ART: Refactor intrinsics slow-paths
Refactor slow paths so that there is a default implementation for
common cases (only arm64 with vixl is special). Write a generic
intrinsic slow-path that can be reused for the specific architectures.
Move helper functions into CodeGenerator so that they are accessible.
Change-Id: Ibd788dce432601c6a9f7e6f13eab31f28dcb8550
diff --git a/compiler/optimizing/code_generator.h b/compiler/optimizing/code_generator.h
index a54dbf1..a1c6db0 100644
--- a/compiler/optimizing/code_generator.h
+++ b/compiler/optimizing/code_generator.h
@@ -30,6 +30,7 @@
#include "nodes.h"
#include "optimizing_compiler_stats.h"
#include "stack_map_stream.h"
+#include "utils/label.h"
namespace art {
@@ -105,6 +106,9 @@
virtual const char* GetDescription() const = 0;
+ Label* GetEntryLabel() { return &entry_label_; }
+ Label* GetExitLabel() { return &exit_label_; }
+
protected:
static constexpr size_t kMaximumNumberOfExpectedRegisters = 32;
static constexpr uint32_t kRegisterNotSaved = -1;
@@ -112,6 +116,9 @@
uint32_t saved_fpu_stack_offsets_[kMaximumNumberOfExpectedRegisters];
private:
+ Label entry_label_;
+ Label exit_label_;
+
DISALLOW_COPY_AND_ASSIGN(SlowPathCode);
};
@@ -386,6 +393,14 @@
uint32_t dex_pc,
SlowPathCode* slow_path) = 0;
+ // Generate a call to a static or direct method.
+ virtual void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) = 0;
+ // Generate a call to a virtual method.
+ virtual void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) = 0;
+
+ // Copy the result of a call into the given target.
+ virtual void MoveFromReturnRegister(Location trg, Primitive::Type type) = 0;
+
protected:
// Method patch info used for recording locations of required linker patches and
// target methods. The target method can be used for various purposes, whether for
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 6f89293..55c9214 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -59,7 +59,7 @@
#define __ down_cast<ArmAssembler*>(codegen->GetAssembler())->
#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmWordSize, x).Int32Value()
-class NullCheckSlowPathARM : public SlowPathCodeARM {
+class NullCheckSlowPathARM : public SlowPathCode {
public:
explicit NullCheckSlowPathARM(HNullCheck* instruction) : instruction_(instruction) {}
@@ -83,7 +83,7 @@
DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARM);
};
-class DivZeroCheckSlowPathARM : public SlowPathCodeARM {
+class DivZeroCheckSlowPathARM : public SlowPathCode {
public:
explicit DivZeroCheckSlowPathARM(HDivZeroCheck* instruction) : instruction_(instruction) {}
@@ -107,7 +107,7 @@
DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARM);
};
-class SuspendCheckSlowPathARM : public SlowPathCodeARM {
+class SuspendCheckSlowPathARM : public SlowPathCode {
public:
SuspendCheckSlowPathARM(HSuspendCheck* instruction, HBasicBlock* successor)
: instruction_(instruction), successor_(successor) {}
@@ -148,7 +148,7 @@
DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM);
};
-class BoundsCheckSlowPathARM : public SlowPathCodeARM {
+class BoundsCheckSlowPathARM : public SlowPathCode {
public:
explicit BoundsCheckSlowPathARM(HBoundsCheck* instruction)
: instruction_(instruction) {}
@@ -186,7 +186,7 @@
DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM);
};
-class LoadClassSlowPathARM : public SlowPathCodeARM {
+class LoadClassSlowPathARM : public SlowPathCode {
public:
LoadClassSlowPathARM(HLoadClass* cls,
HInstruction* at,
@@ -239,7 +239,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM);
};
-class LoadStringSlowPathARM : public SlowPathCodeARM {
+class LoadStringSlowPathARM : public SlowPathCode {
public:
explicit LoadStringSlowPathARM(HLoadString* instruction) : instruction_(instruction) {}
@@ -269,7 +269,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM);
};
-class TypeCheckSlowPathARM : public SlowPathCodeARM {
+class TypeCheckSlowPathARM : public SlowPathCode {
public:
explicit TypeCheckSlowPathARM(HInstruction* instruction) : instruction_(instruction) {}
@@ -321,7 +321,7 @@
DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM);
};
-class DeoptimizationSlowPathARM : public SlowPathCodeARM {
+class DeoptimizationSlowPathARM : public SlowPathCode {
public:
explicit DeoptimizationSlowPathARM(HInstruction* instruction)
: instruction_(instruction) {}
@@ -1261,7 +1261,7 @@
}
void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) {
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena())
+ SlowPathCode* slow_path = new (GetGraph()->GetArena())
DeoptimizationSlowPathARM(deoptimize);
codegen_->AddSlowPath(slow_path);
Label* slow_path_entry = slow_path->GetEntryLabel();
@@ -2792,7 +2792,7 @@
}
void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) {
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = instruction->GetLocations();
@@ -3558,7 +3558,7 @@
}
void InstructionCodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) {
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = instruction->GetLocations();
@@ -3920,7 +3920,7 @@
void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) {
LocationSummary* locations = instruction->GetLocations();
- SlowPathCodeARM* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction);
codegen_->AddSlowPath(slow_path);
@@ -4256,7 +4256,7 @@
__ LoadFromOffset(kLoadWord, out, out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex()));
// TODO: We will need a read barrier here.
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM(
cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck());
codegen_->AddSlowPath(slow_path);
__ CompareAndBranchIfZero(out, slow_path->GetEntryLabel());
@@ -4279,7 +4279,7 @@
void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) {
// We assume the class is not null.
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM(
check->GetLoadClass(), check, check->GetDexPc(), true);
codegen_->AddSlowPath(slow_path);
GenerateClassInitializationCheck(slow_path,
@@ -4287,7 +4287,7 @@
}
void InstructionCodeGeneratorARM::GenerateClassInitializationCheck(
- SlowPathCodeARM* slow_path, Register class_reg) {
+ SlowPathCode* slow_path, Register class_reg) {
__ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value());
__ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized));
__ b(slow_path->GetEntryLabel(), LT);
@@ -4305,7 +4305,7 @@
}
void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) {
- SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = load->GetLocations();
@@ -4375,7 +4375,7 @@
Register out = locations->Out().AsRegister<Register>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
Label done, zero;
- SlowPathCodeARM* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
// Return 0 if `obj` is null.
// avoid null check if we know obj is not null.
@@ -4428,8 +4428,7 @@
Register temp = locations->GetTemp(0).AsRegister<Register>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- SlowPathCodeARM* slow_path =
- new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction);
codegen_->AddSlowPath(slow_path);
// avoid null check if we know obj is not null.
@@ -4732,6 +4731,39 @@
// Will be generated at use site.
}
+void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) {
+ if (!trg.IsValid()) {
+ DCHECK(type == Primitive::kPrimVoid);
+ return;
+ }
+
+ DCHECK_NE(type, Primitive::kPrimVoid);
+
+ Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type);
+ if (return_loc.Equals(trg)) {
+ return;
+ }
+
+ // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged
+ // with the last branch.
+ if (type == Primitive::kPrimLong) {
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr);
+ parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+ } else if (type == Primitive::kPrimDouble) {
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr);
+ parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+ } else {
+ // Let the parallel move resolver take care of all of this.
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc, trg, type, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+ }
+}
+
#undef __
#undef QUICK_ENTRY_POINT
diff --git a/compiler/optimizing/code_generator_arm.h b/compiler/optimizing/code_generator_arm.h
index 91cfd00..1d98789 100644
--- a/compiler/optimizing/code_generator_arm.h
+++ b/compiler/optimizing/code_generator_arm.h
@@ -28,7 +28,6 @@
namespace arm {
class CodeGeneratorARM;
-class SlowPathCodeARM;
// Use a local definition to prevent copying mistakes.
static constexpr size_t kArmWordSize = kArmPointerSize;
@@ -118,20 +117,6 @@
DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverARM);
};
-class SlowPathCodeARM : public SlowPathCode {
- public:
- SlowPathCodeARM() : entry_label_(), exit_label_() {}
-
- Label* GetEntryLabel() { return &entry_label_; }
- Label* GetExitLabel() { return &exit_label_; }
-
- private:
- Label entry_label_;
- Label exit_label_;
-
- DISALLOW_COPY_AND_ASSIGN(SlowPathCodeARM);
-};
-
class LocationsBuilderARM : public HGraphVisitor {
public:
LocationsBuilderARM(HGraph* graph, CodeGeneratorARM* codegen)
@@ -187,7 +172,7 @@
// is the block to branch to if the suspend check is not needed, and after
// the suspend call.
void GenerateSuspendCheck(HSuspendCheck* check, HBasicBlock* successor);
- void GenerateClassInitializationCheck(SlowPathCodeARM* slow_path, Register class_reg);
+ void GenerateClassInitializationCheck(SlowPathCode* slow_path, Register class_reg);
void HandleBitwiseOperation(HBinaryOperation* operation);
void HandleShift(HBinaryOperation* operation);
void GenerateMemoryBarrier(MemBarrierKind kind);
@@ -335,8 +320,10 @@
Label* GetFrameEntryLabel() { return &frame_entry_label_; }
- void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
- void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp);
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
+ void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
+
+ void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index 576406e..8967108 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -365,8 +365,13 @@
return false;
}
- void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
- void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp);
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
+ void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
+
+ void MoveFromReturnRegister(Location trg ATTRIBUTE_UNUSED,
+ Primitive::Type type ATTRIBUTE_UNUSED) OVERRIDE {
+ UNIMPLEMENTED(FATAL);
+ }
void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
diff --git a/compiler/optimizing/code_generator_mips64.h b/compiler/optimizing/code_generator_mips64.h
index 8511eb6..f66ecb3 100644
--- a/compiler/optimizing/code_generator_mips64.h
+++ b/compiler/optimizing/code_generator_mips64.h
@@ -302,7 +302,16 @@
bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const { return false; }
- void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
+ void GenerateVirtualCall(HInvokeVirtual* invoke ATTRIBUTE_UNUSED,
+ Location temp ATTRIBUTE_UNUSED) OVERRIDE {
+ UNIMPLEMENTED(FATAL);
+ }
+
+ void MoveFromReturnRegister(Location trg ATTRIBUTE_UNUSED,
+ Primitive::Type type ATTRIBUTE_UNUSED) OVERRIDE {
+ UNIMPLEMENTED(FATAL);
+ }
private:
// Labels for each block that will be compiled.
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 9c5ecc3..09e939d 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -49,7 +49,7 @@
#define __ down_cast<X86Assembler*>(codegen->GetAssembler())->
#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kX86WordSize, x).Int32Value()
-class NullCheckSlowPathX86 : public SlowPathCodeX86 {
+class NullCheckSlowPathX86 : public SlowPathCode {
public:
explicit NullCheckSlowPathX86(HNullCheck* instruction) : instruction_(instruction) {}
@@ -75,7 +75,7 @@
DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathX86);
};
-class DivZeroCheckSlowPathX86 : public SlowPathCodeX86 {
+class DivZeroCheckSlowPathX86 : public SlowPathCode {
public:
explicit DivZeroCheckSlowPathX86(HDivZeroCheck* instruction) : instruction_(instruction) {}
@@ -101,7 +101,7 @@
DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathX86);
};
-class DivRemMinusOneSlowPathX86 : public SlowPathCodeX86 {
+class DivRemMinusOneSlowPathX86 : public SlowPathCode {
public:
DivRemMinusOneSlowPathX86(Register reg, bool is_div) : reg_(reg), is_div_(is_div) {}
@@ -123,7 +123,7 @@
DISALLOW_COPY_AND_ASSIGN(DivRemMinusOneSlowPathX86);
};
-class BoundsCheckSlowPathX86 : public SlowPathCodeX86 {
+class BoundsCheckSlowPathX86 : public SlowPathCode {
public:
explicit BoundsCheckSlowPathX86(HBoundsCheck* instruction) : instruction_(instruction) {}
@@ -161,7 +161,7 @@
DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathX86);
};
-class SuspendCheckSlowPathX86 : public SlowPathCodeX86 {
+class SuspendCheckSlowPathX86 : public SlowPathCode {
public:
SuspendCheckSlowPathX86(HSuspendCheck* instruction, HBasicBlock* successor)
: instruction_(instruction), successor_(successor) {}
@@ -201,7 +201,7 @@
DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathX86);
};
-class LoadStringSlowPathX86 : public SlowPathCodeX86 {
+class LoadStringSlowPathX86 : public SlowPathCode {
public:
explicit LoadStringSlowPathX86(HLoadString* instruction) : instruction_(instruction) {}
@@ -233,7 +233,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathX86);
};
-class LoadClassSlowPathX86 : public SlowPathCodeX86 {
+class LoadClassSlowPathX86 : public SlowPathCode {
public:
LoadClassSlowPathX86(HLoadClass* cls,
HInstruction* at,
@@ -285,7 +285,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathX86);
};
-class TypeCheckSlowPathX86 : public SlowPathCodeX86 {
+class TypeCheckSlowPathX86 : public SlowPathCode {
public:
explicit TypeCheckSlowPathX86(HInstruction* instruction) : instruction_(instruction) {}
@@ -340,7 +340,7 @@
DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathX86);
};
-class DeoptimizationSlowPathX86 : public SlowPathCodeX86 {
+class DeoptimizationSlowPathX86 : public SlowPathCode {
public:
explicit DeoptimizationSlowPathX86(HInstruction* instruction)
: instruction_(instruction) {}
@@ -1197,7 +1197,7 @@
}
void InstructionCodeGeneratorX86::VisitDeoptimize(HDeoptimize* deoptimize) {
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena())
+ SlowPathCode* slow_path = new (GetGraph()->GetArena())
DeoptimizationSlowPathX86(deoptimize);
codegen_->AddSlowPath(slow_path);
Label* slow_path_entry = slow_path->GetEntryLabel();
@@ -2870,7 +2870,7 @@
GenerateDivRemWithAnyConstant(instruction);
}
} else {
- SlowPathCodeX86* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) DivRemMinusOneSlowPathX86(out.AsRegister<Register>(),
is_div);
codegen_->AddSlowPath(slow_path);
@@ -3107,7 +3107,7 @@
}
void InstructionCodeGeneratorX86::VisitDivZeroCheck(HDivZeroCheck* instruction) {
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathX86(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathX86(instruction);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = instruction->GetLocations();
@@ -4049,7 +4049,7 @@
}
void InstructionCodeGeneratorX86::GenerateExplicitNullCheck(HNullCheck* instruction) {
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86(instruction);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = instruction->GetLocations();
@@ -4489,7 +4489,7 @@
LocationSummary* locations = instruction->GetLocations();
Location index_loc = locations->InAt(0);
Location length_loc = locations->InAt(1);
- SlowPathCodeX86* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) BoundsCheckSlowPathX86(instruction);
if (length_loc.IsConstant()) {
@@ -4856,7 +4856,7 @@
__ movl(out, Address(out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())));
// TODO: We will need a read barrier here.
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck());
codegen_->AddSlowPath(slow_path);
__ testl(out, out);
@@ -4880,7 +4880,7 @@
void InstructionCodeGeneratorX86::VisitClinitCheck(HClinitCheck* check) {
// We assume the class to not be null.
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
check->GetLoadClass(), check, check->GetDexPc(), true);
codegen_->AddSlowPath(slow_path);
GenerateClassInitializationCheck(slow_path,
@@ -4888,7 +4888,7 @@
}
void InstructionCodeGeneratorX86::GenerateClassInitializationCheck(
- SlowPathCodeX86* slow_path, Register class_reg) {
+ SlowPathCode* slow_path, Register class_reg) {
__ cmpl(Address(class_reg, mirror::Class::StatusOffset().Int32Value()),
Immediate(mirror::Class::kStatusInitialized));
__ j(kLess, slow_path->GetEntryLabel());
@@ -4904,7 +4904,7 @@
}
void InstructionCodeGeneratorX86::VisitLoadString(HLoadString* load) {
- SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86(load);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86(load);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = load->GetLocations();
@@ -4973,7 +4973,7 @@
Register out = locations->Out().AsRegister<Register>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
NearLabel done, zero;
- SlowPathCodeX86* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
// Return 0 if `obj` is null.
// Avoid null check if we know obj is not null.
@@ -5032,8 +5032,7 @@
Location cls = locations->InAt(1);
Register temp = locations->GetTemp(0).AsRegister<Register>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- SlowPathCodeX86* slow_path =
- new (GetGraph()->GetArena()) TypeCheckSlowPathX86(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86(instruction);
codegen_->AddSlowPath(slow_path);
// Avoid null check if we know obj is not null.
@@ -5467,6 +5466,35 @@
visitor.VisitInsertionOrder();
}
+// TODO: target as memory.
+void CodeGeneratorX86::MoveFromReturnRegister(Location target, Primitive::Type type) {
+ if (!target.IsValid()) {
+ DCHECK(type == Primitive::kPrimVoid);
+ return;
+ }
+
+ DCHECK_NE(type, Primitive::kPrimVoid);
+
+ Location return_loc = InvokeDexCallingConventionVisitorX86().GetReturnLocation(type);
+ if (target.Equals(return_loc)) {
+ return;
+ }
+
+ // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged
+ // with the else branch.
+ if (type == Primitive::kPrimLong) {
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc.ToLow(), target.ToLow(), Primitive::kPrimInt, nullptr);
+ parallel_move.AddMove(return_loc.ToHigh(), target.ToHigh(), Primitive::kPrimInt, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+ } else {
+ // Let the parallel move resolver take care of all of this.
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc, target, type, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+ }
+}
+
#undef __
} // namespace x86
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index f3307cf..f38e1ea 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -31,7 +31,6 @@
static constexpr size_t kX86WordSize = kX86PointerSize;
class CodeGeneratorX86;
-class SlowPathCodeX86;
static constexpr Register kParameterCoreRegisters[] = { ECX, EDX, EBX };
static constexpr RegisterPair kParameterCorePairRegisters[] = { ECX_EDX, EDX_EBX };
@@ -171,7 +170,7 @@
// is the block to branch to if the suspend check is not needed, and after
// the suspend call.
void GenerateSuspendCheck(HSuspendCheck* check, HBasicBlock* successor);
- void GenerateClassInitializationCheck(SlowPathCodeX86* slow_path, Register class_reg);
+ void GenerateClassInitializationCheck(SlowPathCode* slow_path, Register class_reg);
void HandleBitwiseOperation(HBinaryOperation* instruction);
void GenerateDivRemIntegral(HBinaryOperation* instruction);
void DivRemOneOrMinusOne(HBinaryOperation* instruction);
@@ -300,9 +299,11 @@
void Move64(Location destination, Location source);
// Generate a call to a static or direct method.
- void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
// Generate a call to a virtual method.
- void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp);
+ void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
+
+ void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
// Emit linker patches.
void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
@@ -384,20 +385,6 @@
DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86);
};
-class SlowPathCodeX86 : public SlowPathCode {
- public:
- SlowPathCodeX86() : entry_label_(), exit_label_() {}
-
- Label* GetEntryLabel() { return &entry_label_; }
- Label* GetExitLabel() { return &exit_label_; }
-
- private:
- Label entry_label_;
- Label exit_label_;
-
- DISALLOW_COPY_AND_ASSIGN(SlowPathCodeX86);
-};
-
} // namespace x86
} // namespace art
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 134bfed..289ef64 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -50,7 +50,7 @@
#define __ down_cast<X86_64Assembler*>(codegen->GetAssembler())->
#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kX86_64WordSize, x).Int32Value()
-class NullCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
+class NullCheckSlowPathX86_64 : public SlowPathCode {
public:
explicit NullCheckSlowPathX86_64(HNullCheck* instruction) : instruction_(instruction) {}
@@ -76,7 +76,7 @@
DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathX86_64);
};
-class DivZeroCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
+class DivZeroCheckSlowPathX86_64 : public SlowPathCode {
public:
explicit DivZeroCheckSlowPathX86_64(HDivZeroCheck* instruction) : instruction_(instruction) {}
@@ -102,7 +102,7 @@
DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathX86_64);
};
-class DivRemMinusOneSlowPathX86_64 : public SlowPathCodeX86_64 {
+class DivRemMinusOneSlowPathX86_64 : public SlowPathCode {
public:
DivRemMinusOneSlowPathX86_64(Register reg, Primitive::Type type, bool is_div)
: cpu_reg_(CpuRegister(reg)), type_(type), is_div_(is_div) {}
@@ -136,7 +136,7 @@
DISALLOW_COPY_AND_ASSIGN(DivRemMinusOneSlowPathX86_64);
};
-class SuspendCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
+class SuspendCheckSlowPathX86_64 : public SlowPathCode {
public:
SuspendCheckSlowPathX86_64(HSuspendCheck* instruction, HBasicBlock* successor)
: instruction_(instruction), successor_(successor) {}
@@ -176,7 +176,7 @@
DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathX86_64);
};
-class BoundsCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
+class BoundsCheckSlowPathX86_64 : public SlowPathCode {
public:
explicit BoundsCheckSlowPathX86_64(HBoundsCheck* instruction)
: instruction_(instruction) {}
@@ -213,7 +213,7 @@
DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathX86_64);
};
-class LoadClassSlowPathX86_64 : public SlowPathCodeX86_64 {
+class LoadClassSlowPathX86_64 : public SlowPathCode {
public:
LoadClassSlowPathX86_64(HLoadClass* cls,
HInstruction* at,
@@ -266,7 +266,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathX86_64);
};
-class LoadStringSlowPathX86_64 : public SlowPathCodeX86_64 {
+class LoadStringSlowPathX86_64 : public SlowPathCode {
public:
explicit LoadStringSlowPathX86_64(HLoadString* instruction) : instruction_(instruction) {}
@@ -298,7 +298,7 @@
DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathX86_64);
};
-class TypeCheckSlowPathX86_64 : public SlowPathCodeX86_64 {
+class TypeCheckSlowPathX86_64 : public SlowPathCode {
public:
explicit TypeCheckSlowPathX86_64(HInstruction* instruction)
: instruction_(instruction) {}
@@ -355,7 +355,7 @@
DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathX86_64);
};
-class DeoptimizationSlowPathX86_64 : public SlowPathCodeX86_64 {
+class DeoptimizationSlowPathX86_64 : public SlowPathCode {
public:
explicit DeoptimizationSlowPathX86_64(HInstruction* instruction)
: instruction_(instruction) {}
@@ -1181,7 +1181,7 @@
}
void InstructionCodeGeneratorX86_64::VisitDeoptimize(HDeoptimize* deoptimize) {
- SlowPathCodeX86_64* slow_path = new (GetGraph()->GetArena())
+ SlowPathCode* slow_path = new (GetGraph()->GetArena())
DeoptimizationSlowPathX86_64(deoptimize);
codegen_->AddSlowPath(slow_path);
Label* slow_path_entry = slow_path->GetEntryLabel();
@@ -3072,7 +3072,7 @@
GenerateDivRemWithAnyConstant(instruction);
}
} else {
- SlowPathCodeX86_64* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) DivRemMinusOneSlowPathX86_64(
out.AsRegister(), type, is_div);
codegen_->AddSlowPath(slow_path);
@@ -3245,7 +3245,7 @@
}
void InstructionCodeGeneratorX86_64::VisitDivZeroCheck(HDivZeroCheck* instruction) {
- SlowPathCodeX86_64* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) DivZeroCheckSlowPathX86_64(instruction);
codegen_->AddSlowPath(slow_path);
@@ -3813,7 +3813,7 @@
}
void InstructionCodeGeneratorX86_64::GenerateExplicitNullCheck(HNullCheck* instruction) {
- SlowPathCodeX86_64* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86_64(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86_64(instruction);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = instruction->GetLocations();
@@ -4232,7 +4232,7 @@
LocationSummary* locations = instruction->GetLocations();
Location index_loc = locations->InAt(0);
Location length_loc = locations->InAt(1);
- SlowPathCodeX86_64* slow_path =
+ SlowPathCode* slow_path =
new (GetGraph()->GetArena()) BoundsCheckSlowPathX86_64(instruction);
if (length_loc.IsConstant()) {
@@ -4560,7 +4560,7 @@
}
void InstructionCodeGeneratorX86_64::GenerateClassInitializationCheck(
- SlowPathCodeX86_64* slow_path, CpuRegister class_reg) {
+ SlowPathCode* slow_path, CpuRegister class_reg) {
__ cmpl(Address(class_reg, mirror::Class::StatusOffset().Int32Value()),
Immediate(mirror::Class::kStatusInitialized));
__ j(kLess, slow_path->GetEntryLabel());
@@ -4593,7 +4593,7 @@
__ movl(out, Address(out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())));
// TODO: We will need a read barrier here.
- SlowPathCodeX86_64* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86_64(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86_64(
cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck());
codegen_->AddSlowPath(slow_path);
__ testl(out, out);
@@ -4617,7 +4617,7 @@
void InstructionCodeGeneratorX86_64::VisitClinitCheck(HClinitCheck* check) {
// We assume the class to not be null.
- SlowPathCodeX86_64* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86_64(
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86_64(
check->GetLoadClass(), check, check->GetDexPc(), true);
codegen_->AddSlowPath(slow_path);
GenerateClassInitializationCheck(slow_path,
@@ -4632,7 +4632,7 @@
}
void InstructionCodeGeneratorX86_64::VisitLoadString(HLoadString* load) {
- SlowPathCodeX86_64* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86_64(load);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86_64(load);
codegen_->AddSlowPath(slow_path);
LocationSummary* locations = load->GetLocations();
@@ -4701,7 +4701,7 @@
CpuRegister out = locations->Out().AsRegister<CpuRegister>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
NearLabel done, zero;
- SlowPathCodeX86_64* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
// Return 0 if `obj` is null.
// Avoid null check if we know obj is not null.
@@ -4759,8 +4759,7 @@
Location cls = locations->InAt(1);
CpuRegister temp = locations->GetTemp(0).AsRegister<CpuRegister>();
uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
- SlowPathCodeX86_64* slow_path =
- new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(instruction);
+ SlowPathCode* slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86_64(instruction);
codegen_->AddSlowPath(slow_path);
// Avoid null check if we know obj is not null.
@@ -5021,6 +5020,26 @@
return Address::RIP(fixup);
}
+// TODO: trg as memory.
+void CodeGeneratorX86_64::MoveFromReturnRegister(Location trg, Primitive::Type type) {
+ if (!trg.IsValid()) {
+ DCHECK(type == Primitive::kPrimVoid);
+ return;
+ }
+
+ DCHECK_NE(type, Primitive::kPrimVoid);
+
+ Location return_loc = InvokeDexCallingConventionVisitorX86_64().GetReturnLocation(type);
+ if (trg.Equals(return_loc)) {
+ return;
+ }
+
+ // Let the parallel move resolver take care of all of this.
+ HParallelMove parallel_move(GetGraph()->GetArena());
+ parallel_move.AddMove(return_loc, trg, type, nullptr);
+ GetMoveResolver()->EmitNativeCode(¶llel_move);
+}
+
#undef __
} // namespace x86_64
diff --git a/compiler/optimizing/code_generator_x86_64.h b/compiler/optimizing/code_generator_x86_64.h
index 9b2423f..1ec3580 100644
--- a/compiler/optimizing/code_generator_x86_64.h
+++ b/compiler/optimizing/code_generator_x86_64.h
@@ -87,20 +87,6 @@
class CodeGeneratorX86_64;
-class SlowPathCodeX86_64 : public SlowPathCode {
- public:
- SlowPathCodeX86_64() : entry_label_(), exit_label_() {}
-
- Label* GetEntryLabel() { return &entry_label_; }
- Label* GetExitLabel() { return &exit_label_; }
-
- private:
- Label entry_label_;
- Label exit_label_;
-
- DISALLOW_COPY_AND_ASSIGN(SlowPathCodeX86_64);
-};
-
class ParallelMoveResolverX86_64 : public ParallelMoveResolverWithSwap {
public:
ParallelMoveResolverX86_64(ArenaAllocator* allocator, CodeGeneratorX86_64* codegen)
@@ -181,7 +167,7 @@
// is the block to branch to if the suspend check is not needed, and after
// the suspend call.
void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
- void GenerateClassInitializationCheck(SlowPathCodeX86_64* slow_path, CpuRegister class_reg);
+ void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg);
void HandleBitwiseOperation(HBinaryOperation* operation);
void GenerateRemFP(HRem* rem);
void DivRemOneOrMinusOne(HBinaryOperation* instruction);
@@ -311,8 +297,10 @@
return false;
}
- void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
- void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp);
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
+ void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
+
+ void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
diff --git a/compiler/optimizing/intrinsics_arm.cc b/compiler/optimizing/intrinsics_arm.cc
index cc8ddb6..938c78e 100644
--- a/compiler/optimizing/intrinsics_arm.cc
+++ b/compiler/optimizing/intrinsics_arm.cc
@@ -21,6 +21,7 @@
#include "code_generator_arm.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "intrinsics.h"
+#include "intrinsics_utils.h"
#include "mirror/array-inl.h"
#include "mirror/string.h"
#include "thread.h"
@@ -38,99 +39,7 @@
return codegen_->GetGraph()->GetArena();
}
-#define __ codegen->GetAssembler()->
-
-static void MoveFromReturnRegister(Location trg, Primitive::Type type, CodeGeneratorARM* codegen) {
- if (!trg.IsValid()) {
- DCHECK(type == Primitive::kPrimVoid);
- return;
- }
-
- DCHECK_NE(type, Primitive::kPrimVoid);
-
- if (Primitive::IsIntegralType(type) || type == Primitive::kPrimNot) {
- if (type == Primitive::kPrimLong) {
- Register trg_reg_lo = trg.AsRegisterPairLow<Register>();
- Register trg_reg_hi = trg.AsRegisterPairHigh<Register>();
- Register res_reg_lo = R0;
- Register res_reg_hi = R1;
- if (trg_reg_lo != res_reg_hi) {
- if (trg_reg_lo != res_reg_lo) {
- __ mov(trg_reg_lo, ShifterOperand(res_reg_lo));
- __ mov(trg_reg_hi, ShifterOperand(res_reg_hi));
- } else {
- DCHECK_EQ(trg_reg_lo + 1, trg_reg_hi);
- }
- } else {
- __ mov(trg_reg_hi, ShifterOperand(res_reg_hi));
- __ mov(trg_reg_lo, ShifterOperand(res_reg_lo));
- }
- } else {
- Register trg_reg = trg.AsRegister<Register>();
- Register res_reg = R0;
- if (trg_reg != res_reg) {
- __ mov(trg_reg, ShifterOperand(res_reg));
- }
- }
- } else {
- UNIMPLEMENTED(FATAL) << "Floating-point return.";
- }
-}
-
-static void MoveArguments(HInvoke* invoke, CodeGeneratorARM* codegen) {
- InvokeDexCallingConventionVisitorARM calling_convention_visitor;
- IntrinsicVisitor::MoveArguments(invoke, codegen, &calling_convention_visitor);
-}
-
-// Slow-path for fallback (calling the managed code to handle the intrinsic) in an intrinsified
-// call. This will copy the arguments into the positions for a regular call.
-//
-// Note: The actual parameters are required to be in the locations given by the invoke's location
-// summary. If an intrinsic modifies those locations before a slowpath call, they must be
-// restored!
-class IntrinsicSlowPathARM : public SlowPathCodeARM {
- public:
- explicit IntrinsicSlowPathARM(HInvoke* invoke) : invoke_(invoke) { }
-
- void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
- CodeGeneratorARM* codegen = down_cast<CodeGeneratorARM*>(codegen_in);
- __ Bind(GetEntryLabel());
-
- SaveLiveRegisters(codegen, invoke_->GetLocations());
-
- MoveArguments(invoke_, codegen);
-
- if (invoke_->IsInvokeStaticOrDirect()) {
- codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(),
- Location::RegisterLocation(kArtMethodRegister));
- } else {
- codegen->GenerateVirtualCall(invoke_->AsInvokeVirtual(),
- Location::RegisterLocation(kArtMethodRegister));
- }
- codegen->RecordPcInfo(invoke_, invoke_->GetDexPc(), this);
-
- // Copy the result back to the expected output.
- Location out = invoke_->GetLocations()->Out();
- if (out.IsValid()) {
- DCHECK(out.IsRegister()); // TODO: Replace this when we support output in memory.
- DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
- MoveFromReturnRegister(out, invoke_->GetType(), codegen);
- }
-
- RestoreLiveRegisters(codegen, invoke_->GetLocations());
- __ b(GetExitLabel());
- }
-
- const char* GetDescription() const OVERRIDE { return "IntrinsicSlowPathARM"; }
-
- private:
- // The instruction where this slow path is happening.
- HInvoke* const invoke_;
-
- DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathARM);
-};
-
-#undef __
+using IntrinsicSlowPathARM = IntrinsicSlowPath<InvokeDexCallingConventionVisitorARM>;
bool IntrinsicLocationsBuilderARM::TryDispatch(HInvoke* invoke) {
Dispatch(invoke);
@@ -1094,7 +1003,7 @@
// TODO: For simplicity, the index parameter is requested in a register, so different from Quick
// we will not optimize the code for constants (which would save a register).
- SlowPathCodeARM* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
codegen_->AddSlowPath(slow_path);
__ ldr(temp, Address(obj, count_offset.Int32Value())); // temp = str.length.
@@ -1130,7 +1039,7 @@
Register argument = locations->InAt(1).AsRegister<Register>();
__ cmp(argument, ShifterOperand(0));
- SlowPathCodeARM* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
codegen_->AddSlowPath(slow_path);
__ b(slow_path->GetEntryLabel(), EQ);
@@ -1248,7 +1157,7 @@
// Check for code points > 0xFFFF. Either a slow-path check when we don't know statically,
// or directly dispatch if we have a constant.
- SlowPathCodeARM* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
if (invoke->InputAt(1)->IsIntConstant()) {
if (static_cast<uint32_t>(invoke->InputAt(1)->AsIntConstant()->GetValue()) >
std::numeric_limits<uint16_t>::max()) {
@@ -1341,7 +1250,7 @@
Register byte_array = locations->InAt(0).AsRegister<Register>();
__ cmp(byte_array, ShifterOperand(0));
- SlowPathCodeARM* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
codegen_->AddSlowPath(slow_path);
__ b(slow_path->GetEntryLabel(), EQ);
@@ -1387,7 +1296,7 @@
Register string_to_copy = locations->InAt(0).AsRegister<Register>();
__ cmp(string_to_copy, ShifterOperand(0));
- SlowPathCodeARM* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
codegen_->AddSlowPath(slow_path);
__ b(slow_path->GetEntryLabel(), EQ);
diff --git a/compiler/optimizing/intrinsics_utils.h b/compiler/optimizing/intrinsics_utils.h
new file mode 100644
index 0000000..e70afd2
--- /dev/null
+++ b/compiler/optimizing/intrinsics_utils.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_INTRINSICS_UTILS_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_UTILS_H_
+
+#include "base/macros.h"
+#include "code_generator.h"
+#include "locations.h"
+#include "nodes.h"
+#include "utils/assembler.h"
+#include "utils/label.h"
+
+namespace art {
+
+// Default slow-path for fallback (calling the managed code to handle the intrinsic) in an
+// intrinsified call. This will copy the arguments into the positions for a regular call.
+//
+// Note: The actual parameters are required to be in the locations given by the invoke's location
+// summary. If an intrinsic modifies those locations before a slowpath call, they must be
+// restored!
+//
+// Note: If an invoke wasn't sharpened, we will put down an invoke-virtual here. That's potentially
+// sub-optimal (compared to a direct pointer call), but this is a slow-path.
+
+template <typename TDexCallingConvention>
+class IntrinsicSlowPath : public SlowPathCode {
+ public:
+ explicit IntrinsicSlowPath(HInvoke* invoke) : invoke_(invoke) { }
+
+ Location MoveArguments(CodeGenerator* codegen) {
+ TDexCallingConvention calling_convention_visitor;
+ IntrinsicVisitor::MoveArguments(invoke_, codegen, &calling_convention_visitor);
+ return calling_convention_visitor.GetMethodLocation();
+ }
+
+ void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
+ Assembler* assembler = codegen->GetAssembler();
+ assembler->Bind(GetEntryLabel());
+
+ SaveLiveRegisters(codegen, invoke_->GetLocations());
+
+ Location method_loc = MoveArguments(codegen);
+
+ if (invoke_->IsInvokeStaticOrDirect()) {
+ codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(), method_loc);
+ } else {
+ codegen->GenerateVirtualCall(invoke_->AsInvokeVirtual(), method_loc);
+ }
+ codegen->RecordPcInfo(invoke_, invoke_->GetDexPc(), this);
+
+ // Copy the result back to the expected output.
+ Location out = invoke_->GetLocations()->Out();
+ if (out.IsValid()) {
+ DCHECK(out.IsRegister()); // TODO: Replace this when we support output in memory.
+ DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
+ codegen->MoveFromReturnRegister(out, invoke_->GetType());
+ }
+
+ RestoreLiveRegisters(codegen, invoke_->GetLocations());
+ assembler->Jump(GetExitLabel());
+ }
+
+ const char* GetDescription() const OVERRIDE { return "IntrinsicSlowPath"; }
+
+ private:
+ // The instruction where this slow path is happening.
+ HInvoke* const invoke_;
+
+ DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPath);
+};
+
+} // namespace art
+
+#endif // ART_COMPILER_OPTIMIZING_INTRINSICS_UTILS_H_
diff --git a/compiler/optimizing/intrinsics_x86.cc b/compiler/optimizing/intrinsics_x86.cc
index d0bc55c..5becf0f 100644
--- a/compiler/optimizing/intrinsics_x86.cc
+++ b/compiler/optimizing/intrinsics_x86.cc
@@ -24,6 +24,7 @@
#include "code_generator_x86.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "intrinsics.h"
+#include "intrinsics_utils.h"
#include "mirror/array-inl.h"
#include "mirror/string.h"
#include "thread.h"
@@ -57,117 +58,13 @@
return res != nullptr && res->Intrinsified();
}
-#define __ reinterpret_cast<X86Assembler*>(codegen->GetAssembler())->
-
-// TODO: target as memory.
-static void MoveFromReturnRegister(Location target,
- Primitive::Type type,
- CodeGeneratorX86* codegen) {
- if (!target.IsValid()) {
- DCHECK(type == Primitive::kPrimVoid);
- return;
- }
-
- switch (type) {
- case Primitive::kPrimBoolean:
- case Primitive::kPrimByte:
- case Primitive::kPrimChar:
- case Primitive::kPrimShort:
- case Primitive::kPrimInt:
- case Primitive::kPrimNot: {
- Register target_reg = target.AsRegister<Register>();
- if (target_reg != EAX) {
- __ movl(target_reg, EAX);
- }
- break;
- }
- case Primitive::kPrimLong: {
- Register target_reg_lo = target.AsRegisterPairLow<Register>();
- Register target_reg_hi = target.AsRegisterPairHigh<Register>();
- if (target_reg_lo != EAX) {
- __ movl(target_reg_lo, EAX);
- }
- if (target_reg_hi != EDX) {
- __ movl(target_reg_hi, EDX);
- }
- break;
- }
-
- case Primitive::kPrimVoid:
- LOG(FATAL) << "Unexpected void type for valid location " << target;
- UNREACHABLE();
-
- case Primitive::kPrimDouble: {
- XmmRegister target_reg = target.AsFpuRegister<XmmRegister>();
- if (target_reg != XMM0) {
- __ movsd(target_reg, XMM0);
- }
- break;
- }
- case Primitive::kPrimFloat: {
- XmmRegister target_reg = target.AsFpuRegister<XmmRegister>();
- if (target_reg != XMM0) {
- __ movss(target_reg, XMM0);
- }
- break;
- }
- }
-}
-
static void MoveArguments(HInvoke* invoke, CodeGeneratorX86* codegen) {
InvokeDexCallingConventionVisitorX86 calling_convention_visitor;
IntrinsicVisitor::MoveArguments(invoke, codegen, &calling_convention_visitor);
}
-// Slow-path for fallback (calling the managed code to handle the intrinsic) in an intrinsified
-// call. This will copy the arguments into the positions for a regular call.
-//
-// Note: The actual parameters are required to be in the locations given by the invoke's location
-// summary. If an intrinsic modifies those locations before a slowpath call, they must be
-// restored!
-class IntrinsicSlowPathX86 : public SlowPathCodeX86 {
- public:
- explicit IntrinsicSlowPathX86(HInvoke* invoke)
- : invoke_(invoke) { }
+using IntrinsicSlowPathX86 = IntrinsicSlowPath<InvokeDexCallingConventionVisitorX86>;
- void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
- CodeGeneratorX86* codegen = down_cast<CodeGeneratorX86*>(codegen_in);
- __ Bind(GetEntryLabel());
-
- SaveLiveRegisters(codegen, invoke_->GetLocations());
-
- MoveArguments(invoke_, codegen);
-
- if (invoke_->IsInvokeStaticOrDirect()) {
- codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(),
- Location::RegisterLocation(EAX));
- } else {
- codegen->GenerateVirtualCall(invoke_->AsInvokeVirtual(), Location::RegisterLocation(EAX));
- }
- codegen->RecordPcInfo(invoke_, invoke_->GetDexPc(), this);
-
- // Copy the result back to the expected output.
- Location out = invoke_->GetLocations()->Out();
- if (out.IsValid()) {
- DCHECK(out.IsRegister()); // TODO: Replace this when we support output in memory.
- DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
- MoveFromReturnRegister(out, invoke_->GetType(), codegen);
- }
-
- RestoreLiveRegisters(codegen, invoke_->GetLocations());
- __ jmp(GetExitLabel());
- }
-
- const char* GetDescription() const OVERRIDE { return "IntrinsicSlowPathX86"; }
-
- private:
- // The instruction where this slow path is happening.
- HInvoke* const invoke_;
-
- DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathX86);
-};
-
-#undef __
#define __ assembler->
static void CreateFPToIntLocations(ArenaAllocator* arena, HInvoke* invoke, bool is64bit) {
@@ -743,7 +640,7 @@
Location out = invoke->GetLocations()->Out();
if (out.IsValid()) {
DCHECK(out.IsRegister());
- MoveFromReturnRegister(out, invoke->GetType(), codegen);
+ codegen->MoveFromReturnRegister(out, invoke->GetType());
}
}
@@ -902,7 +799,7 @@
// TODO: For simplicity, the index parameter is requested in a register, so different from Quick
// we will not optimize the code for constants (which would save a register).
- SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
codegen_->AddSlowPath(slow_path);
X86Assembler* assembler = GetAssembler();
@@ -971,7 +868,7 @@
Location pos,
Register input,
Register length,
- SlowPathCodeX86* slow_path,
+ SlowPathCode* slow_path,
Register input_len,
Register temp) {
// Where is the length in the String?
@@ -1030,7 +927,7 @@
Register count = locations->GetTemp(2).AsRegister<Register>();
DCHECK_EQ(count, ECX);
- SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
codegen_->AddSlowPath(slow_path);
// Bail out if the source and destination are the same (to handle overlap).
@@ -1114,7 +1011,7 @@
Register argument = locations->InAt(1).AsRegister<Register>();
__ testl(argument, argument);
- SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());
@@ -1259,7 +1156,7 @@
// Check for code points > 0xFFFF. Either a slow-path check when we don't know statically,
// or directly dispatch if we have a constant.
- SlowPathCodeX86* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
if (invoke->InputAt(1)->IsIntConstant()) {
if (static_cast<uint32_t>(invoke->InputAt(1)->AsIntConstant()->GetValue()) >
std::numeric_limits<uint16_t>::max()) {
@@ -1380,7 +1277,7 @@
Register byte_array = locations->InAt(0).AsRegister<Register>();
__ testl(byte_array, byte_array);
- SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());
@@ -1422,7 +1319,7 @@
Register string_to_copy = locations->InAt(0).AsRegister<Register>();
__ testl(string_to_copy, string_to_copy);
- SlowPathCodeX86* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());
diff --git a/compiler/optimizing/intrinsics_x86_64.cc b/compiler/optimizing/intrinsics_x86_64.cc
index 315cbab..2bd86a1 100644
--- a/compiler/optimizing/intrinsics_x86_64.cc
+++ b/compiler/optimizing/intrinsics_x86_64.cc
@@ -24,6 +24,7 @@
#include "code_generator_x86_64.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "intrinsics.h"
+#include "intrinsics_utils.h"
#include "mirror/array-inl.h"
#include "mirror/string.h"
#include "thread.h"
@@ -53,112 +54,13 @@
return res != nullptr && res->Intrinsified();
}
-#define __ reinterpret_cast<X86_64Assembler*>(codegen->GetAssembler())->
-
-// TODO: trg as memory.
-static void MoveFromReturnRegister(Location trg,
- Primitive::Type type,
- CodeGeneratorX86_64* codegen) {
- if (!trg.IsValid()) {
- DCHECK(type == Primitive::kPrimVoid);
- return;
- }
-
- switch (type) {
- case Primitive::kPrimBoolean:
- case Primitive::kPrimByte:
- case Primitive::kPrimChar:
- case Primitive::kPrimShort:
- case Primitive::kPrimInt:
- case Primitive::kPrimNot: {
- CpuRegister trg_reg = trg.AsRegister<CpuRegister>();
- if (trg_reg.AsRegister() != RAX) {
- __ movl(trg_reg, CpuRegister(RAX));
- }
- break;
- }
- case Primitive::kPrimLong: {
- CpuRegister trg_reg = trg.AsRegister<CpuRegister>();
- if (trg_reg.AsRegister() != RAX) {
- __ movq(trg_reg, CpuRegister(RAX));
- }
- break;
- }
-
- case Primitive::kPrimVoid:
- LOG(FATAL) << "Unexpected void type for valid location " << trg;
- UNREACHABLE();
-
- case Primitive::kPrimDouble: {
- XmmRegister trg_reg = trg.AsFpuRegister<XmmRegister>();
- if (trg_reg.AsFloatRegister() != XMM0) {
- __ movsd(trg_reg, XmmRegister(XMM0));
- }
- break;
- }
- case Primitive::kPrimFloat: {
- XmmRegister trg_reg = trg.AsFpuRegister<XmmRegister>();
- if (trg_reg.AsFloatRegister() != XMM0) {
- __ movss(trg_reg, XmmRegister(XMM0));
- }
- break;
- }
- }
-}
-
static void MoveArguments(HInvoke* invoke, CodeGeneratorX86_64* codegen) {
InvokeDexCallingConventionVisitorX86_64 calling_convention_visitor;
IntrinsicVisitor::MoveArguments(invoke, codegen, &calling_convention_visitor);
}
-// Slow-path for fallback (calling the managed code to handle the intrinsic) in an intrinsified
-// call. This will copy the arguments into the positions for a regular call.
-//
-// Note: The actual parameters are required to be in the locations given by the invoke's location
-// summary. If an intrinsic modifies those locations before a slowpath call, they must be
-// restored!
-class IntrinsicSlowPathX86_64 : public SlowPathCodeX86_64 {
- public:
- explicit IntrinsicSlowPathX86_64(HInvoke* invoke) : invoke_(invoke) { }
+using IntrinsicSlowPathX86_64 = IntrinsicSlowPath<InvokeDexCallingConventionVisitorX86_64>;
- void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
- CodeGeneratorX86_64* codegen = down_cast<CodeGeneratorX86_64*>(codegen_in);
- __ Bind(GetEntryLabel());
-
- SaveLiveRegisters(codegen, invoke_->GetLocations());
-
- MoveArguments(invoke_, codegen);
-
- if (invoke_->IsInvokeStaticOrDirect()) {
- codegen->GenerateStaticOrDirectCall(
- invoke_->AsInvokeStaticOrDirect(), Location::RegisterLocation(RDI));
- } else {
- codegen->GenerateVirtualCall(invoke_->AsInvokeVirtual(), Location::RegisterLocation(RDI));
- }
- codegen->RecordPcInfo(invoke_, invoke_->GetDexPc(), this);
-
- // Copy the result back to the expected output.
- Location out = invoke_->GetLocations()->Out();
- if (out.IsValid()) {
- DCHECK(out.IsRegister()); // TODO: Replace this when we support output in memory.
- DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
- MoveFromReturnRegister(out, invoke_->GetType(), codegen);
- }
-
- RestoreLiveRegisters(codegen, invoke_->GetLocations());
- __ jmp(GetExitLabel());
- }
-
- const char* GetDescription() const OVERRIDE { return "IntrinsicSlowPathX86_64"; }
-
- private:
- // The instruction where this slow path is happening.
- HInvoke* const invoke_;
-
- DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathX86_64);
-};
-
-#undef __
#define __ assembler->
static void CreateFPToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
@@ -602,7 +504,7 @@
Location out = invoke->GetLocations()->Out();
if (out.IsValid()) {
DCHECK(out.IsRegister());
- MoveFromReturnRegister(out, invoke->GetType(), codegen);
+ codegen->MoveFromReturnRegister(out, invoke->GetType());
}
}
@@ -811,7 +713,7 @@
// TODO: For simplicity, the index parameter is requested in a register, so different from Quick
// we will not optimize the code for constants (which would save a register).
- SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
codegen_->AddSlowPath(slow_path);
X86_64Assembler* assembler = GetAssembler();
@@ -869,7 +771,7 @@
Location pos,
CpuRegister input,
CpuRegister length,
- SlowPathCodeX86_64* slow_path,
+ SlowPathCode* slow_path,
CpuRegister input_len,
CpuRegister temp) {
// Where is the length in the String?
@@ -928,7 +830,7 @@
CpuRegister count = locations->GetTemp(2).AsRegister<CpuRegister>();
DCHECK_EQ(count.AsRegister(), RCX);
- SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
codegen_->AddSlowPath(slow_path);
// Bail out if the source and destination are the same.
@@ -1010,7 +912,7 @@
CpuRegister argument = locations->InAt(1).AsRegister<CpuRegister>();
__ testl(argument, argument);
- SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());
@@ -1156,7 +1058,7 @@
// Check for code points > 0xFFFF. Either a slow-path check when we don't know statically,
// or directly dispatch if we have a constant.
- SlowPathCodeX86_64* slow_path = nullptr;
+ SlowPathCode* slow_path = nullptr;
if (invoke->InputAt(1)->IsIntConstant()) {
if (static_cast<uint32_t>(invoke->InputAt(1)->AsIntConstant()->GetValue()) >
std::numeric_limits<uint16_t>::max()) {
@@ -1276,7 +1178,7 @@
CpuRegister byte_array = locations->InAt(0).AsRegister<CpuRegister>();
__ testl(byte_array, byte_array);
- SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());
@@ -1320,7 +1222,7 @@
CpuRegister string_to_copy = locations->InAt(0).AsRegister<CpuRegister>();
__ testl(string_to_copy, string_to_copy);
- SlowPathCodeX86_64* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
+ SlowPathCode* slow_path = new (GetAllocator()) IntrinsicSlowPathX86_64(invoke);
codegen_->AddSlowPath(slow_path);
__ j(kEqual, slow_path->GetEntryLabel());