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.h b/compiler/optimizing/code_generator.h
index 9364be3..b8540ba 100644
--- a/compiler/optimizing/code_generator.h
+++ b/compiler/optimizing/code_generator.h
@@ -350,6 +350,16 @@
// accessing the String's `value` field in String intrinsics.
static uint32_t GetArrayDataOffset(HArrayGet* array_get);
+ // Return the entry point offset for ReadBarrierMarkRegX, where X is `reg`.
+ template <size_t pointer_size>
+ static int32_t GetReadBarrierMarkEntryPointsOffset(size_t reg) {
+ DCHECK_LT(reg, 32u);
+ // The ReadBarrierMarkRegX entry points are ordered by increasing
+ // register number in Thread::tls_Ptr_.quick_entrypoints.
+ return QUICK_ENTRYPOINT_OFFSET(pointer_size, pReadBarrierMarkReg00).Int32Value()
+ + pointer_size * reg;
+ }
+
void EmitParallelMoves(Location from1,
Location to1,
Primitive::Type type1,
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 5316d59..e9bc3b3 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -412,8 +412,8 @@
// Slow path marking an object during a read barrier.
class ReadBarrierMarkSlowPathARM : public SlowPathCode {
public:
- ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location out, Location obj)
- : SlowPathCode(instruction), out_(out), obj_(obj) {
+ ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location obj)
+ : SlowPathCode(instruction), obj_(obj) {
DCHECK(kEmitCompilerReadBarrier);
}
@@ -421,9 +421,9 @@
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() ||
@@ -437,24 +437,44 @@
<< instruction_->DebugName();
__ Bind(GetEntryLabel());
+ // Save live registers before the runtime call, and in particular
+ // R0 (if it is live), as it is clobbered by functions
+ // art_quick_read_barrier_mark_regX.
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen);
- arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), obj_);
- arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark),
+ DCHECK_NE(reg, SP);
+ DCHECK_NE(reg, LR);
+ DCHECK_NE(reg, PC);
+ DCHECK(0 <= reg && reg < kNumberOfCoreRegisters) << reg;
+ // "Compact" slow path, saving two moves.
+ //
+ // Instead of using the standard runtime calling convention (input
+ // and output in R0):
+ //
+ // R0 <- obj
+ // R0 <- ReadBarrierMark(R0)
+ // obj <- R0
+ //
+ // 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<kArmWordSize>(reg);
+ // TODO: Do not emit a stack map for this runtime call.
+ arm_codegen->InvokeRuntime(entry_point_offset,
instruction_,
instruction_->GetDexPc(),
this);
- CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>();
- arm_codegen->Move32(out_, Location::RegisterLocation(R0));
RestoreLiveRegisters(codegen, locations);
__ b(GetExitLabel());
}
private:
- const Location out_;
const Location obj_;
DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM);
@@ -6175,7 +6195,7 @@
// Slow path used to mark the GC root `root`.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root, root);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root);
codegen_->AddSlowPath(slow_path);
// IP = Thread::Current()->GetIsGcMarking()
@@ -6324,7 +6344,7 @@
// Slow path used to mark the object `ref` when it is gray.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref, ref);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref);
AddSlowPath(slow_path);
// if (rb_state == ReadBarrier::gray_ptr_)
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index fc2c2c3..e480f12 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -580,8 +580,8 @@
// Slow path marking an object during a read barrier.
class ReadBarrierMarkSlowPathARM64 : public SlowPathCodeARM64 {
public:
- ReadBarrierMarkSlowPathARM64(HInstruction* instruction, Location out, Location obj)
- : SlowPathCodeARM64(instruction), out_(out), obj_(obj) {
+ ReadBarrierMarkSlowPathARM64(HInstruction* instruction, Location obj)
+ : SlowPathCodeARM64(instruction), obj_(obj) {
DCHECK(kEmitCompilerReadBarrier);
}
@@ -589,9 +589,8 @@
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
- Primitive::Type type = Primitive::kPrimNot;
DCHECK(locations->CanCall());
- DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg()));
+ DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(obj_.reg()));
DCHECK(instruction_->IsInstanceFieldGet() ||
instruction_->IsStaticFieldGet() ||
instruction_->IsArrayGet() ||
@@ -605,24 +604,44 @@
<< instruction_->DebugName();
__ Bind(GetEntryLabel());
+ // Save live registers before the runtime call, and in particular
+ // W0 (if it is live), as it is clobbered by functions
+ // art_quick_read_barrier_mark_regX.
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
- arm64_codegen->MoveLocation(LocationFrom(calling_convention.GetRegisterAt(0)), obj_, type);
- arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark),
+ DCHECK_NE(obj_.reg(), LR);
+ DCHECK_NE(obj_.reg(), WSP);
+ DCHECK_NE(obj_.reg(), WZR);
+ DCHECK(0 <= obj_.reg() && obj_.reg() < kNumberOfWRegisters) << obj_.reg();
+ // "Compact" slow path, saving two moves.
+ //
+ // Instead of using the standard runtime calling convention (input
+ // and output in W0):
+ //
+ // W0 <- obj
+ // W0 <- ReadBarrierMark(W0)
+ // obj <- W0
+ //
+ // 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<kArm64WordSize>(obj_.reg());
+ // TODO: Do not emit a stack map for this runtime call.
+ arm64_codegen->InvokeRuntime(entry_point_offset,
instruction_,
instruction_->GetDexPc(),
this);
- CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>();
- arm64_codegen->MoveLocation(out_, calling_convention.GetReturnLocation(type), type);
RestoreLiveRegisters(codegen, locations);
__ B(GetExitLabel());
}
private:
- const Location out_;
const Location obj_;
DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM64);
@@ -5059,7 +5078,7 @@
// Slow path used to mark the GC root `root`.
SlowPathCodeARM64* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, root, root);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, root);
codegen_->AddSlowPath(slow_path);
MacroAssembler* masm = GetVIXLAssembler();
@@ -5267,7 +5286,7 @@
// Slow path used to mark the object `ref` when it is gray.
SlowPathCodeARM64* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, ref, ref);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM64(instruction, ref);
AddSlowPath(slow_path);
// if (rb_state == ReadBarrier::gray_ptr_)
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_)
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 5e30203..a524057 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -451,8 +451,8 @@
// Slow path marking an object during a read barrier.
class ReadBarrierMarkSlowPathX86_64 : public SlowPathCode {
public:
- ReadBarrierMarkSlowPathX86_64(HInstruction* instruction, Location out, Location obj)
- : SlowPathCode(instruction), out_(out), obj_(obj) {
+ ReadBarrierMarkSlowPathX86_64(HInstruction* instruction, Location obj)
+ : SlowPathCode(instruction), obj_(obj) {
DCHECK(kEmitCompilerReadBarrier);
}
@@ -460,9 +460,9 @@
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() ||
@@ -476,24 +476,42 @@
<< instruction_->DebugName();
__ Bind(GetEntryLabel());
+ // Save live registers before the runtime call, and in particular
+ // RDI and/or RAX (if they are live), as they are clobbered by
+ // functions art_quick_read_barrier_mark_regX.
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
- x86_64_codegen->Move(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), obj_);
- x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark),
- instruction_,
- instruction_->GetDexPc(),
- this);
- CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>();
- x86_64_codegen->Move(out_, Location::RegisterLocation(RAX));
+ DCHECK_NE(reg, RSP);
+ DCHECK(0 <= reg && reg < kNumberOfCpuRegisters) << reg;
+ // "Compact" slow path, saving two moves.
+ //
+ // Instead of using the standard runtime calling convention (input
+ // and output in R0):
+ //
+ // RDI <- obj
+ // RAX <- ReadBarrierMark(RDI)
+ // obj <- RAX
+ //
+ // 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<kX86_64WordSize>(reg);
+ // TODO: Do not emit a stack map for this runtime call.
+ x86_64_codegen->InvokeRuntime(entry_point_offset,
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
private:
- const Location out_;
const Location obj_;
DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathX86_64);
@@ -6378,7 +6396,7 @@
// Slow path used to mark the GC root `root`.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86_64(instruction, root, root);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86_64(instruction, root);
codegen_->AddSlowPath(slow_path);
__ gs()->cmpl(Address::Absolute(Thread::IsGcMarkingOffset<kX86_64WordSize>().Int32Value(),
@@ -6509,7 +6527,7 @@
// Slow path used to mark the object `ref` when it is gray.
SlowPathCode* slow_path =
- new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86_64(instruction, ref, ref);
+ new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathX86_64(instruction, ref);
AddSlowPath(slow_path);
// if (rb_state == ReadBarrier::gray_ptr_)