Introduce more compact ReadBarrierMark slow-paths.
Replace entry point ReadBarrierMark with 32
ReadBarrierMarkRegX entry points, using register
number X as input and output (instead of the standard
runtime calling convention) to save two moves in Baker's
read barrier mark slow-path code.
Test: ART host and target (ARM, ARM64) tests.
Bug: 29506760
Bug: 12687968
Change-Id: I73cfb82831cf040b8b018e984163c865cc44ed87
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 1261619..b33cabb 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -430,8 +430,8 @@
// Slow path marking an object during a read barrier.
class ReadBarrierMarkSlowPathX86 : public SlowPathCode {
public:
- ReadBarrierMarkSlowPathX86(HInstruction* instruction, Location out, Location obj)
- : SlowPathCode(instruction), out_(out), obj_(obj) {
+ ReadBarrierMarkSlowPathX86(HInstruction* instruction, Location obj)
+ : SlowPathCode(instruction), obj_(obj) {
DCHECK(kEmitCompilerReadBarrier);
}
@@ -439,9 +439,9 @@
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
- Register reg_out = out_.AsRegister<Register>();
+ Register reg = obj_.AsRegister<Register>();
DCHECK(locations->CanCall());
- DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out));
+ DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg));
DCHECK(instruction_->IsInstanceFieldGet() ||
instruction_->IsStaticFieldGet() ||
instruction_->IsArrayGet() ||
@@ -455,24 +455,42 @@
<< instruction_->DebugName();
__ Bind(GetEntryLabel());
+ // Save live registers before the runtime call, and in particular
+ // EAX (if it is live), as it is clobbered by functions
+ // art_quick_read_barrier_mark_regX.
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
- x86_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), obj_);
- x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark),
+ DCHECK_NE(reg, ESP);
+ DCHECK(0 <= reg && reg < kNumberOfCpuRegisters) << reg;
+ // "Compact" slow path, saving two moves.
+ //
+ // Instead of using the standard runtime calling convention (input
+ // and output in EAX):
+ //
+ // EAX <- obj
+ // EAX <- ReadBarrierMark(EAX)
+ // obj <- EAX
+ //
+ // we just use rX (the register holding `obj`) as input and output
+ // of a dedicated entrypoint:
+ //
+ // rX <- ReadBarrierMarkRegX(rX)
+ //
+ int32_t entry_point_offset =
+ CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kX86WordSize>(reg);
+ // TODO: Do not emit a stack map for this runtime call.
+ x86_codegen->InvokeRuntime(entry_point_offset,
instruction_,
instruction_->GetDexPc(),
this);
- CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>();
- x86_codegen->Move32(out_, Location::RegisterLocation(EAX));
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
private:
- const Location out_;
const Location obj_;
DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathX86);
@@ -6926,7 +6944,7 @@
// Slow path used to mark the GC root `root`.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86(instruction, root, root);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86(instruction, root);
codegen_->AddSlowPath(slow_path);
__ fs()->cmpl(Address::Absolute(Thread::IsGcMarkingOffset<kX86WordSize>().Int32Value()),
@@ -7056,7 +7074,7 @@
// Slow path used to mark the object `ref` when it is gray.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86(instruction, ref, ref);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86(instruction, ref);
AddSlowPath(slow_path);
// if (rb_state == ReadBarrier::gray_ptr_)