HSelectGenerator(graph, stats); - HConstantFolding* fold2 = new (arena) HConstantFolding(graph, "constant_folding_after_inlining"); - HConstantFolding* fold3 = new (arena) HConstantFolding(graph, "constant_folding_after_bce"); + HConstantFolding* fold2 = new (arena) HConstantFolding( + graph, "constant_folding$after_inlining"); + HConstantFolding* fold3 = new (arena) HConstantFolding(graph, "constant_folding$after_bce"); SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects); LICM* licm = new (arena) LICM(graph, *side_effects, stats); @@ -588,9 +717,9 @@ void OptimizingCompiler::RunOptimizations(HGraph* graph, BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, *side_effects, induction); HSharpening* sharpening = new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver); InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier( - graph, stats, "instruction_simplifier_after_bce"); + graph, stats, "instruction_simplifier$after_bce"); InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier( - graph, stats, "instruction_simplifier_before_codegen"); + graph, stats, "instruction_simplifier$before_codegen"); IntrinsicsRecognizer* intrinsics = new (arena) IntrinsicsRecognizer(graph, driver, stats); HOptimization* optimizations1[] = { @@ -626,7 +755,6 @@ void OptimizingCompiler::RunOptimizations(HGraph* graph, RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer); RunArchOptimizations(driver->GetInstructionSet(), graph, codegen, pass_observer); - AllocateRegisters(graph, codegen, pass_observer); } static ArenaVector<LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) { @@ -841,6 +969,10 @@ CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena, &pass_observer, &handles); + RegisterAllocator::Strategy regalloc_strategy = + compiler_options.GetRegisterAllocationStrategy(); + AllocateRegisters(graph, codegen.get(), &pass_observer, regalloc_strategy); + codegen->Compile(code_allocator); pass_observer.DumpDisassembly(); } diff --git a/compiler/optimizing/pc_relative_fixups_mips.h b/compiler/optimizing/pc_relative_fixups_mips.h index 1e8b071bb3..5a7397bf9d 100644 --- a/compiler/optimizing/pc_relative_fixups_mips.h +++ b/compiler/optimizing/pc_relative_fixups_mips.h @@ -32,6 +32,8 @@ class PcRelativeFixups : public HOptimization { : HOptimization(graph, "pc_relative_fixups_mips", stats), codegen_(codegen) {} + static constexpr const char* kPcRelativeFixupsMipsPassName = "pc_relative_fixups_mips"; + void Run() OVERRIDE; private: diff --git a/compiler/optimizing/pc_relative_fixups_x86.h b/compiler/optimizing/pc_relative_fixups_x86.h index 03de2fcece..72fa71ea94 100644 --- a/compiler/optimizing/pc_relative_fixups_x86.h +++ b/compiler/optimizing/pc_relative_fixups_x86.h @@ -29,9 +29,11 @@ namespace x86 { class PcRelativeFixups : public HOptimization { public: PcRelativeFixups(HGraph* graph, CodeGenerator* codegen, OptimizingCompilerStats* stats) - : HOptimization(graph, "pc_relative_fixups_x86", stats), + : HOptimization(graph, kPcRelativeFixupsX86PassName, stats), codegen_(codegen) {} + static constexpr const char* kPcRelativeFixupsX86PassName = "pc_relative_fixups_x86"; + void Run() OVERRIDE; private: diff --git a/compiler/optimizing/x86_memory_gen.cc b/compiler/optimizing/x86_memory_gen.cc index 195159f61b..8aa315a7e3 100644 --- a/compiler/optimizing/x86_memory_gen.cc +++ b/compiler/optimizing/x86_memory_gen.cc @@ -69,8 +69,8 @@ class MemoryOperandVisitor : public HGraphVisitor { }; X86MemoryOperandGeneration::X86MemoryOperandGeneration(HGraph* graph, - OptimizingCompilerStats* stats, - CodeGenerator* codegen) + CodeGenerator* codegen, + OptimizingCompilerStats* stats) : HOptimization(graph, kX86MemoryOperandGenerationPassName, stats), do_implicit_null_checks_(codegen->GetCompilerOptions().GetImplicitNullChecks()) { } diff --git a/compiler/optimizing/x86_memory_gen.h b/compiler/optimizing/x86_memory_gen.h index 7e886819bb..5f15d9f1e6 100644 --- a/compiler/optimizing/x86_memory_gen.h +++ b/compiler/optimizing/x86_memory_gen.h @@ -28,8 +28,8 @@ namespace x86 { class X86MemoryOperandGeneration : public HOptimization { public: X86MemoryOperandGeneration(HGraph* graph, - OptimizingCompilerStats* stats, - CodeGenerator* codegen); + CodeGenerator* codegen, + OptimizingCompilerStats* stats); void Run() OVERRIDE; diff --git a/compiler/utils/jni_macro_assembler.cc b/compiler/utils/jni_macro_assembler.cc index 6fdf09baa3..1acc90ca6f 100644 --- a/compiler/utils/jni_macro_assembler.cc +++ b/compiler/utils/jni_macro_assembler.cc @@ -36,7 +36,7 @@ #include "x86/jni_macro_assembler_x86.h" #endif #ifdef ART_ENABLE_CODEGEN_x86_64 -#include "x86_64/assembler_x86_64.h" +#include "x86_64/jni_macro_assembler_x86_64.h" #endif #include "base/casts.h" #include "globals.h" @@ -98,7 +98,7 @@ MacroAsm64UniquePtr JNIMacroAssembler<PointerSize::k64>::Create( #endif #ifdef ART_ENABLE_CODEGEN_x86_64 case kX86_64: - return MacroAsm64UniquePtr(new (arena) x86_64::X86_64Assembler(arena)); + return MacroAsm64UniquePtr(new (arena) x86_64::X86_64JNIMacroAssembler(arena)); #endif default: LOG(FATAL) << "Unknown/unsupported 8B InstructionSet: " << instruction_set; diff --git a/compiler/utils/x86_64/assembler_x86_64.cc b/compiler/utils/x86_64/assembler_x86_64.cc index ce4ea1d8fd..ddc824425e 100644 --- a/compiler/utils/x86_64/assembler_x86_64.cc +++ b/compiler/utils/x86_64/assembler_x86_64.cc @@ -2639,543 +2639,6 @@ void X86_64Assembler::EmitOptionalByteRegNormalizingRex32(CpuRegister dst, const } } -static dwarf::Reg DWARFReg(Register reg) { - return dwarf::Reg::X86_64Core(static_cast<int>(reg)); -} -static dwarf::Reg DWARFReg(FloatRegister reg) { - return dwarf::Reg::X86_64Fp(static_cast<int>(reg)); -} - -constexpr size_t kFramePointerSize = 8; - -void X86_64Assembler::BuildFrame(size_t frame_size, - ManagedRegister method_reg, - ArrayRef<const ManagedRegister> spill_regs, - const ManagedRegisterEntrySpills& entry_spills) { - DCHECK_EQ(buffer_.Size(), 0U); // Nothing emitted yet. - cfi_.SetCurrentCFAOffset(8); // Return address on stack. - CHECK_ALIGNED(frame_size, kStackAlignment); - int gpr_count = 0; - for (int i = spill_regs.size() - 1; i >= 0; --i) { - x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); - if (spill.IsCpuRegister()) { - pushq(spill.AsCpuRegister()); - gpr_count++; - cfi_.AdjustCFAOffset(kFramePointerSize); - cfi_.RelOffset(DWARFReg(spill.AsCpuRegister().AsRegister()), 0); - } - } - // return address then method on stack. - int64_t rest_of_frame = static_cast<int64_t>(frame_size) - - (gpr_count * kFramePointerSize) - - kFramePointerSize /*return address*/; - subq(CpuRegister(RSP), Immediate(rest_of_frame)); - cfi_.AdjustCFAOffset(rest_of_frame); - - // spill xmms - int64_t offset = rest_of_frame; - for (int i = spill_regs.size() - 1; i >= 0; --i) { - x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); - if (spill.IsXmmRegister()) { - offset -= sizeof(double); - movsd(Address(CpuRegister(RSP), offset), spill.AsXmmRegister()); - cfi_.RelOffset(DWARFReg(spill.AsXmmRegister().AsFloatRegister()), offset); - } - } - - static_assert(static_cast<size_t>(kX86_64PointerSize) == kFramePointerSize, - "Unexpected frame pointer size."); - - movq(Address(CpuRegister(RSP), 0), method_reg.AsX86_64().AsCpuRegister()); - - for (size_t i = 0; i < entry_spills.size(); ++i) { - ManagedRegisterSpill spill = entry_spills.at(i); - if (spill.AsX86_64().IsCpuRegister()) { - if (spill.getSize() == 8) { - movq(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), - spill.AsX86_64().AsCpuRegister()); - } else { - CHECK_EQ(spill.getSize(), 4); - movl(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), spill.AsX86_64().AsCpuRegister()); - } - } else { - if (spill.getSize() == 8) { - movsd(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), spill.AsX86_64().AsXmmRegister()); - } else { - CHECK_EQ(spill.getSize(), 4); - movss(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), spill.AsX86_64().AsXmmRegister()); - } - } - } -} - -void X86_64Assembler::RemoveFrame(size_t frame_size, ArrayRef<const ManagedRegister> spill_regs) { - CHECK_ALIGNED(frame_size, kStackAlignment); - cfi_.RememberState(); - int gpr_count = 0; - // unspill xmms - int64_t offset = static_cast<int64_t>(frame_size) - (spill_regs.size() * kFramePointerSize) - 2 * kFramePointerSize; - for (size_t i = 0; i < spill_regs.size(); ++i) { - x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); - if (spill.IsXmmRegister()) { - offset += sizeof(double); - movsd(spill.AsXmmRegister(), Address(CpuRegister(RSP), offset)); - cfi_.Restore(DWARFReg(spill.AsXmmRegister().AsFloatRegister())); - } else { - gpr_count++; - } - } - int adjust = static_cast<int>(frame_size) - (gpr_count * kFramePointerSize) - kFramePointerSize; - addq(CpuRegister(RSP), Immediate(adjust)); - cfi_.AdjustCFAOffset(-adjust); - for (size_t i = 0; i < spill_regs.size(); ++i) { - x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); - if (spill.IsCpuRegister()) { - popq(spill.AsCpuRegister()); - cfi_.AdjustCFAOffset(-static_cast<int>(kFramePointerSize)); - cfi_.Restore(DWARFReg(spill.AsCpuRegister().AsRegister())); - } - } - ret(); - // The CFI should be restored for any code that follows the exit block. - cfi_.RestoreState(); - cfi_.DefCFAOffset(frame_size); -} - -void X86_64Assembler::IncreaseFrameSize(size_t adjust) { - CHECK_ALIGNED(adjust, kStackAlignment); - addq(CpuRegister(RSP), Immediate(-static_cast<int64_t>(adjust))); - cfi_.AdjustCFAOffset(adjust); -} - -void X86_64Assembler::DecreaseFrameSize(size_t adjust) { - CHECK_ALIGNED(adjust, kStackAlignment); - addq(CpuRegister(RSP), Immediate(adjust)); - cfi_.AdjustCFAOffset(-adjust); -} - -void X86_64Assembler::Store(FrameOffset offs, ManagedRegister msrc, size_t size) { - X86_64ManagedRegister src = msrc.AsX86_64(); - if (src.IsNoRegister()) { - CHECK_EQ(0u, size); - } else if (src.IsCpuRegister()) { - if (size == 4) { - CHECK_EQ(4u, size); - movl(Address(CpuRegister(RSP), offs), src.AsCpuRegister()); - } else { - CHECK_EQ(8u, size); - movq(Address(CpuRegister(RSP), offs), src.AsCpuRegister()); - } - } else if (src.IsRegisterPair()) { - CHECK_EQ(0u, size); - movq(Address(CpuRegister(RSP), offs), src.AsRegisterPairLow()); - movq(Address(CpuRegister(RSP), FrameOffset(offs.Int32Value()+4)), - src.AsRegisterPairHigh()); - } else if (src.IsX87Register()) { - if (size == 4) { - fstps(Address(CpuRegister(RSP), offs)); - } else { - fstpl(Address(CpuRegister(RSP), offs)); - } - } else { - CHECK(src.IsXmmRegister()); - if (size == 4) { - movss(Address(CpuRegister(RSP), offs), src.AsXmmRegister()); - } else { - movsd(Address(CpuRegister(RSP), offs), src.AsXmmRegister()); - } - } -} - -void X86_64Assembler::StoreRef(FrameOffset dest, ManagedRegister msrc) { - X86_64ManagedRegister src = msrc.AsX86_64(); - CHECK(src.IsCpuRegister()); - movl(Address(CpuRegister(RSP), dest), src.AsCpuRegister()); -} - -void X86_64Assembler::StoreRawPtr(FrameOffset dest, ManagedRegister msrc) { - X86_64ManagedRegister src = msrc.AsX86_64(); - CHECK(src.IsCpuRegister()); - movq(Address(CpuRegister(RSP), dest), src.AsCpuRegister()); -} - -void X86_64Assembler::StoreImmediateToFrame(FrameOffset dest, uint32_t imm, - ManagedRegister) { - movl(Address(CpuRegister(RSP), dest), Immediate(imm)); // TODO(64) movq? -} - -void X86_64Assembler::StoreStackOffsetToThread(ThreadOffset64 thr_offs, - FrameOffset fr_offs, - ManagedRegister mscratch) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - CHECK(scratch.IsCpuRegister()); - leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), fr_offs)); - gs()->movq(Address::Absolute(thr_offs, true), scratch.AsCpuRegister()); -} - -void X86_64Assembler::StoreStackPointerToThread(ThreadOffset64 thr_offs) { - gs()->movq(Address::Absolute(thr_offs, true), CpuRegister(RSP)); -} - -void X86_64Assembler::StoreSpanning(FrameOffset /*dst*/, ManagedRegister /*src*/, - FrameOffset /*in_off*/, ManagedRegister /*scratch*/) { - UNIMPLEMENTED(FATAL); // this case only currently exists for ARM -} - -void X86_64Assembler::Load(ManagedRegister mdest, FrameOffset src, size_t size) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - if (dest.IsNoRegister()) { - CHECK_EQ(0u, size); - } else if (dest.IsCpuRegister()) { - if (size == 4) { - CHECK_EQ(4u, size); - movl(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); - } else { - CHECK_EQ(8u, size); - movq(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); - } - } else if (dest.IsRegisterPair()) { - CHECK_EQ(0u, size); - movq(dest.AsRegisterPairLow(), Address(CpuRegister(RSP), src)); - movq(dest.AsRegisterPairHigh(), Address(CpuRegister(RSP), FrameOffset(src.Int32Value()+4))); - } else if (dest.IsX87Register()) { - if (size == 4) { - flds(Address(CpuRegister(RSP), src)); - } else { - fldl(Address(CpuRegister(RSP), src)); - } - } else { - CHECK(dest.IsXmmRegister()); - if (size == 4) { - movss(dest.AsXmmRegister(), Address(CpuRegister(RSP), src)); - } else { - movsd(dest.AsXmmRegister(), Address(CpuRegister(RSP), src)); - } - } -} - -void X86_64Assembler::LoadFromThread(ManagedRegister mdest, ThreadOffset64 src, size_t size) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - if (dest.IsNoRegister()) { - CHECK_EQ(0u, size); - } else if (dest.IsCpuRegister()) { - CHECK_EQ(4u, size); - gs()->movl(dest.AsCpuRegister(), Address::Absolute(src, true)); - } else if (dest.IsRegisterPair()) { - CHECK_EQ(8u, size); - gs()->movq(dest.AsRegisterPairLow(), Address::Absolute(src, true)); - } else if (dest.IsX87Register()) { - if (size == 4) { - gs()->flds(Address::Absolute(src, true)); - } else { - gs()->fldl(Address::Absolute(src, true)); - } - } else { - CHECK(dest.IsXmmRegister()); - if (size == 4) { - gs()->movss(dest.AsXmmRegister(), Address::Absolute(src, true)); - } else { - gs()->movsd(dest.AsXmmRegister(), Address::Absolute(src, true)); - } - } -} - -void X86_64Assembler::LoadRef(ManagedRegister mdest, FrameOffset src) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - CHECK(dest.IsCpuRegister()); - movq(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); -} - -void X86_64Assembler::LoadRef(ManagedRegister mdest, ManagedRegister base, MemberOffset offs, - bool unpoison_reference) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - CHECK(dest.IsCpuRegister() && dest.IsCpuRegister()); - movl(dest.AsCpuRegister(), Address(base.AsX86_64().AsCpuRegister(), offs)); - if (unpoison_reference) { - MaybeUnpoisonHeapReference(dest.AsCpuRegister()); - } -} - -void X86_64Assembler::LoadRawPtr(ManagedRegister mdest, ManagedRegister base, - Offset offs) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - CHECK(dest.IsCpuRegister() && dest.IsCpuRegister()); - movq(dest.AsCpuRegister(), Address(base.AsX86_64().AsCpuRegister(), offs)); -} - -void X86_64Assembler::LoadRawPtrFromThread(ManagedRegister mdest, ThreadOffset64 offs) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - CHECK(dest.IsCpuRegister()); - gs()->movq(dest.AsCpuRegister(), Address::Absolute(offs, true)); -} - -void X86_64Assembler::SignExtend(ManagedRegister mreg, size_t size) { - X86_64ManagedRegister reg = mreg.AsX86_64(); - CHECK(size == 1 || size == 2) << size; - CHECK(reg.IsCpuRegister()) << reg; - if (size == 1) { - movsxb(reg.AsCpuRegister(), reg.AsCpuRegister()); - } else { - movsxw(reg.AsCpuRegister(), reg.AsCpuRegister()); - } -} - -void X86_64Assembler::ZeroExtend(ManagedRegister mreg, size_t size) { - X86_64ManagedRegister reg = mreg.AsX86_64(); - CHECK(size == 1 || size == 2) << size; - CHECK(reg.IsCpuRegister()) << reg; - if (size == 1) { - movzxb(reg.AsCpuRegister(), reg.AsCpuRegister()); - } else { - movzxw(reg.AsCpuRegister(), reg.AsCpuRegister()); - } -} - -void X86_64Assembler::Move(ManagedRegister mdest, ManagedRegister msrc, size_t size) { - X86_64ManagedRegister dest = mdest.AsX86_64(); - X86_64ManagedRegister src = msrc.AsX86_64(); - if (!dest.Equals(src)) { - if (dest.IsCpuRegister() && src.IsCpuRegister()) { - movq(dest.AsCpuRegister(), src.AsCpuRegister()); - } else if (src.IsX87Register() && dest.IsXmmRegister()) { - // Pass via stack and pop X87 register - subl(CpuRegister(RSP), Immediate(16)); - if (size == 4) { - CHECK_EQ(src.AsX87Register(), ST0); - fstps(Address(CpuRegister(RSP), 0)); - movss(dest.AsXmmRegister(), Address(CpuRegister(RSP), 0)); - } else { - CHECK_EQ(src.AsX87Register(), ST0); - fstpl(Address(CpuRegister(RSP), 0)); - movsd(dest.AsXmmRegister(), Address(CpuRegister(RSP), 0)); - } - addq(CpuRegister(RSP), Immediate(16)); - } else { - // TODO: x87, SSE - UNIMPLEMENTED(FATAL) << ": Move " << dest << ", " << src; - } - } -} - -void X86_64Assembler::CopyRef(FrameOffset dest, FrameOffset src, ManagedRegister mscratch) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - CHECK(scratch.IsCpuRegister()); - movl(scratch.AsCpuRegister(), Address(CpuRegister(RSP), src)); - movl(Address(CpuRegister(RSP), dest), scratch.AsCpuRegister()); -} - -void X86_64Assembler::CopyRawPtrFromThread(FrameOffset fr_offs, - ThreadOffset64 thr_offs, - ManagedRegister mscratch) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - CHECK(scratch.IsCpuRegister()); - gs()->movq(scratch.AsCpuRegister(), Address::Absolute(thr_offs, true)); - Store(fr_offs, scratch, 8); -} - -void X86_64Assembler::CopyRawPtrToThread(ThreadOffset64 thr_offs, - FrameOffset fr_offs, - ManagedRegister mscratch) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - CHECK(scratch.IsCpuRegister()); - Load(scratch, fr_offs, 8); - gs()->movq(Address::Absolute(thr_offs, true), scratch.AsCpuRegister()); -} - -void X86_64Assembler::Copy(FrameOffset dest, FrameOffset src, ManagedRegister mscratch, - size_t size) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - if (scratch.IsCpuRegister() && size == 8) { - Load(scratch, src, 4); - Store(dest, scratch, 4); - Load(scratch, FrameOffset(src.Int32Value() + 4), 4); - Store(FrameOffset(dest.Int32Value() + 4), scratch, 4); - } else { - Load(scratch, src, size); - Store(dest, scratch, size); - } -} - -void X86_64Assembler::Copy(FrameOffset /*dst*/, ManagedRegister /*src_base*/, Offset /*src_offset*/, - ManagedRegister /*scratch*/, size_t /*size*/) { - UNIMPLEMENTED(FATAL); -} - -void X86_64Assembler::Copy(ManagedRegister dest_base, Offset dest_offset, FrameOffset src, - ManagedRegister scratch, size_t size) { - CHECK(scratch.IsNoRegister()); - CHECK_EQ(size, 4u); - pushq(Address(CpuRegister(RSP), src)); - popq(Address(dest_base.AsX86_64().AsCpuRegister(), dest_offset)); -} - -void X86_64Assembler::Copy(FrameOffset dest, FrameOffset src_base, Offset src_offset, - ManagedRegister mscratch, size_t size) { - CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); - CHECK_EQ(size, 4u); - movq(scratch, Address(CpuRegister(RSP), src_base)); - movq(scratch, Address(scratch, src_offset)); - movq(Address(CpuRegister(RSP), dest), scratch); -} - -void X86_64Assembler::Copy(ManagedRegister dest, Offset dest_offset, - ManagedRegister src, Offset src_offset, - ManagedRegister scratch, size_t size) { - CHECK_EQ(size, 4u); - CHECK(scratch.IsNoRegister()); - pushq(Address(src.AsX86_64().AsCpuRegister(), src_offset)); - popq(Address(dest.AsX86_64().AsCpuRegister(), dest_offset)); -} - -void X86_64Assembler::Copy(FrameOffset dest, Offset dest_offset, FrameOffset src, Offset src_offset, - ManagedRegister mscratch, size_t size) { - CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); - CHECK_EQ(size, 4u); - CHECK_EQ(dest.Int32Value(), src.Int32Value()); - movq(scratch, Address(CpuRegister(RSP), src)); - pushq(Address(scratch, src_offset)); - popq(Address(scratch, dest_offset)); -} - -void X86_64Assembler::MemoryBarrier(ManagedRegister) { - mfence(); -} - -void X86_64Assembler::CreateHandleScopeEntry(ManagedRegister mout_reg, - FrameOffset handle_scope_offset, - ManagedRegister min_reg, bool null_allowed) { - X86_64ManagedRegister out_reg = mout_reg.AsX86_64(); - X86_64ManagedRegister in_reg = min_reg.AsX86_64(); - if (in_reg.IsNoRegister()) { // TODO(64): && null_allowed - // Use out_reg as indicator of null. - in_reg = out_reg; - // TODO: movzwl - movl(in_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - } - CHECK(in_reg.IsCpuRegister()); - CHECK(out_reg.IsCpuRegister()); - VerifyObject(in_reg, null_allowed); - if (null_allowed) { - Label null_arg; - if (!out_reg.Equals(in_reg)) { - xorl(out_reg.AsCpuRegister(), out_reg.AsCpuRegister()); - } - testl(in_reg.AsCpuRegister(), in_reg.AsCpuRegister()); - j(kZero, &null_arg); - leaq(out_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - Bind(&null_arg); - } else { - leaq(out_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - } -} - -void X86_64Assembler::CreateHandleScopeEntry(FrameOffset out_off, - FrameOffset handle_scope_offset, - ManagedRegister mscratch, - bool null_allowed) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - CHECK(scratch.IsCpuRegister()); - if (null_allowed) { - Label null_arg; - movl(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - testl(scratch.AsCpuRegister(), scratch.AsCpuRegister()); - j(kZero, &null_arg); - leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - Bind(&null_arg); - } else { - leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); - } - Store(out_off, scratch, 8); -} - -// Given a handle scope entry, load the associated reference. -void X86_64Assembler::LoadReferenceFromHandleScope(ManagedRegister mout_reg, - ManagedRegister min_reg) { - X86_64ManagedRegister out_reg = mout_reg.AsX86_64(); - X86_64ManagedRegister in_reg = min_reg.AsX86_64(); - CHECK(out_reg.IsCpuRegister()); - CHECK(in_reg.IsCpuRegister()); - Label null_arg; - if (!out_reg.Equals(in_reg)) { - xorl(out_reg.AsCpuRegister(), out_reg.AsCpuRegister()); - } - testl(in_reg.AsCpuRegister(), in_reg.AsCpuRegister()); - j(kZero, &null_arg); - movq(out_reg.AsCpuRegister(), Address(in_reg.AsCpuRegister(), 0)); - Bind(&null_arg); -} - -void X86_64Assembler::VerifyObject(ManagedRegister /*src*/, bool /*could_be_null*/) { - // TODO: not validating references -} - -void X86_64Assembler::VerifyObject(FrameOffset /*src*/, bool /*could_be_null*/) { - // TODO: not validating references -} - -void X86_64Assembler::Call(ManagedRegister mbase, Offset offset, ManagedRegister) { - X86_64ManagedRegister base = mbase.AsX86_64(); - CHECK(base.IsCpuRegister()); - call(Address(base.AsCpuRegister(), offset.Int32Value())); - // TODO: place reference map on call -} - -void X86_64Assembler::Call(FrameOffset base, Offset offset, ManagedRegister mscratch) { - CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); - movq(scratch, Address(CpuRegister(RSP), base)); - call(Address(scratch, offset)); -} - -void X86_64Assembler::CallFromThread(ThreadOffset64 offset, ManagedRegister /*mscratch*/) { - gs()->call(Address::Absolute(offset, true)); -} - -void X86_64Assembler::GetCurrentThread(ManagedRegister tr) { - gs()->movq(tr.AsX86_64().AsCpuRegister(), - Address::Absolute(Thread::SelfOffset<kX86_64PointerSize>(), true)); -} - -void X86_64Assembler::GetCurrentThread(FrameOffset offset, ManagedRegister mscratch) { - X86_64ManagedRegister scratch = mscratch.AsX86_64(); - gs()->movq(scratch.AsCpuRegister(), - Address::Absolute(Thread::SelfOffset<kX86_64PointerSize>(), true)); - movq(Address(CpuRegister(RSP), offset), scratch.AsCpuRegister()); -} - -// Slowpath entered when Thread::Current()->_exception is non-null -class X86_64ExceptionSlowPath FINAL : public SlowPath { - public: - explicit X86_64ExceptionSlowPath(size_t stack_adjust) : stack_adjust_(stack_adjust) {} - virtual void Emit(Assembler *sp_asm) OVERRIDE; - private: - const size_t stack_adjust_; -}; - -void X86_64Assembler::ExceptionPoll(ManagedRegister /*scratch*/, size_t stack_adjust) { - X86_64ExceptionSlowPath* slow = new (GetArena()) X86_64ExceptionSlowPath(stack_adjust); - buffer_.EnqueueSlowPath(slow); - gs()->cmpl(Address::Absolute(Thread::ExceptionOffset<kX86_64PointerSize>(), true), Immediate(0)); - j(kNotEqual, slow->Entry()); -} - -void X86_64ExceptionSlowPath::Emit(Assembler *sasm) { - X86_64Assembler* sp_asm = down_cast<X86_64Assembler*>(sasm); -#define __ sp_asm-> - __ Bind(&entry_); - // Note: the return value is dead - if (stack_adjust_ != 0) { // Fix up the frame. - __ DecreaseFrameSize(stack_adjust_); - } - // Pass exception as argument in RDI - __ gs()->movq(CpuRegister(RDI), - Address::Absolute(Thread::ExceptionOffset<kX86_64PointerSize>(), true)); - __ gs()->call( - Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86_64PointerSize, pDeliverException), true)); - // this call should never return - __ int3(); -#undef __ -} - void X86_64Assembler::AddConstantArea() { ArrayRef<const int32_t> area = constant_area_.GetBuffer(); for (size_t i = 0, e = area.size(); i < e; i++) { diff --git a/compiler/utils/x86_64/assembler_x86_64.h b/compiler/utils/x86_64/assembler_x86_64.h index d298da2e6c..370f49cb05 100644 --- a/compiler/utils/x86_64/assembler_x86_64.h +++ b/compiler/utils/x86_64/assembler_x86_64.h @@ -333,20 +333,11 @@ class NearLabel : private Label { }; -class X86_64Assembler FINAL : public Assembler, public JNIMacroAssembler<PointerSize::k64> { +class X86_64Assembler FINAL : public Assembler { public: explicit X86_64Assembler(ArenaAllocator* arena) : Assembler(arena), constant_area_(arena) {} virtual ~X86_64Assembler() {} - size_t CodeSize() const OVERRIDE { return Assembler::CodeSize(); } - DebugFrameOpCodeWriterForAssembler& cfi() { return Assembler::cfi(); } - void FinalizeCode() OVERRIDE { - Assembler::FinalizeCode(); - } - void FinalizeInstructions(const MemoryRegion& region) { - Assembler::FinalizeInstructions(region); - } - /* * Emit Machine Instructions. */ @@ -709,124 +700,6 @@ class X86_64Assembler FINAL : public Assembler, public JNIMacroAssembler<Pointer } void Bind(NearLabel* label); - // - // Overridden common assembler high-level functionality - // - - // Emit code that will create an activation on the stack - void BuildFrame(size_t frame_size, - ManagedRegister method_reg, - ArrayRef<const ManagedRegister> callee_save_regs, - const ManagedRegisterEntrySpills& entry_spills) OVERRIDE; - - // Emit code that will remove an activation from the stack - void RemoveFrame(size_t frame_size, ArrayRef<const ManagedRegister> callee_save_regs) - OVERRIDE; - - void IncreaseFrameSize(size_t adjust) OVERRIDE; - void DecreaseFrameSize(size_t adjust) OVERRIDE; - - // Store routines - void Store(FrameOffset offs, ManagedRegister src, size_t size) OVERRIDE; - void StoreRef(FrameOffset dest, ManagedRegister src) OVERRIDE; - void StoreRawPtr(FrameOffset dest, ManagedRegister src) OVERRIDE; - - void StoreImmediateToFrame(FrameOffset dest, uint32_t imm, ManagedRegister scratch) OVERRIDE; - - void StoreStackOffsetToThread(ThreadOffset64 thr_offs, - FrameOffset fr_offs, - ManagedRegister scratch) OVERRIDE; - - void StoreStackPointerToThread(ThreadOffset64 thr_offs) OVERRIDE; - - void StoreSpanning(FrameOffset dest, ManagedRegister src, FrameOffset in_off, - ManagedRegister scratch) OVERRIDE; - - // Load routines - void Load(ManagedRegister dest, FrameOffset src, size_t size) OVERRIDE; - - void LoadFromThread(ManagedRegister dest, ThreadOffset64 src, size_t size) OVERRIDE; - - void LoadRef(ManagedRegister dest, FrameOffset src) OVERRIDE; - - void LoadRef(ManagedRegister dest, ManagedRegister base, MemberOffset offs, - bool unpoison_reference) OVERRIDE; - - void LoadRawPtr(ManagedRegister dest, ManagedRegister base, Offset offs) OVERRIDE; - - void LoadRawPtrFromThread(ManagedRegister dest, ThreadOffset64 offs) OVERRIDE; - - // Copying routines - void Move(ManagedRegister dest, ManagedRegister src, size_t size); - - void CopyRawPtrFromThread(FrameOffset fr_offs, - ThreadOffset64 thr_offs, - ManagedRegister scratch) OVERRIDE; - - void CopyRawPtrToThread(ThreadOffset64 thr_offs, FrameOffset fr_offs, ManagedRegister scratch) - OVERRIDE; - - void CopyRef(FrameOffset dest, FrameOffset src, ManagedRegister scratch) OVERRIDE; - - void Copy(FrameOffset dest, FrameOffset src, ManagedRegister scratch, size_t size) OVERRIDE; - - void Copy(FrameOffset dest, ManagedRegister src_base, Offset src_offset, ManagedRegister scratch, - size_t size) OVERRIDE; - - void Copy(ManagedRegister dest_base, Offset dest_offset, FrameOffset src, ManagedRegister scratch, - size_t size) OVERRIDE; - - void Copy(FrameOffset dest, FrameOffset src_base, Offset src_offset, ManagedRegister scratch, - size_t size) OVERRIDE; - - void Copy(ManagedRegister dest, Offset dest_offset, ManagedRegister src, Offset src_offset, - ManagedRegister scratch, size_t size) OVERRIDE; - - void Copy(FrameOffset dest, Offset dest_offset, FrameOffset src, Offset src_offset, - ManagedRegister scratch, size_t size) OVERRIDE; - - void MemoryBarrier(ManagedRegister) OVERRIDE; - - // Sign extension - void SignExtend(ManagedRegister mreg, size_t size) OVERRIDE; - - // Zero extension - void ZeroExtend(ManagedRegister mreg, size_t size) OVERRIDE; - - // Exploit fast access in managed code to Thread::Current() - void GetCurrentThread(ManagedRegister tr) OVERRIDE; - void GetCurrentThread(FrameOffset dest_offset, ManagedRegister scratch) OVERRIDE; - - // Set up out_reg to hold a Object** into the handle scope, or to be null if the - // value is null and null_allowed. in_reg holds a possibly stale reference - // that can be used to avoid loading the handle scope entry to see if the value is - // null. - void CreateHandleScopeEntry(ManagedRegister out_reg, FrameOffset handlescope_offset, - ManagedRegister in_reg, bool null_allowed) OVERRIDE; - - // Set up out_off to hold a Object** into the handle scope, or to be null if the - // value is null and null_allowed. - void CreateHandleScopeEntry(FrameOffset out_off, FrameOffset handlescope_offset, - ManagedRegister scratch, bool null_allowed) OVERRIDE; - - // src holds a handle scope entry (Object**) load this into dst - virtual void LoadReferenceFromHandleScope(ManagedRegister dst, - ManagedRegister src); - - // Heap::VerifyObject on src. In some cases (such as a reference to this) we - // know that src may not be null. - void VerifyObject(ManagedRegister src, bool could_be_null) OVERRIDE; - void VerifyObject(FrameOffset src, bool could_be_null) OVERRIDE; - - // Call to address held at [base+offset] - void Call(ManagedRegister base, Offset offset, ManagedRegister scratch) OVERRIDE; - void Call(FrameOffset base, Offset offset, ManagedRegister scratch) OVERRIDE; - void CallFromThread(ThreadOffset64 offset, ManagedRegister scratch) OVERRIDE; - - // Generate code to check if Thread::Current()->exception_ is non-null - // and branch to a ExceptionSlowPath if it is. - void ExceptionPoll(ManagedRegister scratch, size_t stack_adjust) OVERRIDE; - // Add a double to the constant area, returning the offset into // the constant area where the literal resides. size_t AddDouble(double v) { return constant_area_.AddDouble(v); } diff --git a/compiler/utils/x86_64/assembler_x86_64_test.cc b/compiler/utils/x86_64/assembler_x86_64_test.cc index 788c7253cf..36c966b3cf 100644 --- a/compiler/utils/x86_64/assembler_x86_64_test.cc +++ b/compiler/utils/x86_64/assembler_x86_64_test.cc @@ -22,7 +22,9 @@ #include "base/bit_utils.h" #include "base/stl_util.h" +#include "jni_macro_assembler_x86_64.h" #include "utils/assembler_test.h" +#include "utils/jni_macro_assembler_test.h" namespace art { @@ -1485,6 +1487,62 @@ TEST_F(AssemblerX86_64Test, SetCC) { DriverFn(&setcc_test_fn, "setcc"); } +TEST_F(AssemblerX86_64Test, MovzxbRegs) { + DriverStr(Repeatrb(&x86_64::X86_64Assembler::movzxb, "movzbl %{reg2}, %{reg1}"), "movzxb"); +} + +TEST_F(AssemblerX86_64Test, MovsxbRegs) { + DriverStr(Repeatrb(&x86_64::X86_64Assembler::movsxb, "movsbl %{reg2}, %{reg1}"), "movsxb"); +} + +TEST_F(AssemblerX86_64Test, Repnescasw) { + GetAssembler()->repne_scasw(); + const char* expected = "repne scasw\n"; + DriverStr(expected, "Repnescasw"); +} + +TEST_F(AssemblerX86_64Test, Repecmpsw) { + GetAssembler()->repe_cmpsw(); + const char* expected = "repe cmpsw\n"; + DriverStr(expected, "Repecmpsw"); +} + +TEST_F(AssemblerX86_64Test, Repecmpsl) { + GetAssembler()->repe_cmpsl(); + const char* expected = "repe cmpsl\n"; + DriverStr(expected, "Repecmpsl"); +} + +TEST_F(AssemblerX86_64Test, Repecmpsq) { + GetAssembler()->repe_cmpsq(); + const char* expected = "repe cmpsq\n"; + DriverStr(expected, "Repecmpsq"); +} + +TEST_F(AssemblerX86_64Test, Cmpb) { + GetAssembler()->cmpb(x86_64::Address(x86_64::CpuRegister(x86_64::RDI), 128), + x86_64::Immediate(0)); + const char* expected = "cmpb $0, 128(%RDI)\n"; + DriverStr(expected, "cmpb"); +} + +class JNIMacroAssemblerX86_64Test : public JNIMacroAssemblerTest<x86_64::X86_64JNIMacroAssembler> { + public: + using Base = JNIMacroAssemblerTest<x86_64::X86_64JNIMacroAssembler>; + + protected: + // Get the typically used name for this architecture, e.g., aarch64, x86-64, ... + std::string GetArchitectureString() OVERRIDE { + return "x86_64"; + } + + std::string GetDisassembleParameters() OVERRIDE { + return " -D -bbinary -mi386:x86-64 -Mx86-64,addr64,data32 --no-show-raw-insn"; + } + + private: +}; + static x86_64::X86_64ManagedRegister ManagedFromCpu(x86_64::Register r) { return x86_64::X86_64ManagedRegister::FromCpuRegister(r); } @@ -1493,8 +1551,8 @@ static x86_64::X86_64ManagedRegister ManagedFromFpu(x86_64::FloatRegister r) { return x86_64::X86_64ManagedRegister::FromXmmRegister(r); } -std::string buildframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, - x86_64::X86_64Assembler* assembler) { +std::string buildframe_test_fn(JNIMacroAssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, + x86_64::X86_64JNIMacroAssembler* assembler) { // TODO: more interesting spill registers / entry spills. // Two random spill regs. @@ -1536,12 +1594,12 @@ std::string buildframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBU return str.str(); } -TEST_F(AssemblerX86_64Test, BuildFrame) { +TEST_F(JNIMacroAssemblerX86_64Test, BuildFrame) { DriverFn(&buildframe_test_fn, "BuildFrame"); } -std::string removeframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, - x86_64::X86_64Assembler* assembler) { +std::string removeframe_test_fn(JNIMacroAssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, + x86_64::X86_64JNIMacroAssembler* assembler) { // TODO: more interesting spill registers / entry spills. // Two random spill regs. @@ -1567,12 +1625,13 @@ std::string removeframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIB return str.str(); } -TEST_F(AssemblerX86_64Test, RemoveFrame) { +TEST_F(JNIMacroAssemblerX86_64Test, RemoveFrame) { DriverFn(&removeframe_test_fn, "RemoveFrame"); } -std::string increaseframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, - x86_64::X86_64Assembler* assembler) { +std::string increaseframe_test_fn( + JNIMacroAssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, + x86_64::X86_64JNIMacroAssembler* assembler) { assembler->IncreaseFrameSize(0U); assembler->IncreaseFrameSize(kStackAlignment); assembler->IncreaseFrameSize(10 * kStackAlignment); @@ -1586,12 +1645,13 @@ std::string increaseframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTR return str.str(); } -TEST_F(AssemblerX86_64Test, IncreaseFrame) { +TEST_F(JNIMacroAssemblerX86_64Test, IncreaseFrame) { DriverFn(&increaseframe_test_fn, "IncreaseFrame"); } -std::string decreaseframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, - x86_64::X86_64Assembler* assembler) { +std::string decreaseframe_test_fn( + JNIMacroAssemblerX86_64Test::Base* assembler_test ATTRIBUTE_UNUSED, + x86_64::X86_64JNIMacroAssembler* assembler) { assembler->DecreaseFrameSize(0U); assembler->DecreaseFrameSize(kStackAlignment); assembler->DecreaseFrameSize(10 * kStackAlignment); @@ -1605,47 +1665,8 @@ std::string decreaseframe_test_fn(AssemblerX86_64Test::Base* assembler_test ATTR return str.str(); } -TEST_F(AssemblerX86_64Test, DecreaseFrame) { +TEST_F(JNIMacroAssemblerX86_64Test, DecreaseFrame) { DriverFn(&decreaseframe_test_fn, "DecreaseFrame"); } -TEST_F(AssemblerX86_64Test, MovzxbRegs) { - DriverStr(Repeatrb(&x86_64::X86_64Assembler::movzxb, "movzbl %{reg2}, %{reg1}"), "movzxb"); -} - -TEST_F(AssemblerX86_64Test, MovsxbRegs) { - DriverStr(Repeatrb(&x86_64::X86_64Assembler::movsxb, "movsbl %{reg2}, %{reg1}"), "movsxb"); -} - -TEST_F(AssemblerX86_64Test, Repnescasw) { - GetAssembler()->repne_scasw(); - const char* expected = "repne scasw\n"; - DriverStr(expected, "Repnescasw"); -} - -TEST_F(AssemblerX86_64Test, Repecmpsw) { - GetAssembler()->repe_cmpsw(); - const char* expected = "repe cmpsw\n"; - DriverStr(expected, "Repecmpsw"); -} - -TEST_F(AssemblerX86_64Test, Repecmpsl) { - GetAssembler()->repe_cmpsl(); - const char* expected = "repe cmpsl\n"; - DriverStr(expected, "Repecmpsl"); -} - -TEST_F(AssemblerX86_64Test, Repecmpsq) { - GetAssembler()->repe_cmpsq(); - const char* expected = "repe cmpsq\n"; - DriverStr(expected, "Repecmpsq"); -} - -TEST_F(AssemblerX86_64Test, Cmpb) { - GetAssembler()->cmpb(x86_64::Address(x86_64::CpuRegister(x86_64::RDI), 128), - x86_64::Immediate(0)); - const char* expected = "cmpb $0, 128(%RDI)\n"; - DriverStr(expected, "cmpb"); -} - } // namespace art diff --git a/compiler/utils/x86_64/jni_macro_assembler_x86_64.cc b/compiler/utils/x86_64/jni_macro_assembler_x86_64.cc new file mode 100644 index 0000000000..47fb59b1d8 --- /dev/null +++ b/compiler/utils/x86_64/jni_macro_assembler_x86_64.cc @@ -0,0 +1,603 @@ +/* + * Copyright (C) 2016 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "jni_macro_assembler_x86_64.h" + +#include "base/casts.h" +#include "entrypoints/quick/quick_entrypoints.h" +#include "memory_region.h" +#include "thread.h" + +namespace art { +namespace x86_64 { + +static dwarf::Reg DWARFReg(Register reg) { + return dwarf::Reg::X86_64Core(static_cast<int>(reg)); +} +static dwarf::Reg DWARFReg(FloatRegister reg) { + return dwarf::Reg::X86_64Fp(static_cast<int>(reg)); +} + +constexpr size_t kFramePointerSize = 8; + +#define __ asm_. + +void X86_64JNIMacroAssembler::BuildFrame(size_t frame_size, + ManagedRegister method_reg, + ArrayRef<const ManagedRegister> spill_regs, + const ManagedRegisterEntrySpills& entry_spills) { + DCHECK_EQ(CodeSize(), 0U); // Nothing emitted yet. + cfi().SetCurrentCFAOffset(8); // Return address on stack. + CHECK_ALIGNED(frame_size, kStackAlignment); + int gpr_count = 0; + for (int i = spill_regs.size() - 1; i >= 0; --i) { + x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); + if (spill.IsCpuRegister()) { + __ pushq(spill.AsCpuRegister()); + gpr_count++; + cfi().AdjustCFAOffset(kFramePointerSize); + cfi().RelOffset(DWARFReg(spill.AsCpuRegister().AsRegister()), 0); + } + } + // return address then method on stack. + int64_t rest_of_frame = static_cast<int64_t>(frame_size) + - (gpr_count * kFramePointerSize) + - kFramePointerSize /*return address*/; + __ subq(CpuRegister(RSP), Immediate(rest_of_frame)); + cfi().AdjustCFAOffset(rest_of_frame); + + // spill xmms + int64_t offset = rest_of_frame; + for (int i = spill_regs.size() - 1; i >= 0; --i) { + x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); + if (spill.IsXmmRegister()) { + offset -= sizeof(double); + __ movsd(Address(CpuRegister(RSP), offset), spill.AsXmmRegister()); + cfi().RelOffset(DWARFReg(spill.AsXmmRegister().AsFloatRegister()), offset); + } + } + + static_assert(static_cast<size_t>(kX86_64PointerSize) == kFramePointerSize, + "Unexpected frame pointer size."); + + __ movq(Address(CpuRegister(RSP), 0), method_reg.AsX86_64().AsCpuRegister()); + + for (size_t i = 0; i < entry_spills.size(); ++i) { + ManagedRegisterSpill spill = entry_spills.at(i); + if (spill.AsX86_64().IsCpuRegister()) { + if (spill.getSize() == 8) { + __ movq(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), + spill.AsX86_64().AsCpuRegister()); + } else { + CHECK_EQ(spill.getSize(), 4); + __ movl(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), + spill.AsX86_64().AsCpuRegister()); + } + } else { + if (spill.getSize() == 8) { + __ movsd(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), + spill.AsX86_64().AsXmmRegister()); + } else { + CHECK_EQ(spill.getSize(), 4); + __ movss(Address(CpuRegister(RSP), frame_size + spill.getSpillOffset()), + spill.AsX86_64().AsXmmRegister()); + } + } + } +} + +void X86_64JNIMacroAssembler::RemoveFrame(size_t frame_size, + ArrayRef<const ManagedRegister> spill_regs) { + CHECK_ALIGNED(frame_size, kStackAlignment); + cfi().RememberState(); + int gpr_count = 0; + // unspill xmms + int64_t offset = static_cast<int64_t>(frame_size) + - (spill_regs.size() * kFramePointerSize) + - 2 * kFramePointerSize; + for (size_t i = 0; i < spill_regs.size(); ++i) { + x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); + if (spill.IsXmmRegister()) { + offset += sizeof(double); + __ movsd(spill.AsXmmRegister(), Address(CpuRegister(RSP), offset)); + cfi().Restore(DWARFReg(spill.AsXmmRegister().AsFloatRegister())); + } else { + gpr_count++; + } + } + int adjust = static_cast<int>(frame_size) - (gpr_count * kFramePointerSize) - kFramePointerSize; + __ addq(CpuRegister(RSP), Immediate(adjust)); + cfi().AdjustCFAOffset(-adjust); + for (size_t i = 0; i < spill_regs.size(); ++i) { + x86_64::X86_64ManagedRegister spill = spill_regs[i].AsX86_64(); + if (spill.IsCpuRegister()) { + __ popq(spill.AsCpuRegister()); + cfi().AdjustCFAOffset(-static_cast<int>(kFramePointerSize)); + cfi().Restore(DWARFReg(spill.AsCpuRegister().AsRegister())); + } + } + __ ret(); + // The CFI should be restored for any code that follows the exit block. + cfi().RestoreState(); + cfi().DefCFAOffset(frame_size); +} + +void X86_64JNIMacroAssembler::IncreaseFrameSize(size_t adjust) { + CHECK_ALIGNED(adjust, kStackAlignment); + __ addq(CpuRegister(RSP), Immediate(-static_cast<int64_t>(adjust))); + cfi().AdjustCFAOffset(adjust); +} + +static void DecreaseFrameSizeImpl(size_t adjust, X86_64Assembler* assembler) { + CHECK_ALIGNED(adjust, kStackAlignment); + assembler->addq(CpuRegister(RSP), Immediate(adjust)); + assembler->cfi().AdjustCFAOffset(-adjust); +} + +void X86_64JNIMacroAssembler::DecreaseFrameSize(size_t adjust) { + DecreaseFrameSizeImpl(adjust, &asm_); +} + +void X86_64JNIMacroAssembler::Store(FrameOffset offs, ManagedRegister msrc, size_t size) { + X86_64ManagedRegister src = msrc.AsX86_64(); + if (src.IsNoRegister()) { + CHECK_EQ(0u, size); + } else if (src.IsCpuRegister()) { + if (size == 4) { + CHECK_EQ(4u, size); + __ movl(Address(CpuRegister(RSP), offs), src.AsCpuRegister()); + } else { + CHECK_EQ(8u, size); + __ movq(Address(CpuRegister(RSP), offs), src.AsCpuRegister()); + } + } else if (src.IsRegisterPair()) { + CHECK_EQ(0u, size); + __ movq(Address(CpuRegister(RSP), offs), src.AsRegisterPairLow()); + __ movq(Address(CpuRegister(RSP), FrameOffset(offs.Int32Value()+4)), + src.AsRegisterPairHigh()); + } else if (src.IsX87Register()) { + if (size == 4) { + __ fstps(Address(CpuRegister(RSP), offs)); + } else { + __ fstpl(Address(CpuRegister(RSP), offs)); + } + } else { + CHECK(src.IsXmmRegister()); + if (size == 4) { + __ movss(Address(CpuRegister(RSP), offs), src.AsXmmRegister()); + } else { + __ movsd(Address(CpuRegister(RSP), offs), src.AsXmmRegister()); + } + } +} + +void X86_64JNIMacroAssembler::StoreRef(FrameOffset dest, ManagedRegister msrc) { + X86_64ManagedRegister src = msrc.AsX86_64(); + CHECK(src.IsCpuRegister()); + __ movl(Address(CpuRegister(RSP), dest), src.AsCpuRegister()); +} + +void X86_64JNIMacroAssembler::StoreRawPtr(FrameOffset dest, ManagedRegister msrc) { + X86_64ManagedRegister src = msrc.AsX86_64(); + CHECK(src.IsCpuRegister()); + __ movq(Address(CpuRegister(RSP), dest), src.AsCpuRegister()); +} + +void X86_64JNIMacroAssembler::StoreImmediateToFrame(FrameOffset dest, + uint32_t imm, + ManagedRegister) { + __ movl(Address(CpuRegister(RSP), dest), Immediate(imm)); // TODO(64) movq? +} + +void X86_64JNIMacroAssembler::StoreStackOffsetToThread(ThreadOffset64 thr_offs, + FrameOffset fr_offs, + ManagedRegister mscratch) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + CHECK(scratch.IsCpuRegister()); + __ leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), fr_offs)); + __ gs()->movq(Address::Absolute(thr_offs, true), scratch.AsCpuRegister()); +} + +void X86_64JNIMacroAssembler::StoreStackPointerToThread(ThreadOffset64 thr_offs) { + __ gs()->movq(Address::Absolute(thr_offs, true), CpuRegister(RSP)); +} + +void X86_64JNIMacroAssembler::StoreSpanning(FrameOffset /*dst*/, + ManagedRegister /*src*/, + FrameOffset /*in_off*/, + ManagedRegister /*scratch*/) { + UNIMPLEMENTED(FATAL); // this case only currently exists for ARM +} + +void X86_64JNIMacroAssembler::Load(ManagedRegister mdest, FrameOffset src, size_t size) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + if (dest.IsNoRegister()) { + CHECK_EQ(0u, size); + } else if (dest.IsCpuRegister()) { + if (size == 4) { + CHECK_EQ(4u, size); + __ movl(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); + } else { + CHECK_EQ(8u, size); + __ movq(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); + } + } else if (dest.IsRegisterPair()) { + CHECK_EQ(0u, size); + __ movq(dest.AsRegisterPairLow(), Address(CpuRegister(RSP), src)); + __ movq(dest.AsRegisterPairHigh(), Address(CpuRegister(RSP), FrameOffset(src.Int32Value()+4))); + } else if (dest.IsX87Register()) { + if (size == 4) { + __ flds(Address(CpuRegister(RSP), src)); + } else { + __ fldl(Address(CpuRegister(RSP), src)); + } + } else { + CHECK(dest.IsXmmRegister()); + if (size == 4) { + __ movss(dest.AsXmmRegister(), Address(CpuRegister(RSP), src)); + } else { + __ movsd(dest.AsXmmRegister(), Address(CpuRegister(RSP), src)); + } + } +} + +void X86_64JNIMacroAssembler::LoadFromThread(ManagedRegister mdest, + ThreadOffset64 src, size_t size) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + if (dest.IsNoRegister()) { + CHECK_EQ(0u, size); + } else if (dest.IsCpuRegister()) { + CHECK_EQ(4u, size); + __ gs()->movl(dest.AsCpuRegister(), Address::Absolute(src, true)); + } else if (dest.IsRegisterPair()) { + CHECK_EQ(8u, size); + __ gs()->movq(dest.AsRegisterPairLow(), Address::Absolute(src, true)); + } else if (dest.IsX87Register()) { + if (size == 4) { + __ gs()->flds(Address::Absolute(src, true)); + } else { + __ gs()->fldl(Address::Absolute(src, true)); + } + } else { + CHECK(dest.IsXmmRegister()); + if (size == 4) { + __ gs()->movss(dest.AsXmmRegister(), Address::Absolute(src, true)); + } else { + __ gs()->movsd(dest.AsXmmRegister(), Address::Absolute(src, true)); + } + } +} + +void X86_64JNIMacroAssembler::LoadRef(ManagedRegister mdest, FrameOffset src) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + CHECK(dest.IsCpuRegister()); + __ movq(dest.AsCpuRegister(), Address(CpuRegister(RSP), src)); +} + +void X86_64JNIMacroAssembler::LoadRef(ManagedRegister mdest, + ManagedRegister base, + MemberOffset offs, + bool unpoison_reference) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + CHECK(dest.IsCpuRegister() && dest.IsCpuRegister()); + __ movl(dest.AsCpuRegister(), Address(base.AsX86_64().AsCpuRegister(), offs)); + if (unpoison_reference) { + __ MaybeUnpoisonHeapReference(dest.AsCpuRegister()); + } +} + +void X86_64JNIMacroAssembler::LoadRawPtr(ManagedRegister mdest, ManagedRegister base, Offset offs) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + CHECK(dest.IsCpuRegister() && dest.IsCpuRegister()); + __ movq(dest.AsCpuRegister(), Address(base.AsX86_64().AsCpuRegister(), offs)); +} + +void X86_64JNIMacroAssembler::LoadRawPtrFromThread(ManagedRegister mdest, ThreadOffset64 offs) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + CHECK(dest.IsCpuRegister()); + __ gs()->movq(dest.AsCpuRegister(), Address::Absolute(offs, true)); +} + +void X86_64JNIMacroAssembler::SignExtend(ManagedRegister mreg, size_t size) { + X86_64ManagedRegister reg = mreg.AsX86_64(); + CHECK(size == 1 || size == 2) << size; + CHECK(reg.IsCpuRegister()) << reg; + if (size == 1) { + __ movsxb(reg.AsCpuRegister(), reg.AsCpuRegister()); + } else { + __ movsxw(reg.AsCpuRegister(), reg.AsCpuRegister()); + } +} + +void X86_64JNIMacroAssembler::ZeroExtend(ManagedRegister mreg, size_t size) { + X86_64ManagedRegister reg = mreg.AsX86_64(); + CHECK(size == 1 || size == 2) << size; + CHECK(reg.IsCpuRegister()) << reg; + if (size == 1) { + __ movzxb(reg.AsCpuRegister(), reg.AsCpuRegister()); + } else { + __ movzxw(reg.AsCpuRegister(), reg.AsCpuRegister()); + } +} + +void X86_64JNIMacroAssembler::Move(ManagedRegister mdest, ManagedRegister msrc, size_t size) { + X86_64ManagedRegister dest = mdest.AsX86_64(); + X86_64ManagedRegister src = msrc.AsX86_64(); + if (!dest.Equals(src)) { + if (dest.IsCpuRegister() && src.IsCpuRegister()) { + __ movq(dest.AsCpuRegister(), src.AsCpuRegister()); + } else if (src.IsX87Register() && dest.IsXmmRegister()) { + // Pass via stack and pop X87 register + __ subl(CpuRegister(RSP), Immediate(16)); + if (size == 4) { + CHECK_EQ(src.AsX87Register(), ST0); + __ fstps(Address(CpuRegister(RSP), 0)); + __ movss(dest.AsXmmRegister(), Address(CpuRegister(RSP), 0)); + } else { + CHECK_EQ(src.AsX87Register(), ST0); + __ fstpl(Address(CpuRegister(RSP), 0)); + __ movsd(dest.AsXmmRegister(), Address(CpuRegister(RSP), 0)); + } + __ addq(CpuRegister(RSP), Immediate(16)); + } else { + // TODO: x87, SSE + UNIMPLEMENTED(FATAL) << ": Move " << dest << ", " << src; + } + } +} + +void X86_64JNIMacroAssembler::CopyRef(FrameOffset dest, FrameOffset src, ManagedRegister mscratch) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + CHECK(scratch.IsCpuRegister()); + __ movl(scratch.AsCpuRegister(), Address(CpuRegister(RSP), src)); + __ movl(Address(CpuRegister(RSP), dest), scratch.AsCpuRegister()); +} + +void X86_64JNIMacroAssembler::CopyRawPtrFromThread(FrameOffset fr_offs, + ThreadOffset64 thr_offs, + ManagedRegister mscratch) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + CHECK(scratch.IsCpuRegister()); + __ gs()->movq(scratch.AsCpuRegister(), Address::Absolute(thr_offs, true)); + Store(fr_offs, scratch, 8); +} + +void X86_64JNIMacroAssembler::CopyRawPtrToThread(ThreadOffset64 thr_offs, + FrameOffset fr_offs, + ManagedRegister mscratch) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + CHECK(scratch.IsCpuRegister()); + Load(scratch, fr_offs, 8); + __ gs()->movq(Address::Absolute(thr_offs, true), scratch.AsCpuRegister()); +} + +void X86_64JNIMacroAssembler::Copy(FrameOffset dest, + FrameOffset src, + ManagedRegister mscratch, + size_t size) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + if (scratch.IsCpuRegister() && size == 8) { + Load(scratch, src, 4); + Store(dest, scratch, 4); + Load(scratch, FrameOffset(src.Int32Value() + 4), 4); + Store(FrameOffset(dest.Int32Value() + 4), scratch, 4); + } else { + Load(scratch, src, size); + Store(dest, scratch, size); + } +} + +void X86_64JNIMacroAssembler::Copy(FrameOffset /*dst*/, + ManagedRegister /*src_base*/, + Offset /*src_offset*/, + ManagedRegister /*scratch*/, + size_t /*size*/) { + UNIMPLEMENTED(FATAL); +} + +void X86_64JNIMacroAssembler::Copy(ManagedRegister dest_base, + Offset dest_offset, + FrameOffset src, + ManagedRegister scratch, + size_t size) { + CHECK(scratch.IsNoRegister()); + CHECK_EQ(size, 4u); + __ pushq(Address(CpuRegister(RSP), src)); + __ popq(Address(dest_base.AsX86_64().AsCpuRegister(), dest_offset)); +} + +void X86_64JNIMacroAssembler::Copy(FrameOffset dest, + FrameOffset src_base, + Offset src_offset, + ManagedRegister mscratch, + size_t size) { + CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); + CHECK_EQ(size, 4u); + __ movq(scratch, Address(CpuRegister(RSP), src_base)); + __ movq(scratch, Address(scratch, src_offset)); + __ movq(Address(CpuRegister(RSP), dest), scratch); +} + +void X86_64JNIMacroAssembler::Copy(ManagedRegister dest, + Offset dest_offset, + ManagedRegister src, + Offset src_offset, + ManagedRegister scratch, + size_t size) { + CHECK_EQ(size, 4u); + CHECK(scratch.IsNoRegister()); + __ pushq(Address(src.AsX86_64().AsCpuRegister(), src_offset)); + __ popq(Address(dest.AsX86_64().AsCpuRegister(), dest_offset)); +} + +void X86_64JNIMacroAssembler::Copy(FrameOffset dest, + Offset dest_offset, + FrameOffset src, + Offset src_offset, + ManagedRegister mscratch, + size_t size) { + CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); + CHECK_EQ(size, 4u); + CHECK_EQ(dest.Int32Value(), src.Int32Value()); + __ movq(scratch, Address(CpuRegister(RSP), src)); + __ pushq(Address(scratch, src_offset)); + __ popq(Address(scratch, dest_offset)); +} + +void X86_64JNIMacroAssembler::MemoryBarrier(ManagedRegister) { + __ mfence(); +} + +void X86_64JNIMacroAssembler::CreateHandleScopeEntry(ManagedRegister mout_reg, + FrameOffset handle_scope_offset, + ManagedRegister min_reg, + bool null_allowed) { + X86_64ManagedRegister out_reg = mout_reg.AsX86_64(); + X86_64ManagedRegister in_reg = min_reg.AsX86_64(); + if (in_reg.IsNoRegister()) { // TODO(64): && null_allowed + // Use out_reg as indicator of null. + in_reg = out_reg; + // TODO: movzwl + __ movl(in_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + } + CHECK(in_reg.IsCpuRegister()); + CHECK(out_reg.IsCpuRegister()); + VerifyObject(in_reg, null_allowed); + if (null_allowed) { + Label null_arg; + if (!out_reg.Equals(in_reg)) { + __ xorl(out_reg.AsCpuRegister(), out_reg.AsCpuRegister()); + } + __ testl(in_reg.AsCpuRegister(), in_reg.AsCpuRegister()); + __ j(kZero, &null_arg); + __ leaq(out_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + __ Bind(&null_arg); + } else { + __ leaq(out_reg.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + } +} + +void X86_64JNIMacroAssembler::CreateHandleScopeEntry(FrameOffset out_off, + FrameOffset handle_scope_offset, + ManagedRegister mscratch, + bool null_allowed) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + CHECK(scratch.IsCpuRegister()); + if (null_allowed) { + Label null_arg; + __ movl(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + __ testl(scratch.AsCpuRegister(), scratch.AsCpuRegister()); + __ j(kZero, &null_arg); + __ leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + __ Bind(&null_arg); + } else { + __ leaq(scratch.AsCpuRegister(), Address(CpuRegister(RSP), handle_scope_offset)); + } + Store(out_off, scratch, 8); +} + +// Given a handle scope entry, load the associated reference. +void X86_64JNIMacroAssembler::LoadReferenceFromHandleScope(ManagedRegister mout_reg, + ManagedRegister min_reg) { + X86_64ManagedRegister out_reg = mout_reg.AsX86_64(); + X86_64ManagedRegister in_reg = min_reg.AsX86_64(); + CHECK(out_reg.IsCpuRegister()); + CHECK(in_reg.IsCpuRegister()); + Label null_arg; + if (!out_reg.Equals(in_reg)) { + __ xorl(out_reg.AsCpuRegister(), out_reg.AsCpuRegister()); + } + __ testl(in_reg.AsCpuRegister(), in_reg.AsCpuRegister()); + __ j(kZero, &null_arg); + __ movq(out_reg.AsCpuRegister(), Address(in_reg.AsCpuRegister(), 0)); + __ Bind(&null_arg); +} + +void X86_64JNIMacroAssembler::VerifyObject(ManagedRegister /*src*/, bool /*could_be_null*/) { + // TODO: not validating references +} + +void X86_64JNIMacroAssembler::VerifyObject(FrameOffset /*src*/, bool /*could_be_null*/) { + // TODO: not validating references +} + +void X86_64JNIMacroAssembler::Call(ManagedRegister mbase, Offset offset, ManagedRegister) { + X86_64ManagedRegister base = mbase.AsX86_64(); + CHECK(base.IsCpuRegister()); + __ call(Address(base.AsCpuRegister(), offset.Int32Value())); + // TODO: place reference map on call +} + +void X86_64JNIMacroAssembler::Call(FrameOffset base, Offset offset, ManagedRegister mscratch) { + CpuRegister scratch = mscratch.AsX86_64().AsCpuRegister(); + __ movq(scratch, Address(CpuRegister(RSP), base)); + __ call(Address(scratch, offset)); +} + +void X86_64JNIMacroAssembler::CallFromThread(ThreadOffset64 offset, ManagedRegister /*mscratch*/) { + __ gs()->call(Address::Absolute(offset, true)); +} + +void X86_64JNIMacroAssembler::GetCurrentThread(ManagedRegister tr) { + __ gs()->movq(tr.AsX86_64().AsCpuRegister(), + Address::Absolute(Thread::SelfOffset<kX86_64PointerSize>(), true)); +} + +void X86_64JNIMacroAssembler::GetCurrentThread(FrameOffset offset, ManagedRegister mscratch) { + X86_64ManagedRegister scratch = mscratch.AsX86_64(); + __ gs()->movq(scratch.AsCpuRegister(), + Address::Absolute(Thread::SelfOffset<kX86_64PointerSize>(), true)); + __ movq(Address(CpuRegister(RSP), offset), scratch.AsCpuRegister()); +} + +// Slowpath entered when Thread::Current()->_exception is non-null +class X86_64ExceptionSlowPath FINAL : public SlowPath { + public: + explicit X86_64ExceptionSlowPath(size_t stack_adjust) : stack_adjust_(stack_adjust) {} + virtual void Emit(Assembler *sp_asm) OVERRIDE; + private: + const size_t stack_adjust_; +}; + +void X86_64JNIMacroAssembler::ExceptionPoll(ManagedRegister /*scratch*/, size_t stack_adjust) { + X86_64ExceptionSlowPath* slow = new (__ GetArena()) X86_64ExceptionSlowPath(stack_adjust); + __ GetBuffer()->EnqueueSlowPath(slow); + __ gs()->cmpl(Address::Absolute(Thread::ExceptionOffset<kX86_64PointerSize>(), true), Immediate(0)); + __ j(kNotEqual, slow->Entry()); +} + +#undef __ + +void X86_64ExceptionSlowPath::Emit(Assembler *sasm) { + X86_64Assembler* sp_asm = down_cast<X86_64Assembler*>(sasm); +#define __ sp_asm-> + __ Bind(&entry_); + // Note: the return value is dead + if (stack_adjust_ != 0) { // Fix up the frame. + DecreaseFrameSizeImpl(stack_adjust_, sp_asm); + } + // Pass exception as argument in RDI + __ gs()->movq(CpuRegister(RDI), + Address::Absolute(Thread::ExceptionOffset<kX86_64PointerSize>(), true)); + __ gs()->call( + Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86_64PointerSize, pDeliverException), true)); + // this call should never return + __ int3(); +#undef __ +} + +} // namespace x86_64 +} // namespace art diff --git a/compiler/utils/x86_64/jni_macro_assembler_x86_64.h b/compiler/utils/x86_64/jni_macro_assembler_x86_64.h new file mode 100644 index 0000000000..cc4e57c999 --- /dev/null +++ b/compiler/utils/x86_64/jni_macro_assembler_x86_64.h @@ -0,0 +1,190 @@ +/* + * Copyright (C) 2016 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef ART_COMPILER_UTILS_X86_64_JNI_MACRO_ASSEMBLER_X86_64_H_ +#define ART_COMPILER_UTILS_X86_64_JNI_MACRO_ASSEMBLER_X86_64_H_ + +#include <vector> + +#include "assembler_x86_64.h" +#include "base/arena_containers.h" +#include "base/enums.h" +#include "base/macros.h" +#include "offsets.h" +#include "utils/array_ref.h" +#include "utils/assembler.h" +#include "utils/jni_macro_assembler.h" + +namespace art { +namespace x86_64 { + +class X86_64JNIMacroAssembler FINAL : public JNIMacroAssemblerFwd<X86_64Assembler, + PointerSize::k64> { + public: + explicit X86_64JNIMacroAssembler(ArenaAllocator* arena) + : JNIMacroAssemblerFwd<X86_64Assembler, PointerSize::k64>(arena) {} + virtual ~X86_64JNIMacroAssembler() {} + + // + // Overridden common assembler high-level functionality + // + + // Emit code that will create an activation on the stack + void BuildFrame(size_t frame_size, + ManagedRegister method_reg, + ArrayRef<const ManagedRegister> callee_save_regs, + const ManagedRegisterEntrySpills& entry_spills) OVERRIDE; + + // Emit code that will remove an activation from the stack + void RemoveFrame(size_t frame_size, ArrayRef<const ManagedRegister> callee_save_regs) + OVERRIDE; + + void IncreaseFrameSize(size_t adjust) OVERRIDE; + void DecreaseFrameSize(size_t adjust) OVERRIDE; + + // Store routines + void Store(FrameOffset offs, ManagedRegister src, size_t size) OVERRIDE; + void StoreRef(FrameOffset dest, ManagedRegister src) OVERRIDE; + void StoreRawPtr(FrameOffset dest, ManagedRegister src) OVERRIDE; + + void StoreImmediateToFrame(FrameOffset dest, uint32_t imm, ManagedRegister scratch) OVERRIDE; + + void StoreStackOffsetToThread(ThreadOffset64 thr_offs, + FrameOffset fr_offs, + ManagedRegister scratch) OVERRIDE; + + void StoreStackPointerToThread(ThreadOffset64 thr_offs) OVERRIDE; + + void StoreSpanning(FrameOffset dest, + ManagedRegister src, + FrameOffset in_off, + ManagedRegister scratch) OVERRIDE; + + // Load routines + void Load(ManagedRegister dest, FrameOffset src, size_t size) OVERRIDE; + + void LoadFromThread(ManagedRegister dest, ThreadOffset64 src, size_t size) OVERRIDE; + + void LoadRef(ManagedRegister dest, FrameOffset src) OVERRIDE; + + void LoadRef(ManagedRegister dest, + ManagedRegister base, + MemberOffset offs, + bool unpoison_reference) OVERRIDE; + + void LoadRawPtr(ManagedRegister dest, ManagedRegister base, Offset offs) OVERRIDE; + + void LoadRawPtrFromThread(ManagedRegister dest, ThreadOffset64 offs) OVERRIDE; + + // Copying routines + void Move(ManagedRegister dest, ManagedRegister src, size_t size); + + void CopyRawPtrFromThread(FrameOffset fr_offs, + ThreadOffset64 thr_offs, + ManagedRegister scratch) OVERRIDE; + + void CopyRawPtrToThread(ThreadOffset64 thr_offs, FrameOffset fr_offs, ManagedRegister scratch) + OVERRIDE; + + void CopyRef(FrameOffset dest, FrameOffset src, ManagedRegister scratch) OVERRIDE; + + void Copy(FrameOffset dest, FrameOffset src, ManagedRegister scratch, size_t size) OVERRIDE; + + void Copy(FrameOffset dest, + ManagedRegister src_base, + Offset src_offset, + ManagedRegister scratch, + size_t size) OVERRIDE; + + void Copy(ManagedRegister dest_base, + Offset dest_offset, + FrameOffset src, + ManagedRegister scratch, + size_t size) OVERRIDE; + + void Copy(FrameOffset dest, + FrameOffset src_base, + Offset src_offset, + ManagedRegister scratch, + size_t size) OVERRIDE; + + void Copy(ManagedRegister dest, + Offset dest_offset, + ManagedRegister src, + Offset src_offset, + ManagedRegister scratch, + size_t size) OVERRIDE; + + void Copy(FrameOffset dest, + Offset dest_offset, + FrameOffset src, + Offset src_offset, + ManagedRegister scratch, + size_t size) OVERRIDE; + + void MemoryBarrier(ManagedRegister) OVERRIDE; + + // Sign extension + void SignExtend(ManagedRegister mreg, size_t size) OVERRIDE; + + // Zero extension + void ZeroExtend(ManagedRegister mreg, size_t size) OVERRIDE; + + // Exploit fast access in managed code to Thread::Current() + void GetCurrentThread(ManagedRegister tr) OVERRIDE; + void GetCurrentThread(FrameOffset dest_offset, ManagedRegister scratch) OVERRIDE; + + // Set up out_reg to hold a Object** into the handle scope, or to be null if the + // value is null and null_allowed. in_reg holds a possibly stale reference + // that can be used to avoid loading the handle scope entry to see if the value is + // null. + void CreateHandleScopeEntry(ManagedRegister out_reg, + FrameOffset handlescope_offset, + ManagedRegister in_reg, + bool null_allowed) OVERRIDE; + + // Set up out_off to hold a Object** into the handle scope, or to be null if the + // value is null and null_allowed. + void CreateHandleScopeEntry(FrameOffset out_off, + FrameOffset handlescope_offset, + ManagedRegister scratch, + bool null_allowed) OVERRIDE; + + // src holds a handle scope entry (Object**) load this into dst + virtual void LoadReferenceFromHandleScope(ManagedRegister dst, ManagedRegister src) OVERRIDE; + + // Heap::VerifyObject on src. In some cases (such as a reference to this) we + // know that src may not be null. + void VerifyObject(ManagedRegister src, bool could_be_null) OVERRIDE; + void VerifyObject(FrameOffset src, bool could_be_null) OVERRIDE; + + // Call to address held at [base+offset] + void Call(ManagedRegister base, Offset offset, ManagedRegister scratch) OVERRIDE; + void Call(FrameOffset base, Offset offset, ManagedRegister scratch) OVERRIDE; + void CallFromThread(ThreadOffset64 offset, ManagedRegister scratch) OVERRIDE; + + // Generate code to check if Thread::Current()->exception_ is non-null + // and branch to a ExceptionSlowPath if it is. + void ExceptionPoll(ManagedRegister scratch, size_t stack_adjust) OVERRIDE; + + private: + DISALLOW_COPY_AND_ASSIGN(X86_64JNIMacroAssembler); +}; + +} // namespace x86_64 +} // namespace art + +#endif // ART_COMPILER_UTILS_X86_64_JNI_MACRO_ASSEMBLER_X86_64_H_ |