• UID623
  • Fans2
  • Follows1
  • Posts72

An Adventure in the ART World: How to Optimize the Compilation Procedure of the Compiler

More Posted time:Oct 17, 2016 9:37 AM
Let’s go back to see how to optimize the compilation procedure of the compiler. First, we’ll check a chart to see what new objects we will encounter.


Let’s first look at the Compile function, the entry point to optimizing the compiler.
CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item,
                                            uint32_t access_flags,
                                            InvokeType invoke_type,
                                            uint16_t class_def_idx,
                                            uint32_t method_idx,
                                            jobject jclass_loader,
                                            const DexFile& dex_file) const {
  CompilerDriver* compiler_driver = GetCompilerDriver();
  CompiledMethod* method = nullptr;
  if (compiler_driver->IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file) &&
      !compiler_driver->GetVerifiedMethod(&dex_file, method_idx)->HasRuntimeThrow()) {

If the check succeeds, we invoke the TryCompile method to complete the compilation.
method = TryCompile(code_item, access_flags, invoke_type, class_def_idx,
                         method_idx, jclass_loader, dex_file);
  } else {
    if (compiler_driver->GetCompilerOptions().VerifyAtRuntime()) {
    } else {

If the compilation succeeds, it will return directly. If it fails, we invoke other methods to compile it again.
if (method != nullptr) {
    return method;
  method = delegate_->Compile(code_item, access_flags, invoke_type, class_def_idx, method_idx,
                              jclass_loader, dex_file);

  if (method != nullptr) {
  return method;

CompiledMethod* OptimizingCompiler::TryCompile(const DexFile::CodeItem* code_item,
                                               uint32_t access_flags,
                                               InvokeType invoke_type,
                                               uint16_t class_def_idx,
                                               uint32_t method_idx,
                                               jobject class_loader,
                                               const DexFile& dex_file) const {
  std::string method_name = PrettyMethod(method_idx, dex_file);
  CompilerDriver* compiler_driver = GetCompilerDriver();

For ARM32 architecture, we use Thumb2 instruction sets by default.
As for ARM instruction sets, we introduced it previously in An Adventure in the ART World (Book 3): Quick Tutorial on the Architecture of ARM64 CPU.
InstructionSet instruction_set = compiler_driver->GetInstructionSet();
  // Always use the thumb2 assembler: some runtime functionality (like implicit stack
  // overflow checks) assume thumb2.
  if (instruction_set == kArm) {
    instruction_set = kThumb2;

If it cannot be optimized, or it does not need to be, it will return directly.
// `run_optimizations_` is set explicitly (either through a compiler filter
  // or the debuggable flag). If it is set, we can run baseline. Otherwise, we
  // fall back to Quick.
  bool should_use_baseline = !run_optimizations_;
  bool can_optimize = CanOptimize(*code_item);
  if (!can_optimize && !should_use_baseline) {
    // We know we will not compile this method. Bail out before doing any work.
    return nullptr;

If it does not support the instruction sets, we do not need to compile it.
// Do not attempt to compile on architectures we do not support.
  if (!IsInstructionSetSupported(instruction_set)) {
    return nullptr;

If it is not worth the compilation, then we do not do it.
How do we determine whether it is worth it or not? If there are too many instructions, or if the virtual registers are over-utilized, we don’t bother with it. Let’s look at the code of IsPathologicalCase.
bool Compiler::IsPathologicalCase(const DexFile::CodeItem& code_item,
                                  uint32_t method_idx,
                                  const DexFile& dex_file) {
   * Skip compilation for pathologically large methods - either by instruction count or num vregs.
   * Dalvik uses 16-bit uints for instruction and register counts.  We'll limit to a quarter
   * of that, which also guarantees we cannot overflow our 16-bit internal Quick SSA name space.
  if (code_item.insns_size_in_code_units_ >= UINT16_MAX / 4) {
    LOG(INFO) << "Method exceeds compiler instruction limit: "
              << code_item.insns_size_in_code_units_
              << " in " << PrettyMethod(method_idx, dex_file);
    return true;
  if (code_item.registers_size_ >= UINT16_MAX / 4) {
    LOG(INFO) << "Method exceeds compiler virtual register limit: "
              << code_item.registers_size_ << " in " << PrettyMethod(method_idx, dex_file);
    return true;
  return false;

UINT16_MAX is 65535, the largest integer of two bytes. Divided by 4, then we got 16384 or so.
if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) {
    return nullptr;

Then we process space options. In order to save space, code items greater than 128 units will not be compiled.
// Implementation of the space filter: do not compile a code item whose size in
  // code units is bigger than 128.
  static constexpr size_t kSpaceFilterOptimizingThreshold = 128;
  const CompilerOptions& compiler_options = compiler_driver->GetCompilerOptions();
  if ((compiler_options.GetCompilerFilter() == CompilerOptions::kSpace)
      && (code_item->insns_size_in_code_units_ > kSpaceFilterOptimizingThreshold)) {
    return nullptr;

Here comes our old friend, DexComplicationUnit.

DexCompilationUnit dex_compilation_unit(
    nullptr, class_loader, art::Runtime::Current()->GetClassLinker(), dex_file, code_item,
    class_def_idx, method_idx, access_flags,
    compiler_driver->GetVerifiedMethod(&dex_file, method_idx));

Next, we construct an HGraph object, which works like the MIRGraph in a quick Compiler.
ArenaAllocator arena(Runtime::Current()->GetArenaPool());
  HGraph* graph = new (&arena) HGraph(
      &arena, dex_file, method_idx, compiler_driver->GetInstructionSet(),

  // For testing purposes, we put a special marker on method names that should be compiled
  // with this compiler. This makes sure we're not regressing.
  bool shouldCompile = method_name.find("$opt$") != std::string::npos;
  bool shouldOptimize = method_name.find("$opt$reg$") != std::string::npos && run_optimizations_;

  std::unique_ptr<CodeGenerator> codegen(
  if (codegen.get() == nullptr) {
    CHECK(!shouldCompile) << "Could not find code generator for optimizing compiler";
    return nullptr;

  PassInfoPrinter pass_info_printer(graph,

Then we construct an HGraphBuilder object and after that we invoke the BuildGraph method in the HGraphBuilder to construct this graph.
HGraphBuilder builder(graph,

  VLOG(compiler) << "Building " << method_name;

    PassInfo pass_info(HGraphBuilder::kBuilderPassName, &pass_info_printer);
    if (!builder.BuildGraph(*code_item)) {
      DCHECK(!(IsCompilingWithCoreImage() && shouldCompile))
          << "Could not build graph in optimizing compiler";
      return nullptr;

Next, we try to construct an SSA to the Hgraph.
bool can_allocate_registers = RegisterAllocator::CanAllocateRegistersFor(*graph, instruction_set);

  if (run_optimizations_ && can_optimize && can_allocate_registers) {
    VLOG(compiler) << "Optimizing " << method_name;

      PassInfo pass_info(SsaBuilder::kSsaBuilderPassName, &pass_info_printer);
      if (!graph->TryBuildingSsa()) {
        // We could not transform the graph to SSA, bailout.
        LOG(INFO) << "Skipping compilation of " << method_name << ": it contains a non natural loop";
        return nullptr;

    return CompileOptimized(graph,
  } else if (shouldOptimize && can_allocate_registers) {
    LOG(FATAL) << "Could not allocate registers in optimizing compiler";
  } else if (should_use_baseline) {

Here we can see the situation without optimization.
VLOG(compiler) << "Compile baseline " << method_name;

    if (!run_optimizations_) {
    } else if (!can_optimize) {
    } else if (!can_allocate_registers) {

    return CompileBaseline(codegen.get(), compiler_driver, dex_compilation_unit);
  } else {
    return nullptr;

Then let’s check the section of optimized compilation.
CompiledMethod* OptimizingCompiler::CompileOptimized(HGraph* graph,
                                                     CodeGenerator* codegen,
                                                     CompilerDriver* compiler_driver,
                                                     const DexFile& dex_file,
                                                     const DexCompilationUnit& dex_compilation_unit,
                                                     PassInfoPrinter* pass_info_printer) const {
  StackHandleScopeCollection handles(Thread::Current());

Next, we begin to invoke RunOptimizations to optimize it, for a total of 15 rounds of optimization.
And AllocateRegisters is the 16th, the liveness optimization.
RunOptimizations(graph, compiler_driver, compilation_stats_.get(),
                   dex_file, dex_compilation_unit, pass_info_printer, &handles);

  AllocateRegisters(graph, codegen, pass_info_printer);

Here is the generation of target code.
CodeVectorAllocator allocator;

  DefaultSrcMap src_mapping_table;
  if (compiler_driver->GetCompilerOptions().GetGenerateDebugInfo()) {

  std::vector<uint8_t> stack_map;


At the end of the process, we allocate space and save the compilation method.
return CompiledMethod::SwapAllocCompiledMethod(
      ArrayRef<const uint8_t>(allocator.GetMemory()),
      // Follow Quick's behavior and set the frame size to zero if it is
      // considered "empty" (see the definition of
      // art::CodeGenerator::HasEmptyFrame).
      codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
      ArrayRef<const uint8_t>(),  // mapping_table.
      ArrayRef<const uint8_t>(stack_map),
      ArrayRef<const uint8_t>(),  // native_gc_map.
      ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()),
      ArrayRef<const LinkerPatch>());

Here comes the optimization process.
1. IntrinsicsRecognizer
2. Here is the first HconstantFolding.
3. And the second InstructionSimplifier.
4. HdeadCodeElimination
5. Hinliner
6. HbooleanSimplifier
7. And the second HconstantFolding.
8. SideEffectsAnalysis
9. GVNOptimization
10. LICM
11. BoundsCheckElimination
12. ReferenceTypePropagation
13. And the second InstructionSimplifier.
14. And the second HdeadCodeElimination.
15. Here the third InstructionSimplifier begins.
static void RunOptimizations(HGraph* graph,
                             CompilerDriver* driver,
                             OptimizingCompilerStats* stats,
                             const DexFile& dex_file,
                             const DexCompilationUnit& dex_compilation_unit,
                             PassInfoPrinter* pass_info_printer,
                             StackHandleScopeCollection* handles) {
  HDeadCodeElimination dce1(graph, stats,
  HDeadCodeElimination dce2(graph, stats,
  HConstantFolding fold1(graph);
  InstructionSimplifier simplify1(graph, stats);
  HBooleanSimplifier boolean_simplify(graph);

  HInliner inliner(graph, dex_compilation_unit, dex_compilation_unit, driver, stats);

  HConstantFolding fold2(graph, "constant_folding_after_inlining");
  SideEffectsAnalysis side_effects(graph);
  GVNOptimization gvn(graph, side_effects);
  LICM licm(graph, side_effects);
  BoundsCheckElimination bce(graph);
  ReferenceTypePropagation type_propagation(graph, dex_file, dex_compilation_unit, handles);
  InstructionSimplifier simplify2(graph, stats, "instruction_simplifier_after_types");
  InstructionSimplifier simplify3(graph, stats, "instruction_simplifier_before_codegen");

  IntrinsicsRecognizer intrinsics(graph, dex_compilation_unit.GetDexFile(), driver);

  HOptimization* optimizations[] = {
    // BooleanSimplifier depends on the InstructionSimplifier removing redundant
    // suspend checks to recognize empty blocks.
    // The codegen has a few assumptions that only the instruction simplifier can
    // satisfy. For example, the code generator does not expect to see a
    // HTypeConversion from a type to the same type.

  RunOptimizations(optimizations, arraysize(optimizations), pass_info_printer);

Let’s begin the liveness optimization process.
static void AllocateRegisters(HGraph* graph,
                              CodeGenerator* codegen,
                              PassInfoPrinter* pass_info_printer) {
  SsaLivenessAnalysis liveness(graph, codegen);
    PassInfo pass_info(SsaLivenessAnalysis::kLivenessPassName, pass_info_printer);
    PassInfo pass_info(RegisterAllocator::kRegisterAllocatorPassName, pass_info_printer);
    RegisterAllocator(graph->GetArena(), codegen, liveness).AllocateRegisters();

We split the main logic into two parts: Initialize() and CompileInternal().
void CodeGenerator::CompileOptimized(CodeAllocator* allocator) {
  // The register allocator already called `InitializeCodeGeneration`,
  // where the frame size has been computed.
  DCHECK(block_order_ != nullptr);
  CompileInternal(allocator, /* is_baseline */ false);


void CodeGenerator::CompileInternal(CodeAllocator* allocator, bool is_baseline) {
  is_baseline_ = is_baseline;
  HGraphVisitor* instruction_visitor = GetInstructionVisitor();
  DCHECK_EQ(current_block_index_, 0u);
  DCHECK_EQ(GetAssembler()->cfi().GetCurrentCFAOffset(), static_cast<int>(frame_size_));
  for (size_t e = block_order_->Size(); current_block_index_ < e; ++current_block_index_) {
    HBasicBlock* block = block_order_->Get(current_block_index_);
    // Don't generate code for an empty block. Its predecessors will branch to its successor
    // directly. Also, the label of that block will not be emitted, so this helps catch
    // errors where we reference that label.
    if (block->IsSingleGoto()) continue;

When we get the Bind, we have already begun the code generation section which is related to the specific architecture. For example, we have the Bind of CodeGeneratorARM64, which is for the arm64 architecture.
void CodeGeneratorARM64::Bind(HBasicBlock* block) {
  __ Bind(GetLabelOf(block));

After the bind, we traverse the instructions in the block.
    for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
      HInstruction* current = it.Current();
      if (is_baseline) {

Next, we’ll handle the operations under the slow path.
// Generate the slow paths.
  for (size_t i = 0, e = slow_paths_.Size(); i < e; ++i) {

Once that’s complete, we call the assembler to generate the instructions
// Finalize instructions in assember;


void CodeGenerator::Finalize(CodeAllocator* allocator) {
  size_t code_size = GetAssembler()->CodeSize();
  uint8_t* buffer = allocator->Allocate(code_size);

  MemoryRegion code(buffer, code_size);

Here we have the FinalizeInstructions for Arm64Assembler. The other chips use the generic version.
void Arm64Assembler::FinalizeInstructions(const MemoryRegion& region) {
  // Copy the instructions from the buffer.
  MemoryRegion from(vixl_masm_->GetStartAddress<void*>(), CodeSize());
  region.CopyFrom(0, from);

CompiledMethod::SwapAllocCompiledMethod Finally, let’s see SwapAllocCompiledMethod. The rest of our work is basically the distribution of space.

CompiledMethod* CompiledMethod::SwapAllocCompiledMethod(
    CompilerDriver* driver,
    InstructionSet instruction_set,
    const ArrayRef<const uint8_t>& quick_code,
    const size_t frame_size_in_bytes,
    const uint32_t core_spill_mask,
    const uint32_t fp_spill_mask,
    DefaultSrcMap* src_mapping_table,
    const ArrayRef<const uint8_t>& mapping_table,
    const ArrayRef<const uint8_t>& vmap_table,
    const ArrayRef<const uint8_t>& native_gc_map,
    const ArrayRef<const uint8_t>& cfi_info,
    const ArrayRef<const LinkerPatch>& patches) {
  SwapAllocator<CompiledMethod> alloc(driver->GetSwapSpaceAllocator());
  CompiledMethod* ret = alloc.allocate(1);
  alloc.construct(ret, driver, instruction_set, quick_code, frame_size_in_bytes, core_spill_mask,
                  fp_spill_mask, src_mapping_table, mapping_table, vmap_table, native_gc_map,
                  cfi_info, patches);
  return ret;

Latest likes: