Change pResolveString entrypoint to kSaveEverything.
Test: Run ART test suite including gcstress on host and Nexus 9.
Test: Run ART test suite including gcstress with baker CC on host and Nexus 9.
Bug: 20323084
Change-Id: I63c21a7d3be8ff7a5765b5003c85b5317635efe6
diff --git a/compiler/linker/arm64/relative_patcher_arm64.cc b/compiler/linker/arm64/relative_patcher_arm64.cc
index 3b77880..4a9de7f 100644
--- a/compiler/linker/arm64/relative_patcher_arm64.cc
+++ b/compiler/linker/arm64/relative_patcher_arm64.cc
@@ -214,7 +214,7 @@
DCHECK(patch.GetType() == LinkerPatch::Type::kStringRelative ||
patch.GetType() == LinkerPatch::Type::kTypeRelative) << patch.GetType();
} else {
- // With the read barrier (non-baker) enabled, it could be kDexCacheArray in the
+ // With the read barrier (non-Baker) enabled, it could be kDexCacheArray in the
// HLoadString::LoadKind::kDexCachePcRelative case of VisitLoadString().
DCHECK(patch.GetType() == LinkerPatch::Type::kStringRelative ||
patch.GetType() == LinkerPatch::Type::kTypeRelative ||
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 9870876..80b4907 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -429,34 +429,50 @@
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
+ HLoadString* load = instruction_->AsLoadString();
+ const uint32_t string_index = load->GetStringIndex();
+ Register out = locations->Out().AsRegister<Register>();
+ Register temp = locations->GetTemp(0).AsRegister<Register>();
+ constexpr bool call_saves_everything_except_r0 = (!kUseReadBarrier || kUseBakerReadBarrier);
CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen);
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
InvokeRuntimeCallingConvention calling_convention;
- HLoadString* load = instruction_->AsLoadString();
- const uint32_t string_index = load->GetStringIndex();
+ // In the unlucky case that the `temp` is R0, we preserve the address in `out` across
+ // the kSaveEverything call (or use `out` for the address after non-kSaveEverything call).
+ bool temp_is_r0 = (temp == calling_convention.GetRegisterAt(0));
+ Register entry_address = temp_is_r0 ? out : temp;
+ DCHECK_NE(entry_address, calling_convention.GetRegisterAt(0));
+ if (call_saves_everything_except_r0 && temp_is_r0) {
+ __ mov(entry_address, ShifterOperand(temp));
+ }
+
__ LoadImmediate(calling_convention.GetRegisterAt(0), string_index);
arm_codegen->InvokeRuntime(kQuickResolveString, instruction_, instruction_->GetDexPc(), this);
CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
+
+ // Store the resolved String to the .bss entry.
+ if (call_saves_everything_except_r0) {
+ // The string entry address was preserved in `entry_address` thanks to kSaveEverything.
+ __ str(R0, Address(entry_address));
+ } else {
+ // For non-Baker read barrier, we need to re-calculate the address of the string entry.
+ CodeGeneratorARM::PcRelativePatchInfo* labels =
+ arm_codegen->NewPcRelativeStringPatch(load->GetDexFile(), string_index);
+ __ BindTrackedLabel(&labels->movw_label);
+ __ movw(entry_address, /* placeholder */ 0u);
+ __ BindTrackedLabel(&labels->movt_label);
+ __ movt(entry_address, /* placeholder */ 0u);
+ __ BindTrackedLabel(&labels->add_pc_label);
+ __ add(entry_address, entry_address, ShifterOperand(PC));
+ __ str(R0, Address(entry_address));
+ }
+
arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0));
-
RestoreLiveRegisters(codegen, locations);
- // Store the resolved String to the BSS entry.
- // TODO: Change art_quick_resolve_string to kSaveEverything and use a temporary for the
- // .bss entry address in the fast path, so that we can avoid another calculation here.
- CodeGeneratorARM::PcRelativePatchInfo* labels =
- arm_codegen->NewPcRelativeStringPatch(load->GetDexFile(), string_index);
- __ BindTrackedLabel(&labels->movw_label);
- __ movw(IP, /* placeholder */ 0u);
- __ BindTrackedLabel(&labels->movt_label);
- __ movt(IP, /* placeholder */ 0u);
- __ BindTrackedLabel(&labels->add_pc_label);
- __ add(IP, IP, ShifterOperand(PC));
- __ str(locations->Out().AsRegister<Register>(), Address(IP));
-
__ b(GetExitLabel());
}
@@ -5704,10 +5720,25 @@
HLoadString::LoadKind load_kind = load->GetLoadKind();
if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod) {
- locations->SetInAt(0, Location::RequiresRegister());
locations->SetOut(Location::RegisterLocation(R0));
} else {
locations->SetOut(Location::RequiresRegister());
+ if (load_kind == HLoadString::LoadKind::kBssEntry) {
+ if (!kUseReadBarrier || kUseBakerReadBarrier) {
+ // Rely on the pResolveString and/or marking to save everything, including temps.
+ // Note that IP may theoretically be clobbered by saving/restoring the live register
+ // (only one thanks to the custom calling convention), so we request a different temp.
+ locations->AddTemp(Location::RequiresRegister());
+ RegisterSet caller_saves = RegisterSet::Empty();
+ InvokeRuntimeCallingConvention calling_convention;
+ caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
+ // TODO: Add GetReturnLocation() to the calling convention so that we can DCHECK()
+ // that the the kPrimNot result register is the same as the first argument register.
+ locations->SetCustomSlowPathCallerSaves(caller_saves);
+ } else {
+ // For non-Baker read barrier we have a temp-clobbering call.
+ }
+ }
}
}
@@ -5743,15 +5774,16 @@
}
case HLoadString::LoadKind::kBssEntry: {
DCHECK(!codegen_->GetCompilerOptions().IsBootImage());
+ Register temp = locations->GetTemp(0).AsRegister<Register>();
CodeGeneratorARM::PcRelativePatchInfo* labels =
codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex());
__ BindTrackedLabel(&labels->movw_label);
- __ movw(out, /* placeholder */ 0u);
+ __ movw(temp, /* placeholder */ 0u);
__ BindTrackedLabel(&labels->movt_label);
- __ movt(out, /* placeholder */ 0u);
+ __ movt(temp, /* placeholder */ 0u);
__ BindTrackedLabel(&labels->add_pc_label);
- __ add(out, out, ShifterOperand(PC));
- GenerateGcRootFieldLoad(load, out_loc, out, 0);
+ __ add(temp, temp, ShifterOperand(PC));
+ GenerateGcRootFieldLoad(load, out_loc, temp, 0);
SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load);
codegen_->AddSlowPath(slow_path);
__ CompareAndBranchIfZero(out, slow_path->GetEntryLabel());
@@ -5765,6 +5797,7 @@
// TODO: Consider re-adding the compiler code to do string dex cache lookup again.
DCHECK(load_kind == HLoadString::LoadKind::kDexCacheViaMethod);
InvokeRuntimeCallingConvention calling_convention;
+ DCHECK_EQ(calling_convention.GetRegisterAt(0), out);
__ LoadImmediate(calling_convention.GetRegisterAt(0), load->GetStringIndex());
codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc());
CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 969d653..8197787 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -331,13 +331,20 @@
class LoadStringSlowPathARM64 : public SlowPathCodeARM64 {
public:
- explicit LoadStringSlowPathARM64(HLoadString* instruction) : SlowPathCodeARM64(instruction) {}
+ LoadStringSlowPathARM64(HLoadString* instruction, Register temp, vixl::aarch64::Label* adrp_label)
+ : SlowPathCodeARM64(instruction),
+ temp_(temp),
+ adrp_label_(adrp_label) {}
void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
LocationSummary* locations = instruction_->GetLocations();
DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen);
+ // temp_ is a scratch register. Make sure it's not used for saving/restoring registers.
+ UseScratchRegisterScope temps(arm64_codegen->GetVIXLAssembler());
+ temps.Exclude(temp_);
+
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
@@ -352,21 +359,21 @@
RestoreLiveRegisters(codegen, locations);
// Store the resolved String to the BSS entry.
- UseScratchRegisterScope temps(arm64_codegen->GetVIXLAssembler());
- Register temp = temps.AcquireX();
const DexFile& dex_file = instruction_->AsLoadString()->GetDexFile();
- // TODO: Change art_quick_resolve_string to kSaveEverything and use a temporary
- // for the ADRP in the fast path, so that we can avoid the ADRP here.
- vixl::aarch64::Label* adrp_label =
- arm64_codegen->NewPcRelativeStringPatch(dex_file, string_index);
- arm64_codegen->EmitAdrpPlaceholder(adrp_label, temp);
+ if (!kUseReadBarrier || kUseBakerReadBarrier) {
+ // The string entry page address was preserved in temp_ thanks to kSaveEverything.
+ } else {
+ // For non-Baker read barrier, we need to re-calculate the address of the string entry page.
+ adrp_label_ = arm64_codegen->NewPcRelativeStringPatch(dex_file, string_index);
+ arm64_codegen->EmitAdrpPlaceholder(adrp_label_, temp_);
+ }
vixl::aarch64::Label* strp_label =
- arm64_codegen->NewPcRelativeStringPatch(dex_file, string_index, adrp_label);
+ arm64_codegen->NewPcRelativeStringPatch(dex_file, string_index, adrp_label_);
{
SingleEmissionCheckScope guard(arm64_codegen->GetVIXLAssembler());
__ Bind(strp_label);
__ str(RegisterFrom(locations->Out(), Primitive::kPrimNot),
- MemOperand(temp, /* offset placeholder */ 0));
+ MemOperand(temp_, /* offset placeholder */ 0));
}
__ B(GetExitLabel());
@@ -375,6 +382,9 @@
const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM64"; }
private:
+ const Register temp_;
+ vixl::aarch64::Label* adrp_label_;
+
DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM64);
};
@@ -4238,11 +4248,24 @@
: LocationSummary::kNoCall;
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind);
if (load->GetLoadKind() == HLoadString::LoadKind::kDexCacheViaMethod) {
- locations->SetInAt(0, Location::RequiresRegister());
InvokeRuntimeCallingConvention calling_convention;
locations->SetOut(calling_convention.GetReturnLocation(load->GetType()));
} else {
locations->SetOut(Location::RequiresRegister());
+ if (load->GetLoadKind() == HLoadString::LoadKind::kBssEntry) {
+ if (!kUseReadBarrier || kUseBakerReadBarrier) {
+ // Rely on the pResolveString and/or marking to save everything, including temps.
+ RegisterSet caller_saves = RegisterSet::Empty();
+ InvokeRuntimeCallingConvention calling_convention;
+ caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0).GetCode()));
+ DCHECK_EQ(calling_convention.GetRegisterAt(0).GetCode(),
+ RegisterFrom(calling_convention.GetReturnLocation(Primitive::kPrimNot),
+ Primitive::kPrimNot).GetCode());
+ locations->SetCustomSlowPathCallerSaves(caller_saves);
+ } else {
+ // For non-Baker read barrier we have a temp-clobbering call.
+ }
+ }
}
}
@@ -4277,18 +4300,21 @@
const DexFile& dex_file = load->GetDexFile();
uint32_t string_index = load->GetStringIndex();
DCHECK(!codegen_->GetCompilerOptions().IsBootImage());
+ UseScratchRegisterScope temps(codegen_->GetVIXLAssembler());
+ Register temp = temps.AcquireX();
vixl::aarch64::Label* adrp_label = codegen_->NewPcRelativeStringPatch(dex_file, string_index);
- codegen_->EmitAdrpPlaceholder(adrp_label, out.X());
+ codegen_->EmitAdrpPlaceholder(adrp_label, temp);
// Add LDR with its PC-relative String patch.
vixl::aarch64::Label* ldr_label =
codegen_->NewPcRelativeStringPatch(dex_file, string_index, adrp_label);
// /* GcRoot<mirror::Class> */ out = *(base_address + offset) /* PC-relative */
GenerateGcRootFieldLoad(load,
load->GetLocations()->Out(),
- out.X(),
+ temp,
/* placeholder */ 0u,
ldr_label);
- SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM64(load);
+ SlowPathCodeARM64* slow_path =
+ new (GetGraph()->GetArena()) LoadStringSlowPathARM64(load, temp, adrp_label);
codegen_->AddSlowPath(slow_path);
__ Cbz(out.X(), slow_path->GetEntryLabel());
__ Bind(slow_path->GetExitLabel());
@@ -4300,6 +4326,7 @@
// TODO: Re-add the compiler code to do string dex cache lookup again.
InvokeRuntimeCallingConvention calling_convention;
+ DCHECK_EQ(calling_convention.GetRegisterAt(0).GetCode(), out.GetCode());
__ Mov(calling_convention.GetRegisterAt(0).W(), load->GetStringIndex());
codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc());
CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 0b23599..ab60671 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -6059,8 +6059,7 @@
: LocationSummary::kNoCall;
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind);
HLoadString::LoadKind load_kind = load->GetLoadKind();
- if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod ||
- load_kind == HLoadString::LoadKind::kBootImageLinkTimePcRelative ||
+ if (load_kind == HLoadString::LoadKind::kBootImageLinkTimePcRelative ||
load_kind == HLoadString::LoadKind::kBssEntry) {
locations->SetInAt(0, Location::RequiresRegister());
}
@@ -6068,6 +6067,17 @@
locations->SetOut(Location::RegisterLocation(EAX));
} else {
locations->SetOut(Location::RequiresRegister());
+ if (load_kind == HLoadString::LoadKind::kBssEntry) {
+ if (!kUseReadBarrier || kUseBakerReadBarrier) {
+ // Rely on the pResolveString and/or marking to save everything.
+ RegisterSet caller_saves = RegisterSet::Empty();
+ InvokeRuntimeCallingConvention calling_convention;
+ caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
+ locations->SetCustomSlowPathCallerSaves(caller_saves);
+ } else {
+ // For non-Baker read barrier we have a temp-clobbering call.
+ }
+ }
}
}
@@ -6114,6 +6124,7 @@
// TODO: Re-add the compiler code to do string dex cache lookup again.
InvokeRuntimeCallingConvention calling_convention;
+ DCHECK_EQ(calling_convention.GetRegisterAt(0), out);
__ movl(calling_convention.GetRegisterAt(0), Immediate(load->GetStringIndex()));
codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc());
CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>();
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 28638d7..6518c32 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -299,9 +299,9 @@
__ Bind(GetEntryLabel());
SaveLiveRegisters(codegen, locations);
- InvokeRuntimeCallingConvention calling_convention;
const uint32_t string_index = instruction_->AsLoadString()->GetStringIndex();
- __ movl(CpuRegister(calling_convention.GetRegisterAt(0)), Immediate(string_index));
+ // Custom calling convention: RAX serves as both input and output.
+ __ movl(CpuRegister(RAX), Immediate(string_index));
x86_64_codegen->InvokeRuntime(kQuickResolveString,
instruction_,
instruction_->GetDexPc(),
@@ -5450,10 +5450,20 @@
: LocationSummary::kNoCall;
LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind);
if (load->GetLoadKind() == HLoadString::LoadKind::kDexCacheViaMethod) {
- locations->SetInAt(0, Location::RequiresRegister());
locations->SetOut(Location::RegisterLocation(RAX));
} else {
locations->SetOut(Location::RequiresRegister());
+ if (load->GetLoadKind() == HLoadString::LoadKind::kBssEntry) {
+ if (!kUseReadBarrier || kUseBakerReadBarrier) {
+ // Rely on the pResolveString and/or marking to save everything.
+ // Custom calling convention: RAX serves as both input and output.
+ RegisterSet caller_saves = RegisterSet::Empty();
+ caller_saves.Add(Location::RegisterLocation(RAX));
+ locations->SetCustomSlowPathCallerSaves(caller_saves);
+ } else {
+ // For non-Baker read barrier we have a temp-clobbering call.
+ }
+ }
}
}
@@ -5493,9 +5503,8 @@
}
// TODO: Re-add the compiler code to do string dex cache lookup again.
- InvokeRuntimeCallingConvention calling_convention;
- __ movl(CpuRegister(calling_convention.GetRegisterAt(0)),
- Immediate(load->GetStringIndex()));
+ // Custom calling convention: RAX serves as both input and output.
+ __ movl(CpuRegister(RAX), Immediate(load->GetStringIndex()));
codegen_->InvokeRuntime(kQuickResolveString,
load,
load->GetDexPc());