summaryrefslogtreecommitdiff
path: root/compiler/optimizing
diff options
context:
space:
mode:
Diffstat (limited to 'compiler/optimizing')
-rw-r--r--compiler/optimizing/builder.h15
-rw-r--r--compiler/optimizing/code_generator.cc2
-rw-r--r--compiler/optimizing/inliner.cc118
-rw-r--r--compiler/optimizing/instruction_builder.cc67
-rw-r--r--compiler/optimizing/instruction_builder.h11
-rw-r--r--compiler/optimizing/intrinsics_arm64.cc162
-rw-r--r--compiler/optimizing/nodes.cc69
-rw-r--r--compiler/optimizing/optimizing_compiler.cc28
-rw-r--r--compiler/optimizing/reference_type_propagation.cc44
-rw-r--r--compiler/optimizing/reference_type_propagation.h3
-rw-r--r--compiler/optimizing/reference_type_propagation_test.cc1
-rw-r--r--compiler/optimizing/sharpening.cc6
-rw-r--r--compiler/optimizing/ssa_builder.cc6
-rw-r--r--compiler/optimizing/ssa_builder.h3
-rw-r--r--compiler/optimizing/stack_map_stream.cc225
-rw-r--r--compiler/optimizing/stack_map_stream.h52
-rw-r--r--compiler/optimizing/stack_map_test.cc94
17 files changed, 544 insertions, 362 deletions
diff --git a/compiler/optimizing/builder.h b/compiler/optimizing/builder.h
index 3a4c9dbd16..e4ad4222fb 100644
--- a/compiler/optimizing/builder.h
+++ b/compiler/optimizing/builder.h
@@ -54,10 +54,7 @@ class HGraphBuilder : public ValueObject {
compiler_driver_(driver),
compilation_stats_(compiler_stats),
block_builder_(graph, dex_file, code_item),
- ssa_builder_(graph,
- dex_compilation_unit->GetClassLoader(),
- dex_compilation_unit->GetDexCache(),
- handles),
+ ssa_builder_(graph, dex_compilation_unit->GetDexCache(), handles),
instruction_builder_(graph,
&block_builder_,
&ssa_builder_,
@@ -83,12 +80,10 @@ class HGraphBuilder : public ValueObject {
code_item_(code_item),
dex_compilation_unit_(nullptr),
compiler_driver_(nullptr),
+ null_dex_cache_(),
compilation_stats_(nullptr),
block_builder_(graph, nullptr, code_item),
- ssa_builder_(graph,
- handles->NewHandle<mirror::ClassLoader>(nullptr),
- handles->NewHandle<mirror::DexCache>(nullptr),
- handles),
+ ssa_builder_(graph, null_dex_cache_, handles),
instruction_builder_(graph,
&block_builder_,
&ssa_builder_,
@@ -101,7 +96,7 @@ class HGraphBuilder : public ValueObject {
/* code_generator */ nullptr,
/* interpreter_metadata */ nullptr,
/* compiler_stats */ nullptr,
- handles->NewHandle<mirror::DexCache>(nullptr),
+ null_dex_cache_,
handles) {}
GraphAnalysisResult BuildGraph();
@@ -122,6 +117,8 @@ class HGraphBuilder : public ValueObject {
CompilerDriver* const compiler_driver_;
+ ScopedNullHandle<mirror::DexCache> null_dex_cache_;
+
OptimizingCompilerStats* compilation_stats_;
HBasicBlockBuilder block_builder_;
diff --git a/compiler/optimizing/code_generator.cc b/compiler/optimizing/code_generator.cc
index d68aa51b1b..bac16cd5df 100644
--- a/compiler/optimizing/code_generator.cc
+++ b/compiler/optimizing/code_generator.cc
@@ -1419,7 +1419,7 @@ void CodeGenerator::EmitJitRoots(uint8_t* code,
QuickEntrypointEnum CodeGenerator::GetArrayAllocationEntrypoint(Handle<mirror::Class> array_klass) {
ScopedObjectAccess soa(Thread::Current());
- if (array_klass.Get() == nullptr) {
+ if (array_klass == nullptr) {
// This can only happen for non-primitive arrays, as primitive arrays can always
// be resolved.
return kQuickAllocArrayResolved32;
diff --git a/compiler/optimizing/inliner.cc b/compiler/optimizing/inliner.cc
index b56ef0f866..e012a4287f 100644
--- a/compiler/optimizing/inliner.cc
+++ b/compiler/optimizing/inliner.cc
@@ -69,38 +69,32 @@ void HInliner::Run() {
// doing some logic in the runtime to discover if a method could have been inlined.
return;
}
- const ArenaVector<HBasicBlock*>& blocks = graph_->GetReversePostOrder();
+ // Keep a copy of all blocks when starting the visit.
+ ArenaVector<HBasicBlock*> blocks = graph_->GetReversePostOrder();
DCHECK(!blocks.empty());
- HBasicBlock* next_block = blocks[0];
- for (size_t i = 0; i < blocks.size(); ++i) {
- // Because we are changing the graph when inlining, we need to remember the next block.
- // This avoids doing the inlining work again on the inlined blocks.
- if (blocks[i] != next_block) {
- continue;
- }
- HBasicBlock* block = next_block;
- next_block = (i == blocks.size() - 1) ? nullptr : blocks[i + 1];
+ // Because we are changing the graph when inlining,
+ // we just iterate over the blocks of the outer method.
+ // This avoids doing the inlining work again on the inlined blocks.
+ for (HBasicBlock* block : blocks) {
for (HInstruction* instruction = block->GetFirstInstruction(); instruction != nullptr;) {
HInstruction* next = instruction->GetNext();
HInvoke* call = instruction->AsInvoke();
// As long as the call is not intrinsified, it is worth trying to inline.
if (call != nullptr && call->GetIntrinsic() == Intrinsics::kNone) {
- // We use the original invoke type to ensure the resolution of the called method
- // works properly.
- if (!TryInline(call)) {
- if (kIsDebugBuild && IsCompilingWithCoreImage()) {
- std::string callee_name =
- outer_compilation_unit_.GetDexFile()->PrettyMethod(call->GetDexMethodIndex());
- bool should_inline = callee_name.find("$inline$") != std::string::npos;
- CHECK(!should_inline) << "Could not inline " << callee_name;
+ if (kIsDebugBuild && IsCompilingWithCoreImage()) {
+ // Debugging case: directives in method names control or assert on inlining.
+ std::string callee_name = outer_compilation_unit_.GetDexFile()->PrettyMethod(
+ call->GetDexMethodIndex(), /* with_signature */ false);
+ // Tests prevent inlining by having $noinline$ in their method names.
+ if (callee_name.find("$noinline$") == std::string::npos) {
+ if (!TryInline(call)) {
+ bool should_have_inlined = (callee_name.find("$inline$") != std::string::npos);
+ CHECK(!should_have_inlined) << "Could not inline " << callee_name;
+ }
}
} else {
- if (kIsDebugBuild && IsCompilingWithCoreImage()) {
- std::string callee_name =
- outer_compilation_unit_.GetDexFile()->PrettyMethod(call->GetDexMethodIndex());
- bool must_not_inline = callee_name.find("$noinline$") != std::string::npos;
- CHECK(!must_not_inline) << "Should not have inlined " << callee_name;
- }
+ // Normal case: try to inline.
+ TryInline(call);
}
}
instruction = next;
@@ -198,9 +192,9 @@ static uint32_t FindMethodIndexIn(ArtMethod* method,
}
static dex::TypeIndex FindClassIndexIn(mirror::Class* cls,
- const DexCompilationUnit& compilation_unit)
+ const DexFile& dex_file,
+ Handle<mirror::DexCache> dex_cache)
REQUIRES_SHARED(Locks::mutator_lock_) {
- const DexFile& dex_file = *compilation_unit.GetDexFile();
dex::TypeIndex index;
if (cls->GetDexCache() == nullptr) {
DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
@@ -209,19 +203,22 @@ static dex::TypeIndex FindClassIndexIn(mirror::Class* cls,
DCHECK(cls->IsProxyClass()) << cls->PrettyClass();
// TODO: deal with proxy classes.
} else if (IsSameDexFile(cls->GetDexFile(), dex_file)) {
- DCHECK_EQ(cls->GetDexCache(), compilation_unit.GetDexCache().Get());
+ DCHECK_EQ(cls->GetDexCache(), dex_cache.Get());
index = cls->GetDexTypeIndex();
+ // Update the dex cache to ensure the class is in. The generated code will
+ // consider it is. We make it safe by updating the dex cache, as other
+ // dex files might also load the class, and there is no guarantee the dex
+ // cache of the dex file of the class will be updated.
+ if (dex_cache->GetResolvedType(index) == nullptr) {
+ dex_cache->SetResolvedType(index, cls);
+ }
} else {
index = cls->FindTypeIndexInOtherDexFile(dex_file);
- // We cannot guarantee the entry will resolve to the same class,
+ // We cannot guarantee the entry in the dex cache will resolve to the same class,
// as there may be different class loaders. So only return the index if it's
- // the right class already resolved with the class loader.
- if (index.IsValid()) {
- ObjPtr<mirror::Class> resolved = ClassLinker::LookupResolvedType(
- index, compilation_unit.GetDexCache().Get(), compilation_unit.GetClassLoader().Get());
- if (resolved != cls) {
- index = dex::TypeIndex::Invalid();
- }
+ // the right class in the dex cache already.
+ if (index.IsValid() && dex_cache->GetResolvedType(index) != cls) {
+ index = dex::TypeIndex::Invalid();
}
}
@@ -377,7 +374,7 @@ bool HInliner::TryInline(HInvoke* invoke_instruction) {
soa.Self(),
class_linker->GetClassRoot(ClassLinker::kClassArrayClass),
InlineCache::kIndividualCacheSize));
- if (inline_cache.Get() == nullptr) {
+ if (inline_cache == nullptr) {
// We got an OOME. Just clear the exception, and don't inline.
DCHECK(soa.Self()->IsExceptionPending());
soa.Self()->ClearException();
@@ -448,8 +445,9 @@ bool HInliner::TryInlineMonomorphicCall(HInvoke* invoke_instruction,
DCHECK(invoke_instruction->IsInvokeVirtual() || invoke_instruction->IsInvokeInterface())
<< invoke_instruction->DebugName();
+ const DexFile& caller_dex_file = *caller_compilation_unit_.GetDexFile();
dex::TypeIndex class_index = FindClassIndexIn(
- GetMonomorphicType(classes), caller_compilation_unit_);
+ GetMonomorphicType(classes), caller_dex_file, caller_compilation_unit_.GetDexCache());
if (!class_index.IsValid()) {
VLOG(compiler) << "Call to " << ArtMethod::PrettyMethod(resolved_method)
<< " from inline cache is not inlined because its class is not"
@@ -492,7 +490,6 @@ bool HInliner::TryInlineMonomorphicCall(HInvoke* invoke_instruction,
// Run type propagation to get the guard typed, and eventually propagate the
// type of the receiver.
ReferenceTypePropagation rtp_fixup(graph_,
- outer_compilation_unit_.GetClassLoader(),
outer_compilation_unit_.GetDexCache(),
handles_,
/* is_first_run */ false);
@@ -563,7 +560,6 @@ HInstruction* HInliner::AddTypeGuard(HInstruction* receiver,
bb_cursor->InsertInstructionAfter(load_class, receiver_class);
load_class->SetLoadKind(kind);
- // TODO: Extend reference type propagation to understand the guard.
HNotEqual* compare = new (graph_->GetArena()) HNotEqual(load_class, receiver_class);
bb_cursor->InsertInstructionAfter(compare, load_class);
if (with_deoptimization) {
@@ -587,6 +583,7 @@ bool HInliner::TryInlinePolymorphicCall(HInvoke* invoke_instruction,
ClassLinker* class_linker = caller_compilation_unit_.GetClassLinker();
PointerSize pointer_size = class_linker->GetImagePointerSize();
+ const DexFile& caller_dex_file = *caller_compilation_unit_.GetDexFile();
bool all_targets_inlined = true;
bool one_target_inlined = false;
@@ -608,7 +605,8 @@ bool HInliner::TryInlinePolymorphicCall(HInvoke* invoke_instruction,
HInstruction* cursor = invoke_instruction->GetPrevious();
HBasicBlock* bb_cursor = invoke_instruction->GetBlock();
- dex::TypeIndex class_index = FindClassIndexIn(handle.Get(), caller_compilation_unit_);
+ dex::TypeIndex class_index = FindClassIndexIn(
+ handle.Get(), caller_dex_file, caller_compilation_unit_.GetDexCache());
HInstruction* return_replacement = nullptr;
if (!class_index.IsValid() ||
!TryBuildAndInline(invoke_instruction,
@@ -664,7 +662,6 @@ bool HInliner::TryInlinePolymorphicCall(HInvoke* invoke_instruction,
// Run type propagation to get the guards typed.
ReferenceTypePropagation rtp_fixup(graph_,
- outer_compilation_unit_.GetClassLoader(),
outer_compilation_unit_.GetDexCache(),
handles_,
/* is_first_run */ false);
@@ -846,7 +843,6 @@ bool HInliner::TryInlinePolymorphicCallToSameTarget(
if (outermost_graph_->IsCompilingOsr()) {
CreateDiamondPatternForPolymorphicInline(compare, return_replacement, invoke_instruction);
} else {
- // TODO: Extend reference type propagation to understand the guard.
HDeoptimize* deoptimize = new (graph_->GetArena()) HDeoptimize(
compare, invoke_instruction->GetDexPc());
bb_cursor->InsertInstructionAfter(deoptimize, compare);
@@ -859,7 +855,6 @@ bool HInliner::TryInlinePolymorphicCallToSameTarget(
// Run type propagation to get the guard typed.
ReferenceTypePropagation rtp_fixup(graph_,
- outer_compilation_unit_.GetClassLoader(),
outer_compilation_unit_.GetDexCache(),
handles_,
/* is_first_run */ false);
@@ -928,7 +923,6 @@ bool HInliner::TryInlineAndReplace(HInvoke* invoke_instruction,
// Actual return value has a more specific type than the method's declared
// return type. Run RTP again on the outer graph to propagate it.
ReferenceTypePropagation(graph_,
- outer_compilation_unit_.GetClassLoader(),
outer_compilation_unit_.GetDexCache(),
handles_,
/* is_first_run */ false).Run();
@@ -1181,11 +1175,7 @@ HInstanceFieldGet* HInliner::CreateInstanceFieldGet(Handle<mirror::DexCache> dex
/* dex_pc */ 0);
if (iget->GetType() == Primitive::kPrimNot) {
// Use the same dex_cache that we used for field lookup as the hint_dex_cache.
- ReferenceTypePropagation rtp(graph_,
- outer_compilation_unit_.GetClassLoader(),
- dex_cache,
- handles_,
- /* is_first_run */ false);
+ ReferenceTypePropagation rtp(graph_, dex_cache, handles_, /* is_first_run */ false);
rtp.Visit(iget);
}
return iget;
@@ -1231,7 +1221,7 @@ bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
resolved_method->GetDeclaringClass()->GetClassLoader()));
DexCompilationUnit dex_compilation_unit(
- class_loader,
+ class_loader.ToJObject(),
class_linker,
callee_dex_file,
code_item,
@@ -1348,7 +1338,6 @@ bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
// are more specific than the declared ones, run RTP again on the inner graph.
if (run_rtp || ArgumentTypesMoreSpecific(invoke_instruction, resolved_method)) {
ReferenceTypePropagation(callee_graph,
- outer_compilation_unit_.GetClassLoader(),
dex_compilation_unit.GetDexCache(),
handles_,
/* is_first_run */ false).Run();
@@ -1359,9 +1348,6 @@ bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
RunOptimizations(callee_graph, code_item, dex_compilation_unit);
number_of_instructions_budget += number_of_inlined_instructions;
- // TODO: We should abort only if all predecessors throw. However,
- // HGraph::InlineInto currently does not handle an exit block with
- // a throw predecessor.
HBasicBlock* exit_block = callee_graph->GetExitBlock();
if (exit_block == nullptr) {
VLOG(compiler) << "Method " << callee_dex_file.PrettyMethod(method_index)
@@ -1369,16 +1355,30 @@ bool HInliner::TryBuildAndInlineHelper(HInvoke* invoke_instruction,
return false;
}
- bool has_throw_predecessor = false;
+ bool has_one_return = false;
for (HBasicBlock* predecessor : exit_block->GetPredecessors()) {
if (predecessor->GetLastInstruction()->IsThrow()) {
- has_throw_predecessor = true;
- break;
+ if (invoke_instruction->GetBlock()->IsTryBlock()) {
+ // TODO(ngeoffray): Support adding HTryBoundary in Hgraph::InlineInto.
+ VLOG(compiler) << "Method " << callee_dex_file.PrettyMethod(method_index)
+ << " could not be inlined because one branch always throws and"
+ << " caller is in a try/catch block";
+ return false;
+ } else if (graph_->GetExitBlock() == nullptr) {
+ // TODO(ngeoffray): Support adding HExit in the caller graph.
+ VLOG(compiler) << "Method " << callee_dex_file.PrettyMethod(method_index)
+ << " could not be inlined because one branch always throws and"
+ << " caller does not have an exit block";
+ return false;
+ }
+ } else {
+ has_one_return = true;
}
}
- if (has_throw_predecessor) {
+
+ if (!has_one_return) {
VLOG(compiler) << "Method " << callee_dex_file.PrettyMethod(method_index)
- << " could not be inlined because one branch always throws";
+ << " could not be inlined because it always throws";
return false;
}
diff --git a/compiler/optimizing/instruction_builder.cc b/compiler/optimizing/instruction_builder.cc
index 3aaf2ca102..3374e42955 100644
--- a/compiler/optimizing/instruction_builder.cc
+++ b/compiler/optimizing/instruction_builder.cc
@@ -669,17 +669,18 @@ static InvokeType GetInvokeTypeFromOpCode(Instruction::Code opcode) {
ArtMethod* HInstructionBuilder::ResolveMethod(uint16_t method_idx, InvokeType invoke_type) {
ScopedObjectAccess soa(Thread::Current());
- StackHandleScope<2> hs(soa.Self());
+ StackHandleScope<3> hs(soa.Self());
ClassLinker* class_linker = dex_compilation_unit_->GetClassLinker();
- Handle<mirror::ClassLoader> class_loader = dex_compilation_unit_->GetClassLoader();
+ Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(dex_compilation_unit_->GetClassLoader())));
Handle<mirror::Class> compiling_class(hs.NewHandle(GetCompilingClass()));
// We fetch the referenced class eagerly (that is, the class pointed by in the MethodId
// at method_idx), as `CanAccessResolvedMethod` expects it be be in the dex cache.
Handle<mirror::Class> methods_class(hs.NewHandle(class_linker->ResolveReferencedClassOfMethod(
method_idx, dex_compilation_unit_->GetDexCache(), class_loader)));
- if (UNLIKELY(methods_class.Get() == nullptr)) {
+ if (UNLIKELY(methods_class == nullptr)) {
// Clean up any exception left by type resolution.
soa.Self()->ClearException();
return nullptr;
@@ -701,7 +702,7 @@ ArtMethod* HInstructionBuilder::ResolveMethod(uint16_t method_idx, InvokeType in
// Check access. The class linker has a fast path for looking into the dex cache
// and does not check the access if it hits it.
- if (compiling_class.Get() == nullptr) {
+ if (compiling_class == nullptr) {
if (!resolved_method->IsPublic()) {
return nullptr;
}
@@ -717,7 +718,7 @@ ArtMethod* HInstructionBuilder::ResolveMethod(uint16_t method_idx, InvokeType in
// make this an invoke-unresolved to handle cross-dex invokes or abstract super methods, both of
// which require runtime handling.
if (invoke_type == kSuper) {
- if (compiling_class.Get() == nullptr) {
+ if (compiling_class == nullptr) {
// We could not determine the method's class we need to wait until runtime.
DCHECK(Runtime::Current()->IsAotCompiler());
return nullptr;
@@ -953,7 +954,7 @@ bool HInstructionBuilder::BuildNewInstance(dex::TypeIndex type_index, uint32_t d
}
// Consider classes we haven't resolved as potentially finalizable.
- bool finalizable = (klass.Get() == nullptr) || klass->IsFinalizable();
+ bool finalizable = (klass == nullptr) || klass->IsFinalizable();
AppendInstruction(new (arena_) HNewInstance(
cls,
@@ -971,7 +972,7 @@ static bool IsSubClass(mirror::Class* to_test, mirror::Class* super_class)
}
bool HInstructionBuilder::IsInitialized(Handle<mirror::Class> cls) const {
- if (cls.Get() == nullptr) {
+ if (cls == nullptr) {
return false;
}
@@ -1259,7 +1260,9 @@ bool HInstructionBuilder::BuildInstanceFieldAccess(const Instruction& instructio
static mirror::Class* GetClassFrom(CompilerDriver* driver,
const DexCompilationUnit& compilation_unit) {
ScopedObjectAccess soa(Thread::Current());
- Handle<mirror::ClassLoader> class_loader = compilation_unit.GetClassLoader();
+ StackHandleScope<1> hs(soa.Self());
+ Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(compilation_unit.GetClassLoader())));
Handle<mirror::DexCache> dex_cache = compilation_unit.GetDexCache();
return driver->ResolveCompilingMethodsClass(soa, dex_cache, class_loader, &compilation_unit);
@@ -1275,9 +1278,10 @@ mirror::Class* HInstructionBuilder::GetCompilingClass() const {
bool HInstructionBuilder::IsOutermostCompilingClass(dex::TypeIndex type_index) const {
ScopedObjectAccess soa(Thread::Current());
- StackHandleScope<2> hs(soa.Self());
+ StackHandleScope<3> hs(soa.Self());
Handle<mirror::DexCache> dex_cache = dex_compilation_unit_->GetDexCache();
- Handle<mirror::ClassLoader> class_loader = dex_compilation_unit_->GetClassLoader();
+ Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(dex_compilation_unit_->GetClassLoader())));
Handle<mirror::Class> cls(hs.NewHandle(compiler_driver_->ResolveClass(
soa, dex_cache, class_loader, type_index, dex_compilation_unit_)));
Handle<mirror::Class> outer_class(hs.NewHandle(GetOutermostCompilingClass()));
@@ -1288,7 +1292,7 @@ bool HInstructionBuilder::IsOutermostCompilingClass(dex::TypeIndex type_index) c
// When this happens we cannot establish a direct relation between the current
// class and the outer class, so we return false.
// (Note that this is only used for optimizing invokes and field accesses)
- return (cls.Get() != nullptr) && (outer_class.Get() == cls.Get());
+ return (cls != nullptr) && (outer_class.Get() == cls.Get());
}
void HInstructionBuilder::BuildUnresolvedStaticFieldAccess(const Instruction& instruction,
@@ -1313,7 +1317,8 @@ ArtField* HInstructionBuilder::ResolveField(uint16_t field_idx, bool is_static,
StackHandleScope<2> hs(soa.Self());
ClassLinker* class_linker = dex_compilation_unit_->GetClassLinker();
- Handle<mirror::ClassLoader> class_loader = dex_compilation_unit_->GetClassLoader();
+ Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(dex_compilation_unit_->GetClassLoader())));
Handle<mirror::Class> compiling_class(hs.NewHandle(GetCompilingClass()));
ArtField* resolved_field = class_linker->ResolveField(*dex_compilation_unit_->GetDexFile(),
@@ -1335,7 +1340,7 @@ ArtField* HInstructionBuilder::ResolveField(uint16_t field_idx, bool is_static,
}
// Check access.
- if (compiling_class.Get() == nullptr) {
+ if (compiling_class == nullptr) {
if (!resolved_field->IsPublic()) {
return nullptr;
}
@@ -1607,7 +1612,7 @@ void HInstructionBuilder::BuildFillWideArrayData(HInstruction* object,
static TypeCheckKind ComputeTypeCheckKind(Handle<mirror::Class> cls)
REQUIRES_SHARED(Locks::mutator_lock_) {
- if (cls.Get() == nullptr) {
+ if (cls == nullptr) {
return TypeCheckKind::kUnresolvedCheck;
} else if (cls->IsInterface()) {
return TypeCheckKind::kInterfaceCheck;
@@ -1630,13 +1635,15 @@ static TypeCheckKind ComputeTypeCheckKind(Handle<mirror::Class> cls)
HLoadClass* HInstructionBuilder::BuildLoadClass(dex::TypeIndex type_index, uint32_t dex_pc) {
ScopedObjectAccess soa(Thread::Current());
+ StackHandleScope<2> hs(soa.Self());
const DexFile& dex_file = *dex_compilation_unit_->GetDexFile();
- Handle<mirror::ClassLoader> class_loader = dex_compilation_unit_->GetClassLoader();
+ Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(dex_compilation_unit_->GetClassLoader())));
Handle<mirror::Class> klass = handles_->NewHandle(compiler_driver_->ResolveClass(
soa, dex_compilation_unit_->GetDexCache(), class_loader, type_index, dex_compilation_unit_));
bool needs_access_check = true;
- if (klass.Get() != nullptr) {
+ if (klass != nullptr) {
if (klass->IsPublic()) {
needs_access_check = false;
} else {
@@ -1672,7 +1679,7 @@ HLoadClass* HInstructionBuilder::BuildLoadClass(dex::TypeIndex type_index,
type_index,
*actual_dex_file,
klass,
- klass.Get() != nullptr && (klass.Get() == GetOutermostCompilingClass()),
+ klass != nullptr && (klass.Get() == GetOutermostCompilingClass()),
dex_pc,
needs_access_check);
@@ -1715,9 +1722,17 @@ void HInstructionBuilder::BuildTypeCheck(const Instruction& instruction,
}
}
-bool HInstructionBuilder::NeedsAccessCheck(dex::TypeIndex type_index, bool* finalizable) const {
+bool HInstructionBuilder::NeedsAccessCheck(dex::TypeIndex type_index,
+ Handle<mirror::DexCache> dex_cache,
+ bool* finalizable) const {
return !compiler_driver_->CanAccessInstantiableTypeWithoutChecks(
- LookupReferrerClass(), LookupResolvedType(type_index, *dex_compilation_unit_), finalizable);
+ dex_compilation_unit_->GetDexMethodIndex(), dex_cache, type_index, finalizable);
+}
+
+bool HInstructionBuilder::NeedsAccessCheck(dex::TypeIndex type_index, bool* finalizable) const {
+ ScopedObjectAccess soa(Thread::Current());
+ Handle<mirror::DexCache> dex_cache = dex_compilation_unit_->GetDexCache();
+ return NeedsAccessCheck(type_index, dex_cache, finalizable);
}
bool HInstructionBuilder::CanDecodeQuickenedInfo() const {
@@ -2757,18 +2772,4 @@ bool HInstructionBuilder::ProcessDexInstruction(const Instruction& instruction,
return true;
} // NOLINT(readability/fn_size)
-ObjPtr<mirror::Class> HInstructionBuilder::LookupResolvedType(
- dex::TypeIndex type_index,
- const DexCompilationUnit& compilation_unit) const {
- return ClassLinker::LookupResolvedType(
- type_index, compilation_unit.GetDexCache().Get(), compilation_unit.GetClassLoader().Get());
-}
-
-ObjPtr<mirror::Class> HInstructionBuilder::LookupReferrerClass() const {
- // TODO: Cache the result in a Handle<mirror::Class>.
- const DexFile::MethodId& method_id =
- dex_compilation_unit_->GetDexFile()->GetMethodId(dex_compilation_unit_->GetDexMethodIndex());
- return LookupResolvedType(method_id.class_idx_, *dex_compilation_unit_);
-}
-
} // namespace art
diff --git a/compiler/optimizing/instruction_builder.h b/compiler/optimizing/instruction_builder.h
index e735a0c46d..3bb680ce44 100644
--- a/compiler/optimizing/instruction_builder.h
+++ b/compiler/optimizing/instruction_builder.h
@@ -106,8 +106,11 @@ class HInstructionBuilder : public ValueObject {
// Returns whether the current method needs access check for the type.
// Output parameter finalizable is set to whether the type is finalizable.
- bool NeedsAccessCheck(dex::TypeIndex type_index, /*out*/bool* finalizable) const
+ bool NeedsAccessCheck(dex::TypeIndex type_index,
+ Handle<mirror::DexCache> dex_cache,
+ /*out*/bool* finalizable) const
REQUIRES_SHARED(Locks::mutator_lock_);
+ bool NeedsAccessCheck(dex::TypeIndex type_index, /*out*/bool* finalizable) const;
template<typename T>
void Unop_12x(const Instruction& instruction, Primitive::Type type, uint32_t dex_pc);
@@ -297,12 +300,6 @@ class HInstructionBuilder : public ValueObject {
// be found.
ArtField* ResolveField(uint16_t field_idx, bool is_static, bool is_put);
- ObjPtr<mirror::Class> LookupResolvedType(dex::TypeIndex type_index,
- const DexCompilationUnit& compilation_unit) const
- REQUIRES_SHARED(Locks::mutator_lock_);
-
- ObjPtr<mirror::Class> LookupReferrerClass() const REQUIRES_SHARED(Locks::mutator_lock_);
-
ArenaAllocator* const arena_;
HGraph* const graph_;
VariableSizedHandleScope* handles_;
diff --git a/compiler/optimizing/intrinsics_arm64.cc b/compiler/optimizing/intrinsics_arm64.cc
index 1047d3beb6..86e54294ae 100644
--- a/compiler/optimizing/intrinsics_arm64.cc
+++ b/compiler/optimizing/intrinsics_arm64.cc
@@ -23,7 +23,7 @@
#include "entrypoints/quick/quick_entrypoints.h"
#include "intrinsics.h"
#include "mirror/array-inl.h"
-#include "mirror/string.h"
+#include "mirror/string-inl.h"
#include "thread.h"
#include "utils/arm64/assembler_arm64.h"
@@ -1450,16 +1450,47 @@ void IntrinsicCodeGeneratorARM64::VisitStringCompareTo(HInvoke* invoke) {
}
}
+// The cut off for unrolling the loop in String.equals() intrinsic for const strings.
+// The normal loop plus the pre-header is 9 instructions without string compression and 12
+// instructions with string compression. We can compare up to 8 bytes in 4 instructions
+// (LDR+LDR+CMP+BNE) and up to 16 bytes in 5 instructions (LDP+LDP+CMP+CCMP+BNE). Allow up
+// to 10 instructions for the unrolled loop.
+constexpr size_t kShortConstStringEqualsCutoffInBytes = 32;
+
+static const char* GetConstString(HInstruction* candidate, uint32_t* utf16_length) {
+ if (candidate->IsLoadString()) {
+ HLoadString* load_string = candidate->AsLoadString();
+ const DexFile& dex_file = load_string->GetDexFile();
+ return dex_file.StringDataAndUtf16LengthByIdx(load_string->GetStringIndex(), utf16_length);
+ }
+ return nullptr;
+}
+
void IntrinsicLocationsBuilderARM64::VisitStringEquals(HInvoke* invoke) {
LocationSummary* locations = new (arena_) LocationSummary(invoke,
LocationSummary::kNoCall,
kIntrinsified);
locations->SetInAt(0, Location::RequiresRegister());
locations->SetInAt(1, Location::RequiresRegister());
- // Temporary registers to store lengths of strings and for calculations.
- locations->AddTemp(Location::RequiresRegister());
- locations->AddTemp(Location::RequiresRegister());
+ // For the generic implementation and for long const strings we need a temporary.
+ // We do not need it for short const strings, up to 8 bytes, see code generation below.
+ uint32_t const_string_length = 0u;
+ const char* const_string = GetConstString(invoke->InputAt(0), &const_string_length);
+ if (const_string == nullptr) {
+ const_string = GetConstString(invoke->InputAt(1), &const_string_length);
+ }
+ bool is_compressed =
+ mirror::kUseStringCompression &&
+ const_string != nullptr &&
+ mirror::String::DexFileStringAllASCII(const_string, const_string_length);
+ if (const_string == nullptr || const_string_length > (is_compressed ? 8u : 4u)) {
+ locations->AddTemp(Location::RequiresRegister());
+ }
+
+ // TODO: If the String.equals() is used only for an immediately following HIf, we can
+ // mark it as emitted-at-use-site and emit branches directly to the appropriate blocks.
+ // Then we shall need an extra temporary register instead of the output register.
locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
}
@@ -1473,8 +1504,7 @@ void IntrinsicCodeGeneratorARM64::VisitStringEquals(HInvoke* invoke) {
UseScratchRegisterScope scratch_scope(masm);
Register temp = scratch_scope.AcquireW();
- Register temp1 = WRegisterFrom(locations->GetTemp(0));
- Register temp2 = WRegisterFrom(locations->GetTemp(1));
+ Register temp1 = scratch_scope.AcquireW();
vixl::aarch64::Label loop;
vixl::aarch64::Label end;
@@ -1510,46 +1540,98 @@ void IntrinsicCodeGeneratorARM64::VisitStringEquals(HInvoke* invoke) {
__ B(&return_false, ne);
}
- // Load `count` fields of this and argument strings.
- __ Ldr(temp, MemOperand(str.X(), count_offset));
- __ Ldr(temp1, MemOperand(arg.X(), count_offset));
- // Check if `count` fields are equal, return false if they're not.
- // Also compares the compression style, if differs return false.
- __ Cmp(temp, temp1);
- __ B(&return_false, ne);
- // Return true if both strings are empty. Even with string compression `count == 0` means empty.
- static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u,
- "Expecting 0=compressed, 1=uncompressed");
- __ Cbz(temp, &return_true);
+ // Check if one of the inputs is a const string. Do not special-case both strings
+ // being const, such cases should be handled by constant folding if needed.
+ uint32_t const_string_length = 0u;
+ const char* const_string = GetConstString(invoke->InputAt(0), &const_string_length);
+ if (const_string == nullptr) {
+ const_string = GetConstString(invoke->InputAt(1), &const_string_length);
+ if (const_string != nullptr) {
+ std::swap(str, arg); // Make sure the const string is in `str`.
+ }
+ }
+ bool is_compressed =
+ mirror::kUseStringCompression &&
+ const_string != nullptr &&
+ mirror::String::DexFileStringAllASCII(const_string, const_string_length);
+
+ if (const_string != nullptr) {
+ // Load `count` field of the argument string and check if it matches the const string.
+ // Also compares the compression style, if differs return false.
+ __ Ldr(temp, MemOperand(arg.X(), count_offset));
+ __ Cmp(temp, Operand(mirror::String::GetFlaggedCount(const_string_length, is_compressed)));
+ __ B(&return_false, ne);
+ } else {
+ // Load `count` fields of this and argument strings.
+ __ Ldr(temp, MemOperand(str.X(), count_offset));
+ __ Ldr(temp1, MemOperand(arg.X(), count_offset));
+ // Check if `count` fields are equal, return false if they're not.
+ // Also compares the compression style, if differs return false.
+ __ Cmp(temp, temp1);
+ __ B(&return_false, ne);
+ }
// Assertions that must hold in order to compare strings 8 bytes at a time.
DCHECK_ALIGNED(value_offset, 8);
static_assert(IsAligned<8>(kObjectAlignment), "String of odd length is not zero padded");
- if (mirror::kUseStringCompression) {
- // For string compression, calculate the number of bytes to compare (not chars).
- // This could in theory exceed INT32_MAX, so treat temp as unsigned.
- __ Lsr(temp, temp, 1u); // Extract length.
- __ And(temp1, temp1, Operand(1)); // Extract compression flag.
- __ Lsl(temp, temp, temp1); // Calculate number of bytes to compare.
- }
-
- // Store offset of string value in preparation for comparison loop
- __ Mov(temp1, value_offset);
+ if (const_string != nullptr &&
+ const_string_length < (is_compressed ? kShortConstStringEqualsCutoffInBytes
+ : kShortConstStringEqualsCutoffInBytes / 2u)) {
+ // Load and compare the contents. Though we know the contents of the short const string
+ // at compile time, materializing constants may be more code than loading from memory.
+ int32_t offset = value_offset;
+ size_t remaining_bytes =
+ RoundUp(is_compressed ? const_string_length : const_string_length * 2u, 8u);
+ temp = temp.X();
+ temp1 = temp1.X();
+ while (remaining_bytes > 8u) {
+ Register temp2 = XRegisterFrom(locations->GetTemp(0));
+ __ Ldp(temp, temp1, MemOperand(str.X(), offset));
+ __ Ldp(temp2, out, MemOperand(arg.X(), offset));
+ __ Cmp(temp, temp2);
+ __ Ccmp(temp1, out, NoFlag, eq);
+ __ B(&return_false, ne);
+ offset += 2u * sizeof(uint64_t);
+ remaining_bytes -= 2u * sizeof(uint64_t);
+ }
+ if (remaining_bytes != 0u) {
+ __ Ldr(temp, MemOperand(str.X(), offset));
+ __ Ldr(temp1, MemOperand(arg.X(), offset));
+ __ Cmp(temp, temp1);
+ __ B(&return_false, ne);
+ }
+ } else {
+ // Return true if both strings are empty. Even with string compression `count == 0` means empty.
+ static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u,
+ "Expecting 0=compressed, 1=uncompressed");
+ __ Cbz(temp, &return_true);
+
+ if (mirror::kUseStringCompression) {
+ // For string compression, calculate the number of bytes to compare (not chars).
+ // This could in theory exceed INT32_MAX, so treat temp as unsigned.
+ __ And(temp1, temp, Operand(1)); // Extract compression flag.
+ __ Lsr(temp, temp, 1u); // Extract length.
+ __ Lsl(temp, temp, temp1); // Calculate number of bytes to compare.
+ }
- temp1 = temp1.X();
- temp2 = temp2.X();
- // Loop to compare strings 8 bytes at a time starting at the front of the string.
- // Ok to do this because strings are zero-padded to kObjectAlignment.
- __ Bind(&loop);
- __ Ldr(out, MemOperand(str.X(), temp1));
- __ Ldr(temp2, MemOperand(arg.X(), temp1));
- __ Add(temp1, temp1, Operand(sizeof(uint64_t)));
- __ Cmp(out, temp2);
- __ B(&return_false, ne);
- // With string compression, we have compared 8 bytes, otherwise 4 chars.
- __ Sub(temp, temp, Operand(mirror::kUseStringCompression ? 8 : 4), SetFlags);
- __ B(&loop, hi);
+ // Store offset of string value in preparation for comparison loop
+ __ Mov(temp1, value_offset);
+
+ temp1 = temp1.X();
+ Register temp2 = XRegisterFrom(locations->GetTemp(0));
+ // Loop to compare strings 8 bytes at a time starting at the front of the string.
+ // Ok to do this because strings are zero-padded to kObjectAlignment.
+ __ Bind(&loop);
+ __ Ldr(out, MemOperand(str.X(), temp1));
+ __ Ldr(temp2, MemOperand(arg.X(), temp1));
+ __ Add(temp1, temp1, Operand(sizeof(uint64_t)));
+ __ Cmp(out, temp2);
+ __ B(&return_false, ne);
+ // With string compression, we have compared 8 bytes, otherwise 4 chars.
+ __ Sub(temp, temp, Operand(mirror::kUseStringCompression ? 8 : 4), SetFlags);
+ __ B(&loop, hi);
+ }
// Return true and exit the function.
// If loop does not result in returning false, we return true.
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index abbb91a1a9..71a26ebe79 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -2038,6 +2038,8 @@ HInstruction* HGraph::InlineInto(HGraph* outer_graph, HInvoke* invoke) {
HInstruction* return_value = nullptr;
if (GetBlocks().size() == 3) {
+ // Inliner already made sure we don't inline methods that always throw.
+ DCHECK(!GetBlocks()[1]->GetLastInstruction()->IsThrow());
// Simple case of an entry block, a body block, and an exit block.
// Put the body block's instruction into `invoke`'s block.
HBasicBlock* body = GetBlocks()[1];
@@ -2119,33 +2121,60 @@ HInstruction* HGraph::InlineInto(HGraph* outer_graph, HInvoke* invoke) {
UpdateLoopAndTryInformationOfNewBlock(to, at, /* replace_if_back_edge */ true);
// Update all predecessors of the exit block (now the `to` block)
- // to not `HReturn` but `HGoto` instead.
- bool returns_void = to->GetPredecessors()[0]->GetLastInstruction()->IsReturnVoid();
- if (to->GetPredecessors().size() == 1) {
- HBasicBlock* predecessor = to->GetPredecessors()[0];
+ // to not `HReturn` but `HGoto` instead. Special case throwing blocks
+ // to now get the outer graph exit block as successor. Note that the inliner
+ // currently doesn't support inlining methods with try/catch.
+ HPhi* return_value_phi = nullptr;
+ bool rerun_dominance = false;
+ bool rerun_loop_analysis = false;
+ for (size_t pred = 0; pred < to->GetPredecessors().size(); ++pred) {
+ HBasicBlock* predecessor = to->GetPredecessors()[pred];
HInstruction* last = predecessor->GetLastInstruction();
- if (!returns_void) {
- return_value = last->InputAt(0);
- }
- predecessor->AddInstruction(new (allocator) HGoto(last->GetDexPc()));
- predecessor->RemoveInstruction(last);
- } else {
- if (!returns_void) {
- // There will be multiple returns.
- return_value = new (allocator) HPhi(
- allocator, kNoRegNumber, 0, HPhi::ToPhiType(invoke->GetType()), to->GetDexPc());
- to->AddPhi(return_value->AsPhi());
- }
- for (HBasicBlock* predecessor : to->GetPredecessors()) {
- HInstruction* last = predecessor->GetLastInstruction();
- if (!returns_void) {
+ if (last->IsThrow()) {
+ DCHECK(!at->IsTryBlock());
+ predecessor->ReplaceSuccessor(to, outer_graph->GetExitBlock());
+ --pred;
+ // We need to re-run dominance information, as the exit block now has
+ // a new dominator.
+ rerun_dominance = true;
+ if (predecessor->GetLoopInformation() != nullptr) {
+ // The exit block and blocks post dominated by the exit block do not belong
+ // to any loop. Because we do not compute the post dominators, we need to re-run
+ // loop analysis to get the loop information correct.
+ rerun_loop_analysis = true;
+ }
+ } else {
+ if (last->IsReturnVoid()) {
+ DCHECK(return_value == nullptr);
+ DCHECK(return_value_phi == nullptr);
+ } else {
DCHECK(last->IsReturn());
- return_value->AsPhi()->AddInput(last->InputAt(0));
+ if (return_value_phi != nullptr) {
+ return_value_phi->AddInput(last->InputAt(0));
+ } else if (return_value == nullptr) {
+ return_value = last->InputAt(0);
+ } else {
+ // There will be multiple returns.
+ return_value_phi = new (allocator) HPhi(
+ allocator, kNoRegNumber, 0, HPhi::ToPhiType(invoke->GetType()), to->GetDexPc());
+ to->AddPhi(return_value_phi);
+ return_value_phi->AddInput(return_value);
+ return_value_phi->AddInput(last->InputAt(0));
+ return_value = return_value_phi;
+ }
}
predecessor->AddInstruction(new (allocator) HGoto(last->GetDexPc()));
predecessor->RemoveInstruction(last);
}
}
+ if (rerun_loop_analysis) {
+ outer_graph->ClearLoopInformation();
+ outer_graph->ClearDominanceInformation();
+ outer_graph->BuildDominatorTree();
+ } else if (rerun_dominance) {
+ outer_graph->ClearDominanceInformation();
+ outer_graph->ComputeDominanceInformation();
+ }
}
// Walk over the entry block and:
diff --git a/compiler/optimizing/optimizing_compiler.cc b/compiler/optimizing/optimizing_compiler.cc
index 0375c66e42..8638e346fb 100644
--- a/compiler/optimizing/optimizing_compiler.cc
+++ b/compiler/optimizing/optimizing_compiler.cc
@@ -306,7 +306,7 @@ class OptimizingCompiler FINAL : public Compiler {
InvokeType invoke_type,
uint16_t class_def_idx,
uint32_t method_idx,
- Handle<mirror::ClassLoader> class_loader,
+ jobject class_loader,
const DexFile& dex_file,
Handle<mirror::DexCache> dex_cache) const OVERRIDE;
@@ -375,7 +375,7 @@ class OptimizingCompiler FINAL : public Compiler {
InvokeType invoke_type,
uint16_t class_def_idx,
uint32_t method_idx,
- Handle<mirror::ClassLoader> class_loader,
+ jobject class_loader,
const DexFile& dex_file,
Handle<mirror::DexCache> dex_cache,
ArtMethod* method,
@@ -875,7 +875,7 @@ CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena,
InvokeType invoke_type,
uint16_t class_def_idx,
uint32_t method_idx,
- Handle<mirror::ClassLoader> class_loader,
+ jobject class_loader,
const DexFile& dex_file,
Handle<mirror::DexCache> dex_cache,
ArtMethod* method,
@@ -946,8 +946,11 @@ CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena,
const uint8_t* interpreter_metadata = nullptr;
if (method == nullptr) {
ScopedObjectAccess soa(Thread::Current());
+ StackHandleScope<1> hs(soa.Self());
+ Handle<mirror::ClassLoader> loader(hs.NewHandle(
+ soa.Decode<mirror::ClassLoader>(class_loader)));
method = compiler_driver->ResolveMethod(
- soa, dex_cache, class_loader, &dex_compilation_unit, method_idx, invoke_type);
+ soa, dex_cache, loader, &dex_compilation_unit, method_idx, invoke_type);
}
// For AOT compilation, we may not get a method, for example if its class is erroneous.
// JIT should always have a method.
@@ -956,6 +959,16 @@ CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena,
graph->SetArtMethod(method);
ScopedObjectAccess soa(Thread::Current());
interpreter_metadata = method->GetQuickenedInfo(class_linker->GetImagePointerSize());
+ dex::TypeIndex type_index = method->GetDeclaringClass()->GetDexTypeIndex();
+
+ // Update the dex cache if the type is not in it yet. Note that under AOT,
+ // the verifier must have set it, but under JIT, there's no guarantee, as we
+ // don't necessarily run the verifier.
+ // The compiler and the compiler driver assume the compiling class is
+ // in the dex cache.
+ if (dex_cache->GetResolvedType(type_index) == nullptr) {
+ dex_cache->SetResolvedType(type_index, method->GetDeclaringClass());
+ }
}
std::unique_ptr<CodeGenerator> codegen(
@@ -1036,7 +1049,7 @@ CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item,
InvokeType invoke_type,
uint16_t class_def_idx,
uint32_t method_idx,
- Handle<mirror::ClassLoader> jclass_loader,
+ jobject jclass_loader,
const DexFile& dex_file,
Handle<mirror::DexCache> dex_cache) const {
CompilerDriver* compiler_driver = GetCompilerDriver();
@@ -1150,6 +1163,7 @@ bool OptimizingCompiler::JitCompile(Thread* self,
Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache()));
DCHECK(method->IsCompilable());
+ jobject jclass_loader = class_loader.ToJObject();
const DexFile* dex_file = method->GetDexFile();
const uint16_t class_def_idx = method->GetClassDefIndex();
const DexFile::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset());
@@ -1173,7 +1187,7 @@ bool OptimizingCompiler::JitCompile(Thread* self,
invoke_type,
class_def_idx,
method_idx,
- class_loader,
+ jclass_loader,
*dex_file,
dex_cache,
method,
@@ -1200,7 +1214,7 @@ bool OptimizingCompiler::JitCompile(Thread* self,
Handle<mirror::ObjectArray<mirror::Object>> roots(
hs.NewHandle(mirror::ObjectArray<mirror::Object>::Alloc(
self, class_linker->GetClassRoot(ClassLinker::kObjectArrayClass), number_of_roots)));
- if (roots.Get() == nullptr) {
+ if (roots == nullptr) {
// Out of memory, just clear the exception to avoid any Java exception uncaught problems.
DCHECK(self->IsExceptionPending());
self->ClearException();
diff --git a/compiler/optimizing/reference_type_propagation.cc b/compiler/optimizing/reference_type_propagation.cc
index 6e332ca59b..c55fccc7d3 100644
--- a/compiler/optimizing/reference_type_propagation.cc
+++ b/compiler/optimizing/reference_type_propagation.cc
@@ -65,13 +65,11 @@ ReferenceTypeInfo::TypeHandle ReferenceTypePropagation::HandleCache::GetThrowabl
class ReferenceTypePropagation::RTPVisitor : public HGraphDelegateVisitor {
public:
RTPVisitor(HGraph* graph,
- Handle<mirror::ClassLoader> class_loader,
Handle<mirror::DexCache> hint_dex_cache,
HandleCache* handle_cache,
ArenaVector<HInstruction*>* worklist,
bool is_first_run)
: HGraphDelegateVisitor(graph),
- class_loader_(class_loader),
hint_dex_cache_(hint_dex_cache),
handle_cache_(handle_cache),
worklist_(worklist),
@@ -103,7 +101,6 @@ class ReferenceTypePropagation::RTPVisitor : public HGraphDelegateVisitor {
bool is_exact);
private:
- Handle<mirror::ClassLoader> class_loader_;
Handle<mirror::DexCache> hint_dex_cache_;
HandleCache* handle_cache_;
ArenaVector<HInstruction*>* worklist_;
@@ -111,13 +108,11 @@ class ReferenceTypePropagation::RTPVisitor : public HGraphDelegateVisitor {
};
ReferenceTypePropagation::ReferenceTypePropagation(HGraph* graph,
- Handle<mirror::ClassLoader> class_loader,
Handle<mirror::DexCache> hint_dex_cache,
VariableSizedHandleScope* handles,
bool is_first_run,
const char* name)
: HOptimization(graph, name),
- class_loader_(class_loader),
hint_dex_cache_(hint_dex_cache),
handle_cache_(handles),
worklist_(graph->GetArena()->Adapter(kArenaAllocReferenceTypePropagation)),
@@ -152,12 +147,7 @@ void ReferenceTypePropagation::ValidateTypes() {
}
void ReferenceTypePropagation::Visit(HInstruction* instruction) {
- RTPVisitor visitor(graph_,
- class_loader_,
- hint_dex_cache_,
- &handle_cache_,
- &worklist_,
- is_first_run_);
+ RTPVisitor visitor(graph_, hint_dex_cache_, &handle_cache_, &worklist_, is_first_run_);
instruction->Accept(&visitor);
}
@@ -331,12 +321,7 @@ void ReferenceTypePropagation::Run() {
}
void ReferenceTypePropagation::VisitBasicBlock(HBasicBlock* block) {
- RTPVisitor visitor(graph_,
- class_loader_,
- hint_dex_cache_,
- &handle_cache_,
- &worklist_,
- is_first_run_);
+ RTPVisitor visitor(graph_, hint_dex_cache_, &handle_cache_, &worklist_, is_first_run_);
// Handle Phis first as there might be instructions in the same block who depend on them.
for (HInstructionIterator it(block->GetPhis()); !it.Done(); it.Advance()) {
VisitPhi(it.Current()->AsPhi());
@@ -557,9 +542,8 @@ void ReferenceTypePropagation::RTPVisitor::UpdateReferenceTypeInfo(HInstruction*
ScopedObjectAccess soa(Thread::Current());
ObjPtr<mirror::DexCache> dex_cache = FindDexCacheWithHint(soa.Self(), dex_file, hint_dex_cache_);
- ObjPtr<mirror::Class> klass =
- ClassLinker::LookupResolvedType(type_idx, dex_cache, class_loader_.Get());
- SetClassAsTypeInfo(instr, klass, is_exact);
+ // Get type from dex cache assuming it was populated by the verifier.
+ SetClassAsTypeInfo(instr, dex_cache->GetResolvedType(type_idx), is_exact);
}
void ReferenceTypePropagation::RTPVisitor::VisitNewInstance(HNewInstance* instr) {
@@ -572,13 +556,25 @@ void ReferenceTypePropagation::RTPVisitor::VisitNewArray(HNewArray* instr) {
SetClassAsTypeInfo(instr, instr->GetLoadClass()->GetClass().Get(), /* is_exact */ true);
}
+static mirror::Class* GetClassFromDexCache(Thread* self,
+ const DexFile& dex_file,
+ dex::TypeIndex type_idx,
+ Handle<mirror::DexCache> hint_dex_cache)
+ REQUIRES_SHARED(Locks::mutator_lock_) {
+ ObjPtr<mirror::DexCache> dex_cache = FindDexCacheWithHint(self, dex_file, hint_dex_cache);
+ // Get type from dex cache assuming it was populated by the verifier.
+ return dex_cache->GetResolvedType(type_idx);
+}
+
void ReferenceTypePropagation::RTPVisitor::VisitParameterValue(HParameterValue* instr) {
// We check if the existing type is valid: the inliner may have set it.
if (instr->GetType() == Primitive::kPrimNot && !instr->GetReferenceTypeInfo().IsValid()) {
- UpdateReferenceTypeInfo(instr,
- instr->GetTypeIndex(),
- instr->GetDexFile(),
- /* is_exact */ false);
+ ScopedObjectAccess soa(Thread::Current());
+ mirror::Class* resolved_class = GetClassFromDexCache(soa.Self(),
+ instr->GetDexFile(),
+ instr->GetTypeIndex(),
+ hint_dex_cache_);
+ SetClassAsTypeInfo(instr, resolved_class, /* is_exact */ false);
}
}
diff --git a/compiler/optimizing/reference_type_propagation.h b/compiler/optimizing/reference_type_propagation.h
index 215e96786b..4663471729 100644
--- a/compiler/optimizing/reference_type_propagation.h
+++ b/compiler/optimizing/reference_type_propagation.h
@@ -33,7 +33,6 @@ namespace art {
class ReferenceTypePropagation : public HOptimization {
public:
ReferenceTypePropagation(HGraph* graph,
- Handle<mirror::ClassLoader> class_loader,
Handle<mirror::DexCache> hint_dex_cache,
VariableSizedHandleScope* handles,
bool is_first_run,
@@ -106,8 +105,6 @@ class ReferenceTypePropagation : public HOptimization {
void ValidateTypes();
- Handle<mirror::ClassLoader> class_loader_;
-
// Note: hint_dex_cache_ is usually, but not necessarily, the dex cache associated with
// graph_->GetDexFile(). Since we may look up also in other dex files, it's used only
// as a hint, to reduce the number of calls to the costly ClassLinker::FindDexCache().
diff --git a/compiler/optimizing/reference_type_propagation_test.cc b/compiler/optimizing/reference_type_propagation_test.cc
index 84a4bab1a9..b061c871b0 100644
--- a/compiler/optimizing/reference_type_propagation_test.cc
+++ b/compiler/optimizing/reference_type_propagation_test.cc
@@ -38,7 +38,6 @@ class ReferenceTypePropagationTest : public CommonCompilerTest {
void SetupPropagation(VariableSizedHandleScope* handles) {
graph_->InitializeInexactObjectRTI(handles);
propagation_ = new (&allocator_) ReferenceTypePropagation(graph_,
- Handle<mirror::ClassLoader>(),
Handle<mirror::DexCache>(),
handles,
true,
diff --git a/compiler/optimizing/sharpening.cc b/compiler/optimizing/sharpening.cc
index f07f02a719..be400925d5 100644
--- a/compiler/optimizing/sharpening.cc
+++ b/compiler/optimizing/sharpening.cc
@@ -163,7 +163,7 @@ HLoadClass::LoadKind HSharpening::SharpenClass(HLoadClass* load_class,
if (!compiler_driver->GetSupportBootImageFixup()) {
// compiler_driver_test. Do not sharpen.
desired_load_kind = HLoadClass::LoadKind::kDexCacheViaMethod;
- } else if ((klass.Get() != nullptr) && compiler_driver->IsImageClass(
+ } else if ((klass != nullptr) && compiler_driver->IsImageClass(
dex_file.StringDataByIdx(dex_file.GetTypeId(type_index).descriptor_idx_))) {
is_in_boot_image = true;
desired_load_kind = codegen->GetCompilerOptions().GetCompilePic()
@@ -175,7 +175,7 @@ HLoadClass::LoadKind HSharpening::SharpenClass(HLoadClass* load_class,
desired_load_kind = HLoadClass::LoadKind::kBssEntry;
}
} else {
- is_in_boot_image = (klass.Get() != nullptr) &&
+ is_in_boot_image = (klass != nullptr) &&
runtime->GetHeap()->ObjectIsInBootImageSpace(klass.Get());
if (runtime->UseJitCompilation()) {
// TODO: Make sure we don't set the "compile PIC" flag for JIT as that's bogus.
@@ -183,7 +183,7 @@ HLoadClass::LoadKind HSharpening::SharpenClass(HLoadClass* load_class,
if (is_in_boot_image) {
// TODO: Use direct pointers for all non-moving spaces, not just boot image. Bug: 29530787
desired_load_kind = HLoadClass::LoadKind::kBootImageAddress;
- } else if (klass.Get() != nullptr) {
+ } else if (klass != nullptr) {
desired_load_kind = HLoadClass::LoadKind::kJitTableAddress;
} else {
// Class not loaded yet. This happens when the dex code requesting
diff --git a/compiler/optimizing/ssa_builder.cc b/compiler/optimizing/ssa_builder.cc
index 50ab11bc23..487e4dd498 100644
--- a/compiler/optimizing/ssa_builder.cc
+++ b/compiler/optimizing/ssa_builder.cc
@@ -499,11 +499,7 @@ GraphAnalysisResult SsaBuilder::BuildSsa() {
// 4) Compute type of reference type instructions. The pass assumes that
// NullConstant has been fixed up.
- ReferenceTypePropagation(graph_,
- class_loader_,
- dex_cache_,
- handles_,
- /* is_first_run */ true).Run();
+ ReferenceTypePropagation(graph_, dex_cache_, handles_, /* is_first_run */ true).Run();
// 5) HInstructionBuilder duplicated ArrayGet instructions with ambiguous type
// (int/float or long/double) and marked ArraySets with ambiguous input type.
diff --git a/compiler/optimizing/ssa_builder.h b/compiler/optimizing/ssa_builder.h
index 978f113ec4..45dac54115 100644
--- a/compiler/optimizing/ssa_builder.h
+++ b/compiler/optimizing/ssa_builder.h
@@ -48,11 +48,9 @@ namespace art {
class SsaBuilder : public ValueObject {
public:
SsaBuilder(HGraph* graph,
- Handle<mirror::ClassLoader> class_loader,
Handle<mirror::DexCache> dex_cache,
VariableSizedHandleScope* handles)
: graph_(graph),
- class_loader_(class_loader),
dex_cache_(dex_cache),
handles_(handles),
agets_fixed_(false),
@@ -117,7 +115,6 @@ class SsaBuilder : public ValueObject {
void RemoveRedundantUninitializedStrings();
HGraph* graph_;
- Handle<mirror::ClassLoader> class_loader_;
Handle<mirror::DexCache> dex_cache_;
VariableSizedHandleScope* const handles_;
diff --git a/compiler/optimizing/stack_map_stream.cc b/compiler/optimizing/stack_map_stream.cc
index f8e01b7537..1bcc8e1ace 100644
--- a/compiler/optimizing/stack_map_stream.cc
+++ b/compiler/optimizing/stack_map_stream.cc
@@ -38,19 +38,14 @@ void StackMapStream::BeginStackMapEntry(uint32_t dex_pc,
current_entry_.native_pc_code_offset = CodeOffset::FromOffset(native_pc_offset, instruction_set_);
current_entry_.register_mask = register_mask;
current_entry_.sp_mask = sp_mask;
- current_entry_.num_dex_registers = num_dex_registers;
current_entry_.inlining_depth = inlining_depth;
- current_entry_.dex_register_locations_start_index = dex_register_locations_.size();
current_entry_.inline_infos_start_index = inline_infos_.size();
- current_entry_.dex_register_map_hash = 0;
- current_entry_.same_dex_register_map_as_ = kNoSameDexMapFound;
current_entry_.stack_mask_index = 0;
- if (num_dex_registers != 0) {
- current_entry_.live_dex_registers_mask =
- ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream);
- } else {
- current_entry_.live_dex_registers_mask = nullptr;
- }
+ current_entry_.dex_register_entry.num_dex_registers = num_dex_registers;
+ current_entry_.dex_register_entry.locations_start_index = dex_register_locations_.size();
+ current_entry_.dex_register_entry.live_dex_registers_mask = (num_dex_registers != 0)
+ ? ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream)
+ : nullptr;
if (sp_mask != nullptr) {
stack_mask_max_ = std::max(stack_mask_max_, sp_mask->GetHighestBitSet());
@@ -65,7 +60,7 @@ void StackMapStream::BeginStackMapEntry(uint32_t dex_pc,
}
void StackMapStream::EndStackMapEntry() {
- current_entry_.same_dex_register_map_as_ = FindEntryWithTheSameDexMap();
+ current_entry_.dex_register_map_index = AddDexRegisterMapEntry(current_entry_.dex_register_entry);
stack_maps_.push_back(current_entry_);
current_entry_ = StackMapEntry();
}
@@ -91,19 +86,15 @@ void StackMapStream::AddDexRegisterEntry(DexRegisterLocation::Kind kind, int32_t
dex_register_locations_.push_back(index);
location_catalog_entries_indices_.Insert(std::make_pair(location, index));
}
-
- if (in_inline_frame_) {
- // TODO: Support sharing DexRegisterMap across InlineInfo.
- DCHECK_LT(current_dex_register_, current_inline_info_.num_dex_registers);
- current_inline_info_.live_dex_registers_mask->SetBit(current_dex_register_);
- } else {
- DCHECK_LT(current_dex_register_, current_entry_.num_dex_registers);
- current_entry_.live_dex_registers_mask->SetBit(current_dex_register_);
- current_entry_.dex_register_map_hash += (1 <<
- (current_dex_register_ % (sizeof(current_entry_.dex_register_map_hash) * kBitsPerByte)));
- current_entry_.dex_register_map_hash += static_cast<uint32_t>(value);
- current_entry_.dex_register_map_hash += static_cast<uint32_t>(kind);
- }
+ DexRegisterMapEntry* const entry = in_inline_frame_
+ ? &current_inline_info_.dex_register_entry
+ : &current_entry_.dex_register_entry;
+ DCHECK_LT(current_dex_register_, entry->num_dex_registers);
+ entry->live_dex_registers_mask->SetBit(current_dex_register_);
+ entry->hash += (1 <<
+ (current_dex_register_ % (sizeof(DexRegisterMapEntry::hash) * kBitsPerByte)));
+ entry->hash += static_cast<uint32_t>(value);
+ entry->hash += static_cast<uint32_t>(kind);
}
current_dex_register_++;
}
@@ -124,20 +115,19 @@ void StackMapStream::BeginInlineInfoEntry(ArtMethod* method,
current_inline_info_.method_index = method->GetDexMethodIndexUnchecked();
}
current_inline_info_.dex_pc = dex_pc;
- current_inline_info_.num_dex_registers = num_dex_registers;
- current_inline_info_.dex_register_locations_start_index = dex_register_locations_.size();
- if (num_dex_registers != 0) {
- current_inline_info_.live_dex_registers_mask =
- ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream);
- } else {
- current_inline_info_.live_dex_registers_mask = nullptr;
- }
+ current_inline_info_.dex_register_entry.num_dex_registers = num_dex_registers;
+ current_inline_info_.dex_register_entry.locations_start_index = dex_register_locations_.size();
+ current_inline_info_.dex_register_entry.live_dex_registers_mask = (num_dex_registers != 0)
+ ? ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream)
+ : nullptr;
current_dex_register_ = 0;
}
void StackMapStream::EndInlineInfoEntry() {
+ current_inline_info_.dex_register_map_index =
+ AddDexRegisterMapEntry(current_inline_info_.dex_register_entry);
DCHECK(in_inline_frame_);
- DCHECK_EQ(current_dex_register_, current_inline_info_.num_dex_registers)
+ DCHECK_EQ(current_dex_register_, current_inline_info_.dex_register_entry.num_dex_registers)
<< "Inline information contains less registers than expected";
in_inline_frame_ = false;
inline_infos_.push_back(current_inline_info_);
@@ -193,8 +183,7 @@ size_t StackMapStream::ComputeDexRegisterLocationCatalogSize() const {
return size;
}
-size_t StackMapStream::ComputeDexRegisterMapSize(uint32_t num_dex_registers,
- const BitVector* live_dex_registers_mask) const {
+size_t StackMapStream::DexRegisterMapEntry::ComputeSize(size_t catalog_size) const {
// For num_dex_registers == 0u live_dex_registers_mask may be null.
if (num_dex_registers == 0u) {
return 0u; // No register map will be emitted.
@@ -208,8 +197,7 @@ size_t StackMapStream::ComputeDexRegisterMapSize(uint32_t num_dex_registers,
// Compute the size of the set of live Dex register entries.
size_t number_of_live_dex_registers = live_dex_registers_mask->NumSetBits();
size_t map_entries_size_in_bits =
- DexRegisterMap::SingleEntrySizeInBits(location_catalog_entries_.size())
- * number_of_live_dex_registers;
+ DexRegisterMap::SingleEntrySizeInBits(catalog_size) * number_of_live_dex_registers;
size_t map_entries_size_in_bytes =
RoundUp(map_entries_size_in_bits, kBitsPerByte) / kBitsPerByte;
size += map_entries_size_in_bytes;
@@ -218,18 +206,8 @@ size_t StackMapStream::ComputeDexRegisterMapSize(uint32_t num_dex_registers,
size_t StackMapStream::ComputeDexRegisterMapsSize() const {
size_t size = 0;
- size_t inline_info_index = 0;
- for (const StackMapEntry& entry : stack_maps_) {
- if (entry.same_dex_register_map_as_ == kNoSameDexMapFound) {
- size += ComputeDexRegisterMapSize(entry.num_dex_registers, entry.live_dex_registers_mask);
- } else {
- // Entries with the same dex map will have the same offset.
- }
- for (size_t j = 0; j < entry.inlining_depth; ++j) {
- InlineInfoEntry inline_entry = inline_infos_[inline_info_index++];
- size += ComputeDexRegisterMapSize(inline_entry.num_dex_registers,
- inline_entry.live_dex_registers_mask);
- }
+ for (const DexRegisterMapEntry& entry : dex_register_entries_) {
+ size += entry.ComputeSize(location_catalog_entries_.size());
}
return size;
}
@@ -264,6 +242,30 @@ void StackMapStream::ComputeInlineInfoEncoding(InlineInfoEncoding* encoding,
encoding->SetFromSizes(method_index_max, dex_pc_max, extra_data_max, dex_register_maps_bytes);
}
+size_t StackMapStream::MaybeCopyDexRegisterMap(DexRegisterMapEntry& entry,
+ size_t* current_offset,
+ MemoryRegion dex_register_locations_region) {
+ DCHECK(current_offset != nullptr);
+ if ((entry.num_dex_registers == 0) || (entry.live_dex_registers_mask->NumSetBits() == 0)) {
+ // No dex register map needed.
+ return StackMap::kNoDexRegisterMap;
+ }
+ if (entry.offset == DexRegisterMapEntry::kOffsetUnassigned) {
+ // Not already copied, need to copy and and assign an offset.
+ entry.offset = *current_offset;
+ const size_t entry_size = entry.ComputeSize(location_catalog_entries_.size());
+ DexRegisterMap dex_register_map(
+ dex_register_locations_region.Subregion(entry.offset, entry_size));
+ *current_offset += entry_size;
+ // Fill in the map since it was just added.
+ FillInDexRegisterMap(dex_register_map,
+ entry.num_dex_registers,
+ *entry.live_dex_registers_mask,
+ entry.locations_start_index);
+ }
+ return entry.offset;
+}
+
void StackMapStream::FillIn(MemoryRegion region) {
DCHECK_EQ(0u, current_entry_.dex_pc) << "EndStackMapEntry not called after BeginStackMapEntry";
DCHECK_NE(0u, needed_size_) << "PrepareForFillIn not called before FillIn";
@@ -311,35 +313,10 @@ void StackMapStream::FillIn(MemoryRegion region) {
stack_map.SetRegisterMaskIndex(encoding.stack_map.encoding, entry.register_mask_index);
stack_map.SetStackMaskIndex(encoding.stack_map.encoding, entry.stack_mask_index);
- if (entry.num_dex_registers == 0 || (entry.live_dex_registers_mask->NumSetBits() == 0)) {
- // No dex map available.
- stack_map.SetDexRegisterMapOffset(encoding.stack_map.encoding, StackMap::kNoDexRegisterMap);
- } else {
- // Search for an entry with the same dex map.
- if (entry.same_dex_register_map_as_ != kNoSameDexMapFound) {
- // If we have a hit reuse the offset.
- stack_map.SetDexRegisterMapOffset(
- encoding.stack_map.encoding,
- code_info.GetStackMapAt(entry.same_dex_register_map_as_, encoding)
- .GetDexRegisterMapOffset(encoding.stack_map.encoding));
- } else {
- // New dex registers maps should be added to the stack map.
- MemoryRegion register_region = dex_register_locations_region.Subregion(
- next_dex_register_map_offset,
- ComputeDexRegisterMapSize(entry.num_dex_registers, entry.live_dex_registers_mask));
- next_dex_register_map_offset += register_region.size();
- DexRegisterMap dex_register_map(register_region);
- stack_map.SetDexRegisterMapOffset(
- encoding.stack_map.encoding,
- register_region.begin() - dex_register_locations_region.begin());
-
- // Set the dex register location.
- FillInDexRegisterMap(dex_register_map,
- entry.num_dex_registers,
- *entry.live_dex_registers_mask,
- entry.dex_register_locations_start_index);
- }
- }
+ size_t offset = MaybeCopyDexRegisterMap(dex_register_entries_[entry.dex_register_map_index],
+ &next_dex_register_map_offset,
+ dex_register_locations_region);
+ stack_map.SetDexRegisterMapOffset(encoding.stack_map.encoding, offset);
// Set the inlining info.
if (entry.inlining_depth != 0) {
@@ -371,29 +348,13 @@ void StackMapStream::FillIn(MemoryRegion region) {
inline_info.SetExtraDataAtDepth(encoding.inline_info.encoding, depth, 1);
}
inline_info.SetDexPcAtDepth(encoding.inline_info.encoding, depth, inline_entry.dex_pc);
- if (inline_entry.num_dex_registers == 0) {
- // No dex map available.
- inline_info.SetDexRegisterMapOffsetAtDepth(encoding.inline_info.encoding,
- depth,
- StackMap::kNoDexRegisterMap);
- DCHECK(inline_entry.live_dex_registers_mask == nullptr);
- } else {
- MemoryRegion register_region = dex_register_locations_region.Subregion(
- next_dex_register_map_offset,
- ComputeDexRegisterMapSize(inline_entry.num_dex_registers,
- inline_entry.live_dex_registers_mask));
- next_dex_register_map_offset += register_region.size();
- DexRegisterMap dex_register_map(register_region);
- inline_info.SetDexRegisterMapOffsetAtDepth(
- encoding.inline_info.encoding,
- depth,
- register_region.begin() - dex_register_locations_region.begin());
-
- FillInDexRegisterMap(dex_register_map,
- inline_entry.num_dex_registers,
- *inline_entry.live_dex_registers_mask,
- inline_entry.dex_register_locations_start_index);
- }
+ size_t dex_register_map_offset = MaybeCopyDexRegisterMap(
+ dex_register_entries_[inline_entry.dex_register_map_index],
+ &next_dex_register_map_offset,
+ dex_register_locations_region);
+ inline_info.SetDexRegisterMapOffsetAtDepth(encoding.inline_info.encoding,
+ depth,
+ dex_register_map_offset);
}
} else if (encoding.stack_map.encoding.GetInlineInfoEncoding().BitSize() > 0) {
stack_map.SetInlineInfoIndex(encoding.stack_map.encoding, StackMap::kNoInlineInfo);
@@ -448,34 +409,31 @@ void StackMapStream::FillInDexRegisterMap(DexRegisterMap dex_register_map,
}
}
-size_t StackMapStream::FindEntryWithTheSameDexMap() {
- size_t current_entry_index = stack_maps_.size();
- auto entries_it = dex_map_hash_to_stack_map_indices_.find(current_entry_.dex_register_map_hash);
+size_t StackMapStream::AddDexRegisterMapEntry(const DexRegisterMapEntry& entry) {
+ const size_t current_entry_index = dex_register_entries_.size();
+ auto entries_it = dex_map_hash_to_stack_map_indices_.find(entry.hash);
if (entries_it == dex_map_hash_to_stack_map_indices_.end()) {
// We don't have a perfect hash functions so we need a list to collect all stack maps
// which might have the same dex register map.
ArenaVector<uint32_t> stack_map_indices(allocator_->Adapter(kArenaAllocStackMapStream));
stack_map_indices.push_back(current_entry_index);
- dex_map_hash_to_stack_map_indices_.Put(current_entry_.dex_register_map_hash,
- std::move(stack_map_indices));
- return kNoSameDexMapFound;
- }
-
- // We might have collisions, so we need to check whether or not we really have a match.
- for (uint32_t test_entry_index : entries_it->second) {
- if (HaveTheSameDexMaps(GetStackMap(test_entry_index), current_entry_)) {
- return test_entry_index;
+ dex_map_hash_to_stack_map_indices_.Put(entry.hash, std::move(stack_map_indices));
+ } else {
+ // We might have collisions, so we need to check whether or not we really have a match.
+ for (uint32_t test_entry_index : entries_it->second) {
+ if (DexRegisterMapEntryEquals(dex_register_entries_[test_entry_index], entry)) {
+ return test_entry_index;
+ }
}
+ entries_it->second.push_back(current_entry_index);
}
- entries_it->second.push_back(current_entry_index);
- return kNoSameDexMapFound;
+ dex_register_entries_.push_back(entry);
+ return current_entry_index;
}
-bool StackMapStream::HaveTheSameDexMaps(const StackMapEntry& a, const StackMapEntry& b) const {
- if (a.live_dex_registers_mask == nullptr && b.live_dex_registers_mask == nullptr) {
- return true;
- }
- if (a.live_dex_registers_mask == nullptr || b.live_dex_registers_mask == nullptr) {
+bool StackMapStream::DexRegisterMapEntryEquals(const DexRegisterMapEntry& a,
+ const DexRegisterMapEntry& b) const {
+ if ((a.live_dex_registers_mask == nullptr) != (b.live_dex_registers_mask == nullptr)) {
return false;
}
if (a.num_dex_registers != b.num_dex_registers) {
@@ -489,12 +447,12 @@ bool StackMapStream::HaveTheSameDexMaps(const StackMapEntry& a, const StackMapEn
}
size_t number_of_live_dex_registers = a.live_dex_registers_mask->NumSetBits();
DCHECK_LE(number_of_live_dex_registers, dex_register_locations_.size());
- DCHECK_LE(a.dex_register_locations_start_index,
+ DCHECK_LE(a.locations_start_index,
dex_register_locations_.size() - number_of_live_dex_registers);
- DCHECK_LE(b.dex_register_locations_start_index,
+ DCHECK_LE(b.locations_start_index,
dex_register_locations_.size() - number_of_live_dex_registers);
- auto a_begin = dex_register_locations_.begin() + a.dex_register_locations_start_index;
- auto b_begin = dex_register_locations_.begin() + b.dex_register_locations_start_index;
+ auto a_begin = dex_register_locations_.begin() + a.locations_start_index;
+ auto b_begin = dex_register_locations_.begin() + b.locations_start_index;
if (!std::equal(a_begin, a_begin + number_of_live_dex_registers, b_begin)) {
return false;
}
@@ -597,10 +555,10 @@ void StackMapStream::CheckCodeInfo(MemoryRegion region) const {
CheckDexRegisterMap(code_info,
code_info.GetDexRegisterMapOf(
- stack_map, encoding, entry.num_dex_registers),
- entry.num_dex_registers,
- entry.live_dex_registers_mask,
- entry.dex_register_locations_start_index);
+ stack_map, encoding, entry.dex_register_entry.num_dex_registers),
+ entry.dex_register_entry.num_dex_registers,
+ entry.dex_register_entry.live_dex_registers_mask,
+ entry.dex_register_entry.locations_start_index);
// Check inline info.
DCHECK_EQ(stack_map.HasInlineInfo(stack_map_encoding), (entry.inlining_depth != 0));
@@ -623,10 +581,13 @@ void StackMapStream::CheckCodeInfo(MemoryRegion region) const {
CheckDexRegisterMap(code_info,
code_info.GetDexRegisterMapAtDepth(
- d, inline_info, encoding, inline_entry.num_dex_registers),
- inline_entry.num_dex_registers,
- inline_entry.live_dex_registers_mask,
- inline_entry.dex_register_locations_start_index);
+ d,
+ inline_info,
+ encoding,
+ inline_entry.dex_register_entry.num_dex_registers),
+ inline_entry.dex_register_entry.num_dex_registers,
+ inline_entry.dex_register_entry.live_dex_registers_mask,
+ inline_entry.dex_register_entry.locations_start_index);
}
}
}
diff --git a/compiler/optimizing/stack_map_stream.h b/compiler/optimizing/stack_map_stream.h
index 08c1d3e3c0..bba3d51e62 100644
--- a/compiler/optimizing/stack_map_stream.h
+++ b/compiler/optimizing/stack_map_stream.h
@@ -70,6 +70,7 @@ class StackMapStream : public ValueObject {
inline_infos_(allocator->Adapter(kArenaAllocStackMapStream)),
stack_masks_(allocator->Adapter(kArenaAllocStackMapStream)),
register_masks_(allocator->Adapter(kArenaAllocStackMapStream)),
+ dex_register_entries_(allocator->Adapter(kArenaAllocStackMapStream)),
stack_mask_max_(-1),
dex_pc_max_(0),
register_mask_max_(0),
@@ -89,30 +90,42 @@ class StackMapStream : public ValueObject {
code_info_encoding_.reserve(16);
}
+ // A dex register map entry for a single stack map entry, contains what registers are live as
+ // well as indices into the location catalog.
+ class DexRegisterMapEntry {
+ public:
+ static const size_t kOffsetUnassigned = -1;
+
+ BitVector* live_dex_registers_mask;
+ uint32_t num_dex_registers;
+ size_t locations_start_index;
+ // Computed fields
+ size_t hash = 0;
+ size_t offset = kOffsetUnassigned;
+
+ size_t ComputeSize(size_t catalog_size) const;
+ };
+
// See runtime/stack_map.h to know what these fields contain.
struct StackMapEntry {
uint32_t dex_pc;
CodeOffset native_pc_code_offset;
uint32_t register_mask;
BitVector* sp_mask;
- uint32_t num_dex_registers;
uint8_t inlining_depth;
- size_t dex_register_locations_start_index;
size_t inline_infos_start_index;
- BitVector* live_dex_registers_mask;
- uint32_t dex_register_map_hash;
- size_t same_dex_register_map_as_;
uint32_t stack_mask_index;
uint32_t register_mask_index;
+ DexRegisterMapEntry dex_register_entry;
+ size_t dex_register_map_index;
};
struct InlineInfoEntry {
uint32_t dex_pc; // DexFile::kDexNoIndex for intrinsified native methods.
ArtMethod* method;
uint32_t method_index;
- uint32_t num_dex_registers;
- BitVector* live_dex_registers_mask;
- size_t dex_register_locations_start_index;
+ DexRegisterMapEntry dex_register_entry;
+ size_t dex_register_map_index;
};
void BeginStackMapEntry(uint32_t dex_pc,
@@ -140,7 +153,8 @@ class StackMapStream : public ValueObject {
}
void SetStackMapNativePcOffset(size_t i, uint32_t native_pc_offset) {
- stack_maps_[i].native_pc_code_offset = CodeOffset::FromOffset(native_pc_offset, instruction_set_);
+ stack_maps_[i].native_pc_code_offset =
+ CodeOffset::FromOffset(native_pc_offset, instruction_set_);
}
// Prepares the stream to fill in a memory region. Must be called before FillIn.
@@ -150,8 +164,6 @@ class StackMapStream : public ValueObject {
private:
size_t ComputeDexRegisterLocationCatalogSize() const;
- size_t ComputeDexRegisterMapSize(uint32_t num_dex_registers,
- const BitVector* live_dex_registers_mask) const;
size_t ComputeDexRegisterMapsSize() const;
void ComputeInlineInfoEncoding(InlineInfoEncoding* encoding,
size_t dex_register_maps_bytes);
@@ -164,15 +176,24 @@ class StackMapStream : public ValueObject {
// Returns the number of unique register masks.
size_t PrepareRegisterMasks();
- // Returns the index of an entry with the same dex register map as the current_entry,
- // or kNoSameDexMapFound if no such entry exists.
- size_t FindEntryWithTheSameDexMap();
- bool HaveTheSameDexMaps(const StackMapEntry& a, const StackMapEntry& b) const;
+ // Deduplicate entry if possible and return the corresponding index into dex_register_entries_
+ // array. If entry is not a duplicate, a new entry is added to dex_register_entries_.
+ size_t AddDexRegisterMapEntry(const DexRegisterMapEntry& entry);
+
+ // Return true if the two dex register map entries are equal.
+ bool DexRegisterMapEntryEquals(const DexRegisterMapEntry& a, const DexRegisterMapEntry& b) const;
+
+ // Fill in the corresponding entries of a register map.
void FillInDexRegisterMap(DexRegisterMap dex_register_map,
uint32_t num_dex_registers,
const BitVector& live_dex_registers_mask,
uint32_t start_index_in_dex_register_locations) const;
+ // Returns the offset for the dex register inside of the dex register location region. See FillIn.
+ // Only copies the dex register map if the offset for the entry is not already assigned.
+ size_t MaybeCopyDexRegisterMap(DexRegisterMapEntry& entry,
+ size_t* current_offset,
+ MemoryRegion dex_register_locations_region);
void CheckDexRegisterMap(const CodeInfo& code_info,
const DexRegisterMap& dex_register_map,
size_t num_dex_registers,
@@ -199,6 +220,7 @@ class StackMapStream : public ValueObject {
ArenaVector<InlineInfoEntry> inline_infos_;
ArenaVector<uint8_t> stack_masks_;
ArenaVector<uint32_t> register_masks_;
+ ArenaVector<DexRegisterMapEntry> dex_register_entries_;
int stack_mask_max_;
uint32_t dex_pc_max_;
uint32_t register_mask_max_;
diff --git a/compiler/optimizing/stack_map_test.cc b/compiler/optimizing/stack_map_test.cc
index bd0aa6dea7..041695187b 100644
--- a/compiler/optimizing/stack_map_test.cc
+++ b/compiler/optimizing/stack_map_test.cc
@@ -410,6 +410,100 @@ TEST(StackMapTest, Test2) {
}
}
+TEST(StackMapTest, TestDeduplicateInlineInfoDexRegisterMap) {
+ ArenaPool pool;
+ ArenaAllocator arena(&pool);
+ StackMapStream stream(&arena, kRuntimeISA);
+ ArtMethod art_method;
+
+ ArenaBitVector sp_mask1(&arena, 0, true);
+ sp_mask1.SetBit(2);
+ sp_mask1.SetBit(4);
+ const size_t number_of_dex_registers = 2;
+ const size_t number_of_dex_registers_in_inline_info = 2;
+ stream.BeginStackMapEntry(0, 64, 0x3, &sp_mask1, number_of_dex_registers, 1);
+ stream.AddDexRegisterEntry(Kind::kInStack, 0); // Short location.
+ stream.AddDexRegisterEntry(Kind::kConstant, -2); // Large location.
+ stream.BeginInlineInfoEntry(&art_method, 3, number_of_dex_registers_in_inline_info);
+ stream.AddDexRegisterEntry(Kind::kInStack, 0); // Short location.
+ stream.AddDexRegisterEntry(Kind::kConstant, -2); // Large location.
+ stream.EndInlineInfoEntry();
+ stream.EndStackMapEntry();
+
+ size_t size = stream.PrepareForFillIn();
+ void* memory = arena.Alloc(size, kArenaAllocMisc);
+ MemoryRegion region(memory, size);
+ stream.FillIn(region);
+
+ CodeInfo code_info(region);
+ CodeInfoEncoding encoding = code_info.ExtractEncoding();
+ ASSERT_EQ(1u, code_info.GetNumberOfStackMaps(encoding));
+
+ uint32_t number_of_catalog_entries = code_info.GetNumberOfLocationCatalogEntries(encoding);
+ ASSERT_EQ(2u, number_of_catalog_entries);
+ DexRegisterLocationCatalog location_catalog = code_info.GetDexRegisterLocationCatalog(encoding);
+ // The Dex register location catalog contains:
+ // - one 1-byte short Dex register locations, and
+ // - one 5-byte large Dex register location.
+ const size_t expected_location_catalog_size = 1u + 5u;
+ ASSERT_EQ(expected_location_catalog_size, location_catalog.Size());
+
+ // First stack map.
+ {
+ StackMap stack_map = code_info.GetStackMapAt(0, encoding);
+ ASSERT_TRUE(stack_map.Equals(code_info.GetStackMapForDexPc(0, encoding)));
+ ASSERT_TRUE(stack_map.Equals(code_info.GetStackMapForNativePcOffset(64, encoding)));
+ ASSERT_EQ(0u, stack_map.GetDexPc(encoding.stack_map.encoding));
+ ASSERT_EQ(64u, stack_map.GetNativePcOffset(encoding.stack_map.encoding, kRuntimeISA));
+ ASSERT_EQ(0x3u, code_info.GetRegisterMaskOf(encoding, stack_map));
+
+ ASSERT_TRUE(CheckStackMask(code_info, encoding, stack_map, sp_mask1));
+
+ ASSERT_TRUE(stack_map.HasDexRegisterMap(encoding.stack_map.encoding));
+ DexRegisterMap map(code_info.GetDexRegisterMapOf(stack_map, encoding, number_of_dex_registers));
+ ASSERT_TRUE(map.IsDexRegisterLive(0));
+ ASSERT_TRUE(map.IsDexRegisterLive(1));
+ ASSERT_EQ(2u, map.GetNumberOfLiveDexRegisters(number_of_dex_registers));
+ // The Dex register map contains:
+ // - one 1-byte live bit mask, and
+ // - one 1-byte set of location catalog entry indices composed of two 2-bit values.
+ size_t expected_map_size = 1u + 1u;
+ ASSERT_EQ(expected_map_size, map.Size());
+
+ ASSERT_EQ(Kind::kInStack, map.GetLocationKind(0, number_of_dex_registers, code_info, encoding));
+ ASSERT_EQ(Kind::kConstant,
+ map.GetLocationKind(1, number_of_dex_registers, code_info, encoding));
+ ASSERT_EQ(Kind::kInStack,
+ map.GetLocationInternalKind(0, number_of_dex_registers, code_info, encoding));
+ ASSERT_EQ(Kind::kConstantLargeValue,
+ map.GetLocationInternalKind(1, number_of_dex_registers, code_info, encoding));
+ ASSERT_EQ(0, map.GetStackOffsetInBytes(0, number_of_dex_registers, code_info, encoding));
+ ASSERT_EQ(-2, map.GetConstant(1, number_of_dex_registers, code_info, encoding));
+
+ const size_t index0 =
+ map.GetLocationCatalogEntryIndex(0, number_of_dex_registers, number_of_catalog_entries);
+ const size_t index1 =
+ map.GetLocationCatalogEntryIndex(1, number_of_dex_registers, number_of_catalog_entries);
+ ASSERT_EQ(0u, index0);
+ ASSERT_EQ(1u, index1);
+ DexRegisterLocation location0 = location_catalog.GetDexRegisterLocation(index0);
+ DexRegisterLocation location1 = location_catalog.GetDexRegisterLocation(index1);
+ ASSERT_EQ(Kind::kInStack, location0.GetKind());
+ ASSERT_EQ(Kind::kConstant, location1.GetKind());
+ ASSERT_EQ(Kind::kInStack, location0.GetInternalKind());
+ ASSERT_EQ(Kind::kConstantLargeValue, location1.GetInternalKind());
+ ASSERT_EQ(0, location0.GetValue());
+ ASSERT_EQ(-2, location1.GetValue());
+
+ // Test that the inline info dex register map deduplicated to the same offset as the stack map
+ // one.
+ ASSERT_TRUE(stack_map.HasInlineInfo(encoding.stack_map.encoding));
+ InlineInfo inline_info = code_info.GetInlineInfoOf(stack_map, encoding);
+ EXPECT_EQ(inline_info.GetDexRegisterMapOffsetAtDepth(encoding.inline_info.encoding, 0),
+ stack_map.GetDexRegisterMapOffset(encoding.stack_map.encoding));
+ }
+}
+
TEST(StackMapTest, TestNonLiveDexRegisters) {
ArenaPool pool;
ArenaAllocator arena(&pool);
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2025-11-02 19:51:24 +0000