x86/x86-64 read barrier support for concurrent GC in Optimizing.
This first implementation uses slow paths to instrument heap
reference loads and GC root loads for the concurrent copying
collector, respectively calling the artReadBarrierSlow and
artReadBarrierForRootSlow (new) runtime entry points.
Notes:
- This implementation does not instrument HInvokeVirtual
nor HInvokeInterface instructions (for class reference
loads), as the corresponding read barriers are not stricly
required with the current concurrent copying collector.
- Intrinsics which may eventually call (on slow path) are
disabled when read barriers are enabled, as the current
slow path infrastructure does not support this case.
- When read barriers are enabled, the code generated for a
HArraySet instruction always go into the array set slow
path for object arrays (delegating the operation to the
runtime), as we are lacking a mechanism to keep a
temporary register live accross a runtime call (needed for
the instrumentation of type checking code, which requires
two successive read barriers).
Bug: 12687968
Change-Id: I14cd6107233c326389120336f93955b28ffbb329
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index e2ad667..d55c084 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -34,6 +34,9 @@
namespace art {
+template<class MirrorType>
+class GcRoot;
+
namespace x86_64 {
static constexpr int kCurrentMethodStackOffset = 0;
@@ -52,16 +55,16 @@
explicit NullCheckSlowPathX86_64(HNullCheck* instruction) : instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
if (instruction_->CanThrowIntoCatchBlock()) {
// Live registers will be restored in the catch block if caught.
SaveLiveRegisters(codegen, instruction_->GetLocations());
}
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowNullPointer),
- instruction_,
- instruction_->GetDexPc(),
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowNullPointer),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
}
bool IsFatal() const OVERRIDE { return true; }
@@ -78,16 +81,16 @@
explicit DivZeroCheckSlowPathX86_64(HDivZeroCheck* instruction) : instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
if (instruction_->CanThrowIntoCatchBlock()) {
// Live registers will be restored in the catch block if caught.
SaveLiveRegisters(codegen, instruction_->GetLocations());
}
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowDivZero),
- instruction_,
- instruction_->GetDexPc(),
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowDivZero),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
}
bool IsFatal() const OVERRIDE { return true; }
@@ -139,18 +142,18 @@
: instruction_(instruction), successor_(successor) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, instruction_->GetLocations());
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pTestSuspend),
- instruction_,
- instruction_->GetDexPc(),
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pTestSuspend),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
RestoreLiveRegisters(codegen, instruction_->GetLocations());
if (successor_ == nullptr) {
__ jmp(GetReturnLabel());
} else {
- __ jmp(x64_codegen->GetLabelOf(successor_));
+ __ jmp(x86_64_codegen->GetLabelOf(successor_));
}
}
@@ -180,7 +183,7 @@
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
if (instruction_->CanThrowIntoCatchBlock()) {
// Live registers will be restored in the catch block if caught.
@@ -196,8 +199,10 @@
locations->InAt(1),
Location::RegisterLocation(calling_convention.GetRegisterAt(1)),
Primitive::kPrimInt);
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowArrayBounds),
- instruction_, instruction_->GetDexPc(), this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowArrayBounds),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
}
bool IsFatal() const OVERRIDE { return true; }
@@ -222,22 +227,25 @@
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = at_->GetLocations();
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
__ movl(CpuRegister(calling_convention.GetRegisterAt(0)), Immediate(cls_->GetTypeIndex()));
- x64_codegen->InvokeRuntime(do_clinit_ ? QUICK_ENTRY_POINT(pInitializeStaticStorage)
- : QUICK_ENTRY_POINT(pInitializeType),
- at_, dex_pc_, this);
+ x86_64_codegen->InvokeRuntime(do_clinit_ ?
+ QUICK_ENTRY_POINT(pInitializeStaticStorage) :
+ QUICK_ENTRY_POINT(pInitializeType),
+ at_,
+ dex_pc_,
+ this);
Location out = locations->Out();
// Move the class to the desired location.
if (out.IsValid()) {
DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
- x64_codegen->Move(out, Location::RegisterLocation(RAX));
+ x86_64_codegen->Move(out, Location::RegisterLocation(RAX));
}
RestoreLiveRegisters(codegen, locations);
@@ -271,18 +279,18 @@
LocationSummary* locations = instruction_->GetLocations();
DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
__ movl(CpuRegister(calling_convention.GetRegisterAt(0)),
Immediate(instruction_->GetStringIndex()));
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pResolveString),
- instruction_,
- instruction_->GetDexPc(),
- this);
- x64_codegen->Move(locations->Out(), Location::RegisterLocation(RAX));
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pResolveString),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
+ x86_64_codegen->Move(locations->Out(), Location::RegisterLocation(RAX));
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
@@ -308,18 +316,9 @@
DCHECK(instruction_->IsCheckCast()
|| !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
- if (instruction_->IsCheckCast()) {
- // The codegen for the instruction overwrites `temp`, so put it back in place.
- CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
- CpuRegister temp = locations->GetTemp(0).AsRegister<CpuRegister>();
- uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- __ movl(temp, Address(obj, class_offset));
- __ MaybeUnpoisonHeapReference(temp);
- }
-
if (!is_fatal_) {
SaveLiveRegisters(codegen, locations);
}
@@ -336,21 +335,24 @@
Primitive::kPrimNot);
if (instruction_->IsInstanceOf()) {
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial),
- instruction_,
- dex_pc,
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial),
+ instruction_,
+ dex_pc,
+ this);
+ CheckEntrypointTypes<
+ kQuickInstanceofNonTrivial, uint32_t, const mirror::Class*, const mirror::Class*>();
} else {
DCHECK(instruction_->IsCheckCast());
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast),
- instruction_,
- dex_pc,
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast),
+ instruction_,
+ dex_pc,
+ this);
+ CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>();
}
if (!is_fatal_) {
if (instruction_->IsInstanceOf()) {
- x64_codegen->Move(locations->Out(), Location::RegisterLocation(RAX));
+ x86_64_codegen->Move(locations->Out(), Location::RegisterLocation(RAX));
}
RestoreLiveRegisters(codegen, locations);
@@ -375,15 +377,15 @@
: instruction_(instruction) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, instruction_->GetLocations());
DCHECK(instruction_->IsDeoptimize());
HDeoptimize* deoptimize = instruction_->AsDeoptimize();
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize),
- deoptimize,
- deoptimize->GetDexPc(),
- this);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize),
+ deoptimize,
+ deoptimize->GetDexPc(),
+ this);
}
const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathX86_64"; }
@@ -421,11 +423,11 @@
nullptr);
codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
- CodeGeneratorX86_64* x64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
- x64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject),
- instruction_,
- instruction_->GetDexPc(),
- this);
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
RestoreLiveRegisters(codegen, locations);
__ jmp(GetExitLabel());
}
@@ -438,6 +440,219 @@
DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathX86_64);
};
+// Slow path generating a read barrier for a heap reference.
+class ReadBarrierForHeapReferenceSlowPathX86_64 : public SlowPathCode {
+ public:
+ ReadBarrierForHeapReferenceSlowPathX86_64(HInstruction* instruction,
+ Location out,
+ Location ref,
+ Location obj,
+ uint32_t offset,
+ Location index)
+ : instruction_(instruction),
+ out_(out),
+ ref_(ref),
+ obj_(obj),
+ offset_(offset),
+ index_(index) {
+ DCHECK(kEmitCompilerReadBarrier);
+ // If `obj` is equal to `out` or `ref`, it means the initial
+ // object has been overwritten by (or after) the heap object
+ // reference load to be instrumented, e.g.:
+ //
+ // __ movl(out, Address(out, offset));
+ // codegen_->GenerateReadBarrier(instruction, out_loc, out_loc, out_loc, offset);
+ //
+ // In that case, we have lost the information about the original
+ // object, and the emitted read barrier cannot work properly.
+ DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out;
+ DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref;
+}
+
+ void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+ CodeGeneratorX86_64* x86_64_codegen = down_cast<CodeGeneratorX86_64*>(codegen);
+ LocationSummary* locations = instruction_->GetLocations();
+ CpuRegister reg_out = out_.AsRegister<CpuRegister>();
+ DCHECK(locations->CanCall());
+ DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out.AsRegister())) << out_;
+ DCHECK(!instruction_->IsInvoke() ||
+ (instruction_->IsInvokeStaticOrDirect() &&
+ instruction_->GetLocations()->Intrinsified()));
+
+ __ Bind(GetEntryLabel());
+ SaveLiveRegisters(codegen, locations);
+
+ // We may have to change the index's value, but as `index_` is a
+ // constant member (like other "inputs" of this slow path),
+ // introduce a copy of it, `index`.
+ Location index = index_;
+ if (index_.IsValid()) {
+ // Handle `index_` for HArrayGet and intrinsic UnsafeGetObject.
+ if (instruction_->IsArrayGet()) {
+ // Compute real offset and store it in index_.
+ Register index_reg = index_.AsRegister<CpuRegister>().AsRegister();
+ DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg));
+ if (codegen->IsCoreCalleeSaveRegister(index_reg)) {
+ // We are about to change the value of `index_reg` (see the
+ // calls to art::x86_64::X86_64Assembler::shll and
+ // art::x86_64::X86_64Assembler::AddImmediate below), but it
+ // has not been saved by the previous call to
+ // art::SlowPathCode::SaveLiveRegisters, as it is a
+ // callee-save register --
+ // art::SlowPathCode::SaveLiveRegisters does not consider
+ // callee-save registers, as it has been designed with the
+ // assumption that callee-save registers are supposed to be
+ // handled by the called function. So, as a callee-save
+ // register, `index_reg` _would_ eventually be saved onto
+ // the stack, but it would be too late: we would have
+ // changed its value earlier. Therefore, we manually save
+ // it here into another freely available register,
+ // `free_reg`, chosen of course among the caller-save
+ // registers (as a callee-save `free_reg` register would
+ // exhibit the same problem).
+ //
+ // Note we could have requested a temporary register from
+ // the register allocator instead; but we prefer not to, as
+ // this is a slow path, and we know we can find a
+ // caller-save register that is available.
+ Register free_reg = FindAvailableCallerSaveRegister(codegen).AsRegister();
+ __ movl(CpuRegister(free_reg), CpuRegister(index_reg));
+ index_reg = free_reg;
+ index = Location::RegisterLocation(index_reg);
+ } else {
+ // The initial register stored in `index_` has already been
+ // saved in the call to art::SlowPathCode::SaveLiveRegisters
+ // (as it is not a callee-save register), so we can freely
+ // use it.
+ }
+ // Shifting the index value contained in `index_reg` by the
+ // scale factor (2) cannot overflow in practice, as the
+ // runtime is unable to allocate object arrays with a size
+ // larger than 2^26 - 1 (that is, 2^28 - 4 bytes).
+ __ shll(CpuRegister(index_reg), Immediate(TIMES_4));
+ static_assert(
+ sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
+ "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
+ __ AddImmediate(CpuRegister(index_reg), Immediate(offset_));
+ } else {
+ DCHECK(instruction_->IsInvoke());
+ DCHECK(instruction_->GetLocations()->Intrinsified());
+ DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) ||
+ (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile))
+ << instruction_->AsInvoke()->GetIntrinsic();
+ DCHECK_EQ(offset_, 0U);
+ DCHECK(index_.IsRegister());
+ }
+ }
+
+ // We're moving two or three locations to locations that could
+ // overlap, so we need a parallel move resolver.
+ InvokeRuntimeCallingConvention calling_convention;
+ HParallelMove parallel_move(codegen->GetGraph()->GetArena());
+ parallel_move.AddMove(ref_,
+ Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
+ Primitive::kPrimNot,
+ nullptr);
+ parallel_move.AddMove(obj_,
+ Location::RegisterLocation(calling_convention.GetRegisterAt(1)),
+ Primitive::kPrimNot,
+ nullptr);
+ if (index.IsValid()) {
+ parallel_move.AddMove(index,
+ Location::RegisterLocation(calling_convention.GetRegisterAt(2)),
+ Primitive::kPrimInt,
+ nullptr);
+ codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
+ } else {
+ codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
+ __ movl(CpuRegister(calling_convention.GetRegisterAt(2)), Immediate(offset_));
+ }
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
+ CheckEntrypointTypes<
+ kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>();
+ x86_64_codegen->Move(out_, Location::RegisterLocation(RAX));
+
+ RestoreLiveRegisters(codegen, locations);
+ __ jmp(GetExitLabel());
+ }
+
+ const char* GetDescription() const OVERRIDE {
+ return "ReadBarrierForHeapReferenceSlowPathX86_64";
+ }
+
+ private:
+ CpuRegister FindAvailableCallerSaveRegister(CodeGenerator* codegen) {
+ size_t ref = static_cast<int>(ref_.AsRegister<CpuRegister>().AsRegister());
+ size_t obj = static_cast<int>(obj_.AsRegister<CpuRegister>().AsRegister());
+ for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) {
+ if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) {
+ return static_cast<CpuRegister>(i);
+ }
+ }
+ // We shall never fail to find a free caller-save register, as
+ // there are more than two core caller-save registers on x86-64
+ // (meaning it is possible to find one which is different from
+ // `ref` and `obj`).
+ DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u);
+ LOG(FATAL) << "Could not find a free caller-save register";
+ UNREACHABLE();
+ }
+
+ HInstruction* const instruction_;
+ const Location out_;
+ const Location ref_;
+ const Location obj_;
+ const uint32_t offset_;
+ // An additional location containing an index to an array.
+ // Only used for HArrayGet and the UnsafeGetObject &
+ // UnsafeGetObjectVolatile intrinsics.
+ const Location index_;
+
+ DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathX86_64);
+};
+
+// Slow path generating a read barrier for a GC root.
+class ReadBarrierForRootSlowPathX86_64 : public SlowPathCode {
+ public:
+ ReadBarrierForRootSlowPathX86_64(HInstruction* instruction, Location out, Location root)
+ : instruction_(instruction), out_(out), root_(root) {}
+
+ void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+ LocationSummary* locations = instruction_->GetLocations();
+ DCHECK(locations->CanCall());
+ DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg()));
+ DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString());
+
+ __ Bind(GetEntryLabel());
+ 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)), root_);
+ x86_64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow),
+ instruction_,
+ instruction_->GetDexPc(),
+ this);
+ CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>();
+ x86_64_codegen->Move(out_, Location::RegisterLocation(RAX));
+
+ RestoreLiveRegisters(codegen, locations);
+ __ jmp(GetExitLabel());
+ }
+
+ const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathX86_64"; }
+
+ private:
+ HInstruction* const instruction_;
+ const Location out_;
+ const Location root_;
+
+ DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathX86_64);
+};
+
#undef __
#define __ down_cast<X86_64Assembler*>(GetAssembler())->
@@ -533,7 +748,7 @@
method_reg = reg.AsRegister();
__ movq(reg, Address(CpuRegister(RSP), kCurrentMethodStackOffset));
}
- // temp = temp->dex_cache_resolved_methods_;
+ // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_;
__ movq(reg,
Address(CpuRegister(method_reg),
ArtMethod::DexCacheResolvedMethodsOffset(kX86_64PointerSize).SizeValue()));
@@ -578,10 +793,17 @@
LocationSummary* locations = invoke->GetLocations();
Location receiver = locations->InAt(0);
size_t class_offset = mirror::Object::ClassOffset().SizeValue();
- // temp = object->GetClass();
DCHECK(receiver.IsRegister());
+ // /* HeapReference<Class> */ temp = receiver->klass_
__ movl(temp, Address(receiver.AsRegister<CpuRegister>(), class_offset));
MaybeRecordImplicitNullCheck(invoke);
+ // Instead of simply (possibly) unpoisoning `temp` here, we should
+ // emit a read barrier for the previous class reference load.
+ // However this is not required in practice, as this is an
+ // intermediate/temporary reference and because the current
+ // concurrent copying collector keeps the from-space memory
+ // intact/accessible until the end of the marking phase (the
+ // concurrent copying collector may not in the future).
__ MaybeUnpoisonHeapReference(temp);
// temp = temp->GetMethodAt(method_offset);
__ movq(temp, Address(temp, method_offset));
@@ -672,9 +894,9 @@
// Use a fake return address register to mimic Quick.
static constexpr Register kFakeReturnRegister = Register(kLastCpuRegister + 1);
CodeGeneratorX86_64::CodeGeneratorX86_64(HGraph* graph,
- const X86_64InstructionSetFeatures& isa_features,
- const CompilerOptions& compiler_options,
- OptimizingCompilerStats* stats)
+ const X86_64InstructionSetFeatures& isa_features,
+ const CompilerOptions& compiler_options,
+ OptimizingCompilerStats* stats)
: CodeGenerator(graph,
kNumberOfCpuRegisters,
kNumberOfFloatRegisters,
@@ -728,7 +950,7 @@
LOG(FATAL) << "Unreachable type " << type;
}
- return Location();
+ return Location::NoLocation();
}
void CodeGeneratorX86_64::SetupBlockedRegisters(bool is_baseline) const {
@@ -1836,7 +2058,7 @@
LOG(FATAL) << "Unexpected parameter type " << type;
break;
}
- return Location();
+ return Location::NoLocation();
}
void LocationsBuilderX86_64::VisitInvokeUnresolved(HInvokeUnresolved* invoke) {
@@ -1907,7 +2129,6 @@
}
codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0));
-
DCHECK(!codegen_->IsLeafMethod());
codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
}
@@ -1920,31 +2141,41 @@
void InstructionCodeGeneratorX86_64::VisitInvokeInterface(HInvokeInterface* invoke) {
// TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError.
- CpuRegister temp = invoke->GetLocations()->GetTemp(0).AsRegister<CpuRegister>();
+ LocationSummary* locations = invoke->GetLocations();
+ CpuRegister temp = locations->GetTemp(0).AsRegister<CpuRegister>();
+ CpuRegister hidden_reg = locations->GetTemp(1).AsRegister<CpuRegister>();
uint32_t method_offset = mirror::Class::EmbeddedImTableEntryOffset(
invoke->GetImtIndex() % mirror::Class::kImtSize, kX86_64PointerSize).Uint32Value();
- LocationSummary* locations = invoke->GetLocations();
Location receiver = locations->InAt(0);
size_t class_offset = mirror::Object::ClassOffset().SizeValue();
- // Set the hidden argument.
- CpuRegister hidden_reg = invoke->GetLocations()->GetTemp(1).AsRegister<CpuRegister>();
+ // Set the hidden argument. This is safe to do this here, as RAX
+ // won't be modified thereafter, before the `call` instruction.
+ DCHECK_EQ(RAX, hidden_reg.AsRegister());
codegen_->Load64BitValue(hidden_reg, invoke->GetDexMethodIndex());
- // temp = object->GetClass();
if (receiver.IsStackSlot()) {
__ movl(temp, Address(CpuRegister(RSP), receiver.GetStackIndex()));
+ // /* HeapReference<Class> */ temp = temp->klass_
__ movl(temp, Address(temp, class_offset));
} else {
+ // /* HeapReference<Class> */ temp = receiver->klass_
__ movl(temp, Address(receiver.AsRegister<CpuRegister>(), class_offset));
}
codegen_->MaybeRecordImplicitNullCheck(invoke);
+ // Instead of simply (possibly) unpoisoning `temp` here, we should
+ // emit a read barrier for the previous class reference load.
+ // However this is not required in practice, as this is an
+ // intermediate/temporary reference and because the current
+ // concurrent copying collector keeps the from-space memory
+ // intact/accessible until the end of the marking phase (the
+ // concurrent copying collector may not in the future).
__ MaybeUnpoisonHeapReference(temp);
// temp = temp->GetImtEntryAt(method_offset);
__ movq(temp, Address(temp, method_offset));
// call temp->GetEntryPoint();
- __ call(Address(temp, ArtMethod::EntryPointFromQuickCompiledCodeOffset(
- kX86_64WordSize).SizeValue()));
+ __ call(Address(temp,
+ ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86_64WordSize).SizeValue()));
DCHECK(!codegen_->IsLeafMethod());
codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
@@ -3686,13 +3917,23 @@
void LocationsBuilderX86_64::HandleFieldGet(HInstruction* instruction) {
DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
+ bool object_field_get_with_read_barrier =
+ kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot);
LocationSummary* locations =
- new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
+ new (GetGraph()->GetArena()) LocationSummary(instruction,
+ object_field_get_with_read_barrier ?
+ LocationSummary::kCallOnSlowPath :
+ LocationSummary::kNoCall);
locations->SetInAt(0, Location::RequiresRegister());
if (Primitive::IsFloatingPointType(instruction->GetType())) {
locations->SetOut(Location::RequiresFpuRegister());
} else {
- locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+ // The output overlaps for an object field get when read barriers
+ // are enabled: we do not want the move to overwrite the object's
+ // location, as we need it to emit the read barrier.
+ locations->SetOut(
+ Location::RequiresRegister(),
+ object_field_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap);
}
}
@@ -3701,7 +3942,8 @@
DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
LocationSummary* locations = instruction->GetLocations();
- CpuRegister base = locations->InAt(0).AsRegister<CpuRegister>();
+ Location base_loc = locations->InAt(0);
+ CpuRegister base = base_loc.AsRegister<CpuRegister>();
Location out = locations->Out();
bool is_volatile = field_info.IsVolatile();
Primitive::Type field_type = field_info.GetFieldType();
@@ -3761,7 +4003,7 @@
}
if (field_type == Primitive::kPrimNot) {
- __ MaybeUnpoisonHeapReference(out.AsRegister<CpuRegister>());
+ codegen_->MaybeGenerateReadBarrier(instruction, out, out, base_loc, offset);
}
}
@@ -4079,20 +4321,31 @@
}
void LocationsBuilderX86_64::VisitArrayGet(HArrayGet* instruction) {
+ bool object_array_get_with_read_barrier =
+ kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot);
LocationSummary* locations =
- new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
+ new (GetGraph()->GetArena()) LocationSummary(instruction,
+ object_array_get_with_read_barrier ?
+ LocationSummary::kCallOnSlowPath :
+ LocationSummary::kNoCall);
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
if (Primitive::IsFloatingPointType(instruction->GetType())) {
locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
} else {
- locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+ // The output overlaps for an object array get when read barriers
+ // are enabled: we do not want the move to overwrite the array's
+ // location, as we need it to emit the read barrier.
+ locations->SetOut(
+ Location::RequiresRegister(),
+ object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap);
}
}
void InstructionCodeGeneratorX86_64::VisitArrayGet(HArrayGet* instruction) {
LocationSummary* locations = instruction->GetLocations();
- CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
+ Location obj_loc = locations->InAt(0);
+ CpuRegister obj = obj_loc.AsRegister<CpuRegister>();
Location index = locations->InAt(1);
Primitive::Type type = instruction->GetType();
@@ -4147,8 +4400,9 @@
case Primitive::kPrimInt:
case Primitive::kPrimNot: {
- static_assert(sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
- "art::mirror::HeapReference<mirror::Object> and int32_t have different sizes.");
+ static_assert(
+ sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
+ "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value();
CpuRegister out = locations->Out().AsRegister<CpuRegister>();
if (index.IsConstant()) {
@@ -4203,8 +4457,17 @@
codegen_->MaybeRecordImplicitNullCheck(instruction);
if (type == Primitive::kPrimNot) {
- CpuRegister out = locations->Out().AsRegister<CpuRegister>();
- __ MaybeUnpoisonHeapReference(out);
+ static_assert(
+ sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t),
+ "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes.");
+ uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value();
+ Location out = locations->Out();
+ if (index.IsConstant()) {
+ uint32_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset;
+ codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset);
+ } else {
+ codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, data_offset, index);
+ }
}
}
@@ -4214,10 +4477,14 @@
bool needs_write_barrier =
CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
bool may_need_runtime_call = instruction->NeedsTypeCheck();
+ bool object_array_set_with_read_barrier =
+ kEmitCompilerReadBarrier && (value_type == Primitive::kPrimNot);
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(
instruction,
- may_need_runtime_call ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall);
+ (may_need_runtime_call || object_array_set_with_read_barrier) ?
+ LocationSummary::kCallOnSlowPath :
+ LocationSummary::kNoCall);
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
@@ -4229,18 +4496,24 @@
if (needs_write_barrier) {
// Temporary registers for the write barrier.
- locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too.
+
+ // This first temporary register is possibly used for heap
+ // reference poisoning and/or read barrier emission too.
+ locations->AddTemp(Location::RequiresRegister());
+ // This second temporary register is possibly used for read
+ // barrier emission too.
locations->AddTemp(Location::RequiresRegister());
}
}
void InstructionCodeGeneratorX86_64::VisitArraySet(HArraySet* instruction) {
LocationSummary* locations = instruction->GetLocations();
- CpuRegister array = locations->InAt(0).AsRegister<CpuRegister>();
+ Location array_loc = locations->InAt(0);
+ CpuRegister array = array_loc.AsRegister<CpuRegister>();
Location index = locations->InAt(1);
Location value = locations->InAt(2);
Primitive::Type value_type = instruction->GetComponentType();
- bool may_need_runtime_call = locations->CanCall();
+ bool may_need_runtime_call = instruction->NeedsTypeCheck();
bool needs_write_barrier =
CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
@@ -4284,6 +4557,7 @@
Address address = index.IsConstant()
? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset)
: Address(array, index.AsRegister<CpuRegister>(), TIMES_4, offset);
+
if (!value.IsRegister()) {
// Just setting null.
DCHECK(instruction->InputAt(2)->IsNullConstant());
@@ -4312,22 +4586,62 @@
__ Bind(¬_null);
}
- __ movl(temp, Address(array, class_offset));
- codegen_->MaybeRecordImplicitNullCheck(instruction);
- __ MaybeUnpoisonHeapReference(temp);
- __ movl(temp, Address(temp, component_offset));
- // No need to poison/unpoison, we're comparing two poisoned references.
- __ cmpl(temp, Address(register_value, class_offset));
- if (instruction->StaticTypeOfArrayIsObjectArray()) {
- __ j(kEqual, &do_put);
- __ MaybeUnpoisonHeapReference(temp);
- __ movl(temp, Address(temp, super_offset));
- // No need to unpoison the result, we're comparing against null.
- __ testl(temp, temp);
- __ j(kNotEqual, slow_path->GetEntryLabel());
- __ Bind(&do_put);
+ if (kEmitCompilerReadBarrier) {
+ // When read barriers are enabled, the type checking
+ // instrumentation requires two read barriers:
+ //
+ // __ movl(temp2, temp);
+ // // /* HeapReference<Class> */ temp = temp->component_type_
+ // __ movl(temp, Address(temp, component_offset));
+ // codegen_->GenerateReadBarrier(
+ // instruction, temp_loc, temp_loc, temp2_loc, component_offset);
+ //
+ // // /* HeapReference<Class> */ temp2 = register_value->klass_
+ // __ movl(temp2, Address(register_value, class_offset));
+ // codegen_->GenerateReadBarrier(
+ // instruction, temp2_loc, temp2_loc, value, class_offset, temp_loc);
+ //
+ // __ cmpl(temp, temp2);
+ //
+ // However, the second read barrier may trash `temp`, as it
+ // is a temporary register, and as such would not be saved
+ // along with live registers before calling the runtime (nor
+ // restored afterwards). So in this case, we bail out and
+ // delegate the work to the array set slow path.
+ //
+ // TODO: Extend the register allocator to support a new
+ // "(locally) live temp" location so as to avoid always
+ // going into the slow path when read barriers are enabled.
+ __ jmp(slow_path->GetEntryLabel());
} else {
- __ j(kNotEqual, slow_path->GetEntryLabel());
+ // /* HeapReference<Class> */ temp = array->klass_
+ __ movl(temp, Address(array, class_offset));
+ codegen_->MaybeRecordImplicitNullCheck(instruction);
+ __ MaybeUnpoisonHeapReference(temp);
+
+ // /* HeapReference<Class> */ temp = temp->component_type_
+ __ movl(temp, Address(temp, component_offset));
+ // If heap poisoning is enabled, no need to unpoison `temp`
+ // nor the object reference in `register_value->klass`, as
+ // we are comparing two poisoned references.
+ __ cmpl(temp, Address(register_value, class_offset));
+
+ if (instruction->StaticTypeOfArrayIsObjectArray()) {
+ __ j(kEqual, &do_put);
+ // If heap poisoning is enabled, the `temp` reference has
+ // not been unpoisoned yet; unpoison it now.
+ __ MaybeUnpoisonHeapReference(temp);
+
+ // /* HeapReference<Class> */ temp = temp->super_class_
+ __ movl(temp, Address(temp, super_offset));
+ // If heap poisoning is enabled, no need to unpoison
+ // `temp`, as we are comparing against null below.
+ __ testl(temp, temp);
+ __ j(kNotEqual, slow_path->GetEntryLabel());
+ __ Bind(&do_put);
+ } else {
+ __ j(kNotEqual, slow_path->GetEntryLabel());
+ }
}
}
@@ -4353,6 +4667,7 @@
break;
}
+
case Primitive::kPrimInt: {
uint32_t offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value();
Address address = index.IsConstant()
@@ -4802,7 +5117,8 @@
CodeGenerator::CreateLoadClassLocationSummary(
cls,
Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
- Location::RegisterLocation(RAX));
+ Location::RegisterLocation(RAX),
+ /* code_generator_supports_read_barrier */ true);
}
void InstructionCodeGeneratorX86_64::VisitLoadClass(HLoadClass* cls) {
@@ -4816,18 +5132,40 @@
return;
}
- CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+ Location out_loc = locations->Out();
+ CpuRegister out = out_loc.AsRegister<CpuRegister>();
CpuRegister current_method = locations->InAt(0).AsRegister<CpuRegister>();
+
if (cls->IsReferrersClass()) {
DCHECK(!cls->CanCallRuntime());
DCHECK(!cls->MustGenerateClinitCheck());
- __ movl(out, Address(current_method, ArtMethod::DeclaringClassOffset().Int32Value()));
+ uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value();
+ if (kEmitCompilerReadBarrier) {
+ // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_)
+ __ leaq(out, Address(current_method, declaring_class_offset));
+ // /* mirror::Class* */ out = out->Read()
+ codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc);
+ } else {
+ // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
+ __ movl(out, Address(current_method, declaring_class_offset));
+ }
} else {
DCHECK(cls->CanCallRuntime());
- __ movq(out, Address(
- current_method, ArtMethod::DexCacheResolvedTypesOffset(kX86_64PointerSize).Int32Value()));
- __ movl(out, Address(out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())));
- // TODO: We will need a read barrier here.
+ // /* GcRoot<mirror::Class>[] */ out =
+ // current_method.ptr_sized_fields_->dex_cache_resolved_types_
+ __ movq(out, Address(current_method,
+ ArtMethod::DexCacheResolvedTypesOffset(kX86_64PointerSize).Int32Value()));
+
+ size_t cache_offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex());
+ if (kEmitCompilerReadBarrier) {
+ // /* GcRoot<mirror::Class>* */ out = &out[type_index]
+ __ leaq(out, Address(out, cache_offset));
+ // /* mirror::Class* */ out = out->Read()
+ codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc);
+ } else {
+ // /* GcRoot<mirror::Class> */ out = out[type_index]
+ __ movl(out, Address(out, cache_offset));
+ }
SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86_64(
cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck());
@@ -4872,12 +5210,35 @@
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = load->GetLocations();
- CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+ Location out_loc = locations->Out();
+ CpuRegister out = out_loc.AsRegister<CpuRegister>();
CpuRegister current_method = locations->InAt(0).AsRegister<CpuRegister>();
- __ movl(out, Address(current_method, ArtMethod::DeclaringClassOffset().Int32Value()));
- __ movq(out, Address(out, mirror::Class::DexCacheStringsOffset().Int32Value()));
- __ movl(out, Address(out, CodeGenerator::GetCacheOffset(load->GetStringIndex())));
- // TODO: We will need a read barrier here.
+
+ uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value();
+ if (kEmitCompilerReadBarrier) {
+ // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_)
+ __ leaq(out, Address(current_method, declaring_class_offset));
+ // /* mirror::Class* */ out = out->Read()
+ codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc);
+ } else {
+ // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_
+ __ movl(out, Address(current_method, declaring_class_offset));
+ }
+
+ // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_
+ __ movq(out, Address(out, mirror::Class::DexCacheStringsOffset().Uint32Value()));
+
+ size_t cache_offset = CodeGenerator::GetCacheOffset(load->GetStringIndex());
+ if (kEmitCompilerReadBarrier) {
+ // /* GcRoot<mirror::String>* */ out = &out[string_index]
+ __ leaq(out, Address(out, cache_offset));
+ // /* mirror::String* */ out = out->Read()
+ codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc);
+ } else {
+ // /* GcRoot<mirror::String> */ out = out[string_index]
+ __ movl(out, Address(out, cache_offset));
+ }
+
__ testl(out, out);
__ j(kEqual, slow_path->GetEntryLabel());
__ Bind(slow_path->GetExitLabel());
@@ -4921,40 +5282,44 @@
void LocationsBuilderX86_64::VisitInstanceOf(HInstanceOf* instruction) {
LocationSummary::CallKind call_kind = LocationSummary::kNoCall;
- switch (instruction->GetTypeCheckKind()) {
+ TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
+ switch (type_check_kind) {
case TypeCheckKind::kExactCheck:
case TypeCheckKind::kAbstractClassCheck:
case TypeCheckKind::kClassHierarchyCheck:
case TypeCheckKind::kArrayObjectCheck:
- call_kind = LocationSummary::kNoCall;
- break;
- case TypeCheckKind::kUnresolvedCheck:
- case TypeCheckKind::kInterfaceCheck:
- call_kind = LocationSummary::kCall;
+ call_kind =
+ kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall;
break;
case TypeCheckKind::kArrayCheck:
+ case TypeCheckKind::kUnresolvedCheck:
+ case TypeCheckKind::kInterfaceCheck:
call_kind = LocationSummary::kCallOnSlowPath;
break;
}
+
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind);
- if (call_kind != LocationSummary::kCall) {
- locations->SetInAt(0, Location::RequiresRegister());
- locations->SetInAt(1, Location::Any());
- // Note that TypeCheckSlowPathX86_64 uses this register too.
- locations->SetOut(Location::RequiresRegister());
- } else {
- InvokeRuntimeCallingConvention calling_convention;
- locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
- locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
- locations->SetOut(Location::RegisterLocation(RAX));
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::Any());
+ // Note that TypeCheckSlowPathX86_64 uses this "out" register too.
+ locations->SetOut(Location::RequiresRegister());
+ // When read barriers are enabled, we need a temporary register for
+ // some cases.
+ if (kEmitCompilerReadBarrier &&
+ (type_check_kind == TypeCheckKind::kAbstractClassCheck ||
+ type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
+ type_check_kind == TypeCheckKind::kArrayObjectCheck)) {
+ locations->AddTemp(Location::RequiresRegister());
}
}
void InstructionCodeGeneratorX86_64::VisitInstanceOf(HInstanceOf* instruction) {
LocationSummary* locations = instruction->GetLocations();
- CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
+ Location obj_loc = locations->InAt(0);
+ CpuRegister obj = obj_loc.AsRegister<CpuRegister>();
Location cls = locations->InAt(1);
- CpuRegister out = locations->Out().AsRegister<CpuRegister>();
+ Location out_loc = locations->Out();
+ CpuRegister out = out_loc.AsRegister<CpuRegister>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
@@ -4969,15 +5334,9 @@
__ j(kEqual, &zero);
}
- // In case of an interface/unresolved check, we put the object class into the object register.
- // This is safe, as the register is caller-save, and the object must be in another
- // register if it survives the runtime call.
- CpuRegister target = (instruction->GetTypeCheckKind() == TypeCheckKind::kInterfaceCheck) ||
- (instruction->GetTypeCheckKind() == TypeCheckKind::kUnresolvedCheck)
- ? obj
- : out;
- __ movl(target, Address(obj, class_offset));
- __ MaybeUnpoisonHeapReference(target);
+ // /* HeapReference<Class> */ out = obj->klass_
+ __ movl(out, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, obj_loc, class_offset);
switch (instruction->GetTypeCheckKind()) {
case TypeCheckKind::kExactCheck: {
@@ -4999,13 +5358,23 @@
}
break;
}
+
case TypeCheckKind::kAbstractClassCheck: {
// If the class is abstract, we eagerly fetch the super class of the
// object to avoid doing a comparison we know will fail.
NearLabel loop, success;
__ Bind(&loop);
+ Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `out` into `temp` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp = temp_loc.AsRegister<CpuRegister>();
+ __ movl(temp, out);
+ }
+ // /* HeapReference<Class> */ out = out->super_class_
__ movl(out, Address(out, super_offset));
- __ MaybeUnpoisonHeapReference(out);
+ codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset);
__ testl(out, out);
// If `out` is null, we use it for the result, and jump to `done`.
__ j(kEqual, &done);
@@ -5022,6 +5391,7 @@
}
break;
}
+
case TypeCheckKind::kClassHierarchyCheck: {
// Walk over the class hierarchy to find a match.
NearLabel loop, success;
@@ -5033,8 +5403,17 @@
__ cmpl(out, Address(CpuRegister(RSP), cls.GetStackIndex()));
}
__ j(kEqual, &success);
+ Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `out` into `temp` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp = temp_loc.AsRegister<CpuRegister>();
+ __ movl(temp, out);
+ }
+ // /* HeapReference<Class> */ out = out->super_class_
__ movl(out, Address(out, super_offset));
- __ MaybeUnpoisonHeapReference(out);
+ codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset);
__ testl(out, out);
__ j(kNotEqual, &loop);
// If `out` is null, we use it for the result, and jump to `done`.
@@ -5046,6 +5425,7 @@
}
break;
}
+
case TypeCheckKind::kArrayObjectCheck: {
// Do an exact check.
NearLabel exact_check;
@@ -5056,9 +5436,18 @@
__ cmpl(out, Address(CpuRegister(RSP), cls.GetStackIndex()));
}
__ j(kEqual, &exact_check);
- // Otherwise, we need to check that the object's class is a non primitive array.
+ // Otherwise, we need to check that the object's class is a non-primitive array.
+ Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `out` into `temp` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp = temp_loc.AsRegister<CpuRegister>();
+ __ movl(temp, out);
+ }
+ // /* HeapReference<Class> */ out = out->component_type_
__ movl(out, Address(out, component_offset));
- __ MaybeUnpoisonHeapReference(out);
+ codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, component_offset);
__ testl(out, out);
// If `out` is null, we use it for the result, and jump to `done`.
__ j(kEqual, &done);
@@ -5069,6 +5458,7 @@
__ jmp(&done);
break;
}
+
case TypeCheckKind::kArrayCheck: {
if (cls.IsRegister()) {
__ cmpl(out, cls.AsRegister<CpuRegister>());
@@ -5077,8 +5467,8 @@
__ cmpl(out, Address(CpuRegister(RSP), cls.GetStackIndex()));
}
DCHECK(locations->OnlyCallsOnSlowPath());
- slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(
- instruction, /* is_fatal */ false);
+ slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(instruction,
+ /* is_fatal */ false);
codegen_->AddSlowPath(slow_path);
__ j(kNotEqual, slow_path->GetEntryLabel());
__ movl(out, Immediate(1));
@@ -5087,13 +5477,25 @@
}
break;
}
+
case TypeCheckKind::kUnresolvedCheck:
- case TypeCheckKind::kInterfaceCheck:
- default: {
- codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial),
- instruction,
- instruction->GetDexPc(),
- nullptr);
+ case TypeCheckKind::kInterfaceCheck: {
+ // Note that we indeed only call on slow path, but we always go
+ // into the slow path for the unresolved & interface check
+ // cases.
+ //
+ // We cannot directly call the InstanceofNonTrivial runtime
+ // entry point without resorting to a type checking slow path
+ // here (i.e. by calling InvokeRuntime directly), as it would
+ // require to assign fixed registers for the inputs of this
+ // HInstanceOf instruction (following the runtime calling
+ // convention), which might be cluttered by the potential first
+ // read barrier emission at the beginning of this method.
+ DCHECK(locations->OnlyCallsOnSlowPath());
+ slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(instruction,
+ /* is_fatal */ false);
+ codegen_->AddSlowPath(slow_path);
+ __ jmp(slow_path->GetEntryLabel());
if (zero.IsLinked()) {
__ jmp(&done);
}
@@ -5118,58 +5520,60 @@
void LocationsBuilderX86_64::VisitCheckCast(HCheckCast* instruction) {
LocationSummary::CallKind call_kind = LocationSummary::kNoCall;
bool throws_into_catch = instruction->CanThrowIntoCatchBlock();
-
- switch (instruction->GetTypeCheckKind()) {
+ TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
+ switch (type_check_kind) {
case TypeCheckKind::kExactCheck:
case TypeCheckKind::kAbstractClassCheck:
case TypeCheckKind::kClassHierarchyCheck:
case TypeCheckKind::kArrayObjectCheck:
- call_kind = throws_into_catch
- ? LocationSummary::kCallOnSlowPath
- : LocationSummary::kNoCall;
- break;
- case TypeCheckKind::kUnresolvedCheck:
- case TypeCheckKind::kInterfaceCheck:
- call_kind = LocationSummary::kCall;
+ call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ?
+ LocationSummary::kCallOnSlowPath :
+ LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path.
break;
case TypeCheckKind::kArrayCheck:
+ case TypeCheckKind::kUnresolvedCheck:
+ case TypeCheckKind::kInterfaceCheck:
call_kind = LocationSummary::kCallOnSlowPath;
break;
}
-
- LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(
- instruction, call_kind);
- if (call_kind != LocationSummary::kCall) {
- locations->SetInAt(0, Location::RequiresRegister());
- locations->SetInAt(1, Location::Any());
- // Note that TypeCheckSlowPathX86_64 uses this register too.
+ LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::Any());
+ // Note that TypeCheckSlowPathX86_64 uses this "temp" register too.
+ locations->AddTemp(Location::RequiresRegister());
+ // When read barriers are enabled, we need an additional temporary
+ // register for some cases.
+ if (kEmitCompilerReadBarrier &&
+ (type_check_kind == TypeCheckKind::kAbstractClassCheck ||
+ type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
+ type_check_kind == TypeCheckKind::kArrayObjectCheck)) {
locations->AddTemp(Location::RequiresRegister());
- } else {
- InvokeRuntimeCallingConvention calling_convention;
- locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
- locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
}
}
void InstructionCodeGeneratorX86_64::VisitCheckCast(HCheckCast* instruction) {
LocationSummary* locations = instruction->GetLocations();
- CpuRegister obj = locations->InAt(0).AsRegister<CpuRegister>();
+ Location obj_loc = locations->InAt(0);
+ CpuRegister obj = obj_loc.AsRegister<CpuRegister>();
Location cls = locations->InAt(1);
- CpuRegister temp = locations->WillCall()
- ? CpuRegister(kNoRegister)
- : locations->GetTemp(0).AsRegister<CpuRegister>();
-
+ Location temp_loc = locations->GetTemp(0);
+ CpuRegister temp = temp_loc.AsRegister<CpuRegister>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value();
uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value();
uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value();
- SlowPathCode* slow_path = nullptr;
- if (!locations->WillCall()) {
- slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(
- instruction, !locations->CanCall());
- codegen_->AddSlowPath(slow_path);
- }
+ TypeCheckKind type_check_kind = instruction->GetTypeCheckKind();
+ bool is_type_check_slow_path_fatal =
+ (type_check_kind == TypeCheckKind::kExactCheck ||
+ type_check_kind == TypeCheckKind::kAbstractClassCheck ||
+ type_check_kind == TypeCheckKind::kClassHierarchyCheck ||
+ type_check_kind == TypeCheckKind::kArrayObjectCheck) &&
+ !instruction->CanThrowIntoCatchBlock();
+ SlowPathCode* type_check_slow_path =
+ new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(instruction,
+ is_type_check_slow_path_fatal);
+ codegen_->AddSlowPath(type_check_slow_path);
NearLabel done;
// Avoid null check if we know obj is not null.
@@ -5178,15 +5582,11 @@
__ j(kEqual, &done);
}
- if (locations->WillCall()) {
- __ movl(obj, Address(obj, class_offset));
- __ MaybeUnpoisonHeapReference(obj);
- } else {
- __ movl(temp, Address(obj, class_offset));
- __ MaybeUnpoisonHeapReference(temp);
- }
+ // /* HeapReference<Class> */ temp = obj->klass_
+ __ movl(temp, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
- switch (instruction->GetTypeCheckKind()) {
+ switch (type_check_kind) {
case TypeCheckKind::kExactCheck:
case TypeCheckKind::kArrayCheck: {
if (cls.IsRegister()) {
@@ -5197,19 +5597,44 @@
}
// Jump to slow path for throwing the exception or doing a
// more involved array check.
- __ j(kNotEqual, slow_path->GetEntryLabel());
+ __ j(kNotEqual, type_check_slow_path->GetEntryLabel());
break;
}
+
case TypeCheckKind::kAbstractClassCheck: {
// If the class is abstract, we eagerly fetch the super class of the
// object to avoid doing a comparison we know will fail.
- NearLabel loop;
+ NearLabel loop, compare_classes;
__ Bind(&loop);
+ Location temp2_loc =
+ kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `temp` into `temp2` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp2 = temp2_loc.AsRegister<CpuRegister>();
+ __ movl(temp2, temp);
+ }
+ // /* HeapReference<Class> */ temp = temp->super_class_
__ movl(temp, Address(temp, super_offset));
- __ MaybeUnpoisonHeapReference(temp);
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset);
+
+ // If the class reference currently in `temp` is not null, jump
+ // to the `compare_classes` label to compare it with the checked
+ // class.
__ testl(temp, temp);
- // Jump to the slow path to throw the exception.
- __ j(kEqual, slow_path->GetEntryLabel());
+ __ j(kNotEqual, &compare_classes);
+ // Otherwise, jump to the slow path to throw the exception.
+ //
+ // But before, move back the object's class into `temp` before
+ // going into the slow path, as it has been overwritten in the
+ // meantime.
+ // /* HeapReference<Class> */ temp = obj->klass_
+ __ movl(temp, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+ __ jmp(type_check_slow_path->GetEntryLabel());
+
+ __ Bind(&compare_classes);
if (cls.IsRegister()) {
__ cmpl(temp, cls.AsRegister<CpuRegister>());
} else {
@@ -5219,6 +5644,7 @@
__ j(kNotEqual, &loop);
break;
}
+
case TypeCheckKind::kClassHierarchyCheck: {
// Walk over the class hierarchy to find a match.
NearLabel loop;
@@ -5230,16 +5656,39 @@
__ cmpl(temp, Address(CpuRegister(RSP), cls.GetStackIndex()));
}
__ j(kEqual, &done);
+
+ Location temp2_loc =
+ kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `temp` into `temp2` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp2 = temp2_loc.AsRegister<CpuRegister>();
+ __ movl(temp2, temp);
+ }
+ // /* HeapReference<Class> */ temp = temp->super_class_
__ movl(temp, Address(temp, super_offset));
- __ MaybeUnpoisonHeapReference(temp);
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset);
+
+ // If the class reference currently in `temp` is not null, jump
+ // back at the beginning of the loop.
__ testl(temp, temp);
__ j(kNotEqual, &loop);
- // Jump to the slow path to throw the exception.
- __ jmp(slow_path->GetEntryLabel());
+ // Otherwise, jump to the slow path to throw the exception.
+ //
+ // But before, move back the object's class into `temp` before
+ // going into the slow path, as it has been overwritten in the
+ // meantime.
+ // /* HeapReference<Class> */ temp = obj->klass_
+ __ movl(temp, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+ __ jmp(type_check_slow_path->GetEntryLabel());
break;
}
+
case TypeCheckKind::kArrayObjectCheck: {
// Do an exact check.
+ NearLabel check_non_primitive_component_type;
if (cls.IsRegister()) {
__ cmpl(temp, cls.AsRegister<CpuRegister>());
} else {
@@ -5247,29 +5696,67 @@
__ cmpl(temp, Address(CpuRegister(RSP), cls.GetStackIndex()));
}
__ j(kEqual, &done);
- // Otherwise, we need to check that the object's class is a non primitive array.
+
+ // Otherwise, we need to check that the object's class is a non-primitive array.
+ Location temp2_loc =
+ kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation();
+ if (kEmitCompilerReadBarrier) {
+ // Save the value of `temp` into `temp2` before overwriting it
+ // in the following move operation, as we will need it for the
+ // read barrier below.
+ CpuRegister temp2 = temp2_loc.AsRegister<CpuRegister>();
+ __ movl(temp2, temp);
+ }
+ // /* HeapReference<Class> */ temp = temp->component_type_
__ movl(temp, Address(temp, component_offset));
- __ MaybeUnpoisonHeapReference(temp);
+ codegen_->MaybeGenerateReadBarrier(
+ instruction, temp_loc, temp_loc, temp2_loc, component_offset);
+
+ // If the component type is not null (i.e. the object is indeed
+ // an array), jump to label `check_non_primitive_component_type`
+ // to further check that this component type is not a primitive
+ // type.
__ testl(temp, temp);
- __ j(kEqual, slow_path->GetEntryLabel());
+ __ j(kNotEqual, &check_non_primitive_component_type);
+ // Otherwise, jump to the slow path to throw the exception.
+ //
+ // But before, move back the object's class into `temp` before
+ // going into the slow path, as it has been overwritten in the
+ // meantime.
+ // /* HeapReference<Class> */ temp = obj->klass_
+ __ movl(temp, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+ __ jmp(type_check_slow_path->GetEntryLabel());
+
+ __ Bind(&check_non_primitive_component_type);
__ cmpw(Address(temp, primitive_offset), Immediate(Primitive::kPrimNot));
- __ j(kNotEqual, slow_path->GetEntryLabel());
+ __ j(kEqual, &done);
+ // Same comment as above regarding `temp` and the slow path.
+ // /* HeapReference<Class> */ temp = obj->klass_
+ __ movl(temp, Address(obj, class_offset));
+ codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset);
+ __ jmp(type_check_slow_path->GetEntryLabel());
break;
}
+
case TypeCheckKind::kUnresolvedCheck:
case TypeCheckKind::kInterfaceCheck:
- default:
- codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast),
- instruction,
- instruction->GetDexPc(),
- nullptr);
+ // We always go into the type check slow path for the unresolved &
+ // interface check cases.
+ //
+ // We cannot directly call the CheckCast runtime entry point
+ // without resorting to a type checking slow path here (i.e. by
+ // calling InvokeRuntime directly), as it would require to
+ // assign fixed registers for the inputs of this HInstanceOf
+ // instruction (following the runtime calling convention), which
+ // might be cluttered by the potential first read barrier
+ // emission at the beginning of this method.
+ __ jmp(type_check_slow_path->GetEntryLabel());
break;
}
__ Bind(&done);
- if (slow_path != nullptr) {
- __ Bind(slow_path->GetExitLabel());
- }
+ __ Bind(type_check_slow_path->GetExitLabel());
}
void LocationsBuilderX86_64::VisitMonitorOperation(HMonitorOperation* instruction) {
@@ -5402,6 +5889,82 @@
}
}
+void CodeGeneratorX86_64::GenerateReadBarrier(HInstruction* instruction,
+ Location out,
+ Location ref,
+ Location obj,
+ uint32_t offset,
+ Location index) {
+ DCHECK(kEmitCompilerReadBarrier);
+
+ // If heap poisoning is enabled, the unpoisoning of the loaded
+ // reference will be carried out by the runtime within the slow
+ // path.
+ //
+ // Note that `ref` currently does not get unpoisoned (when heap
+ // poisoning is enabled), which is alright as the `ref` argument is
+ // not used by the artReadBarrierSlow entry point.
+ //
+ // TODO: Unpoison `ref` when it is used by artReadBarrierSlow.
+ SlowPathCode* slow_path = new (GetGraph()->GetArena())
+ ReadBarrierForHeapReferenceSlowPathX86_64(instruction, out, ref, obj, offset, index);
+ AddSlowPath(slow_path);
+
+ // TODO: When read barrier has a fast path, add it here.
+ /* Currently the read barrier call is inserted after the original load.
+ * However, if we have a fast path, we need to perform the load of obj.LockWord *before* the
+ * original load. This load-load ordering is required by the read barrier.
+ * The fast path/slow path (for Baker's algorithm) should look like:
+ *
+ * bool isGray = obj.LockWord & kReadBarrierMask;
+ * lfence; // load fence or artificial data dependence to prevent load-load reordering
+ * ref = obj.field; // this is the original load
+ * if (isGray) {
+ * ref = Mark(ref); // ideally the slow path just does Mark(ref)
+ * }
+ */
+
+ __ jmp(slow_path->GetEntryLabel());
+ __ Bind(slow_path->GetExitLabel());
+}
+
+void CodeGeneratorX86_64::MaybeGenerateReadBarrier(HInstruction* instruction,
+ Location out,
+ Location ref,
+ Location obj,
+ uint32_t offset,
+ Location index) {
+ if (kEmitCompilerReadBarrier) {
+ // If heap poisoning is enabled, unpoisoning will be taken care of
+ // by the runtime within the slow path.
+ GenerateReadBarrier(instruction, out, ref, obj, offset, index);
+ } else if (kPoisonHeapReferences) {
+ __ UnpoisonHeapReference(out.AsRegister<CpuRegister>());
+ }
+}
+
+void CodeGeneratorX86_64::GenerateReadBarrierForRoot(HInstruction* instruction,
+ Location out,
+ Location root) {
+ DCHECK(kEmitCompilerReadBarrier);
+
+ // Note that GC roots are not affected by heap poisoning, so we do
+ // not need to do anything special for this here.
+ SlowPathCode* slow_path =
+ new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathX86_64(instruction, out, root);
+ AddSlowPath(slow_path);
+
+ // TODO: Implement a fast path for ReadBarrierForRoot, performing
+ // the following operation (for Baker's algorithm):
+ //
+ // if (thread.tls32_.is_gc_marking) {
+ // root = Mark(root);
+ // }
+
+ __ jmp(slow_path->GetEntryLabel());
+ __ Bind(slow_path->GetExitLabel());
+}
+
void LocationsBuilderX86_64::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) {
// Nothing to do, this should be removed during prepare for register allocator.
LOG(FATAL) << "Unreachable";