Inlining a few small methods based on profiling dex2oat with perf.
Test: m test-art-host
Change-Id: I6313158e59592d8d132154523be9c82dda3c7eb8
diff --git a/compiler/optimizing/instruction_builder.cc b/compiler/optimizing/instruction_builder.cc
index c60f6e5..7dc2fa8 100644
--- a/compiler/optimizing/instruction_builder.cc
+++ b/compiler/optimizing/instruction_builder.cc
@@ -37,37 +37,45 @@
return block_builder_->GetBlockAt(dex_pc);
}
-ArenaVector<HInstruction*>* HInstructionBuilder::GetLocalsFor(HBasicBlock* block) {
+inline ArenaVector<HInstruction*>* HInstructionBuilder::GetLocalsFor(HBasicBlock* block) {
ArenaVector<HInstruction*>* locals = &locals_for_[block->GetBlockId()];
const size_t vregs = graph_->GetNumberOfVRegs();
- if (locals->size() != vregs) {
- locals->resize(vregs, nullptr);
+ if (locals->size() == vregs) {
+ return locals;
+ }
+ return GetLocalsForWithAllocation(block, locals, vregs);
+}
- if (block->IsCatchBlock()) {
- // We record incoming inputs of catch phis at throwing instructions and
- // must therefore eagerly create the phis. Phis for undefined vregs will
- // be deleted when the first throwing instruction with the vreg undefined
- // is encountered. Unused phis will be removed by dead phi analysis.
- for (size_t i = 0; i < vregs; ++i) {
- // No point in creating the catch phi if it is already undefined at
- // the first throwing instruction.
- HInstruction* current_local_value = (*current_locals_)[i];
- if (current_local_value != nullptr) {
- HPhi* phi = new (arena_) HPhi(
- arena_,
- i,
- 0,
- current_local_value->GetType());
- block->AddPhi(phi);
- (*locals)[i] = phi;
- }
+ArenaVector<HInstruction*>* HInstructionBuilder::GetLocalsForWithAllocation(
+ HBasicBlock* block,
+ ArenaVector<HInstruction*>* locals,
+ const size_t vregs) {
+ DCHECK_NE(locals->size(), vregs);
+ locals->resize(vregs, nullptr);
+ if (block->IsCatchBlock()) {
+ // We record incoming inputs of catch phis at throwing instructions and
+ // must therefore eagerly create the phis. Phis for undefined vregs will
+ // be deleted when the first throwing instruction with the vreg undefined
+ // is encountered. Unused phis will be removed by dead phi analysis.
+ for (size_t i = 0; i < vregs; ++i) {
+ // No point in creating the catch phi if it is already undefined at
+ // the first throwing instruction.
+ HInstruction* current_local_value = (*current_locals_)[i];
+ if (current_local_value != nullptr) {
+ HPhi* phi = new (arena_) HPhi(
+ arena_,
+ i,
+ 0,
+ current_local_value->GetType());
+ block->AddPhi(phi);
+ (*locals)[i] = phi;
}
}
}
return locals;
}
-HInstruction* HInstructionBuilder::ValueOfLocalAt(HBasicBlock* block, size_t local) {
+inline HInstruction* HInstructionBuilder::ValueOfLocalAt(HBasicBlock* block, size_t local) {
ArenaVector<HInstruction*>* locals = GetLocalsFor(block);
return (*locals)[local];
}
diff --git a/compiler/optimizing/instruction_builder.h b/compiler/optimizing/instruction_builder.h
index e735a0c..7fdc188 100644
--- a/compiler/optimizing/instruction_builder.h
+++ b/compiler/optimizing/instruction_builder.h
@@ -93,6 +93,10 @@
HBasicBlock* FindBlockStartingAt(uint32_t dex_pc) const;
ArenaVector<HInstruction*>* GetLocalsFor(HBasicBlock* block);
+ // Out of line version of GetLocalsFor(), which has a fast path that is
+ // beneficial to get inlined by callers.
+ ArenaVector<HInstruction*>* GetLocalsForWithAllocation(
+ HBasicBlock* block, ArenaVector<HInstruction*>* locals, const size_t vregs);
HInstruction* ValueOfLocalAt(HBasicBlock* block, size_t local);
HInstruction* LoadLocal(uint32_t register_index, Primitive::Type type) const;
HInstruction* LoadNullCheckedLocal(uint32_t register_index, uint32_t dex_pc);
diff --git a/compiler/optimizing/loop_optimization.cc b/compiler/optimizing/loop_optimization.cc
index 26c9ab8..bffda93 100644
--- a/compiler/optimizing/loop_optimization.cc
+++ b/compiler/optimizing/loop_optimization.cc
@@ -69,7 +69,7 @@
}
void HLoopOptimization::Run() {
- // Well-behaved loops only.
+ // Skip if there is no loop or the graph has try-catch/irreducible loops.
// TODO: make this less of a sledgehammer.
if (!graph_->HasLoops() || graph_->HasTryCatch() || graph_->HasIrreducibleLoops()) {
return;
@@ -85,6 +85,7 @@
LocalRun();
if (top_loop_ == nullptr) {
+ // All loops have been eliminated.
graph_->SetHasLoops(false);
}
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index c39aed2..11a0d35 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1734,11 +1734,11 @@
// A HEnvironment object contains the values of virtual registers at a given location.
class HEnvironment : public ArenaObject<kArenaAllocEnvironment> {
public:
- HEnvironment(ArenaAllocator* arena,
- size_t number_of_vregs,
- ArtMethod* method,
- uint32_t dex_pc,
- HInstruction* holder)
+ ALWAYS_INLINE HEnvironment(ArenaAllocator* arena,
+ size_t number_of_vregs,
+ ArtMethod* method,
+ uint32_t dex_pc,
+ HInstruction* holder)
: vregs_(number_of_vregs, arena->Adapter(kArenaAllocEnvironmentVRegs)),
locations_(number_of_vregs, arena->Adapter(kArenaAllocEnvironmentLocations)),
parent_(nullptr),
@@ -1747,7 +1747,7 @@
holder_(holder) {
}
- HEnvironment(ArenaAllocator* arena, const HEnvironment& to_copy, HInstruction* holder)
+ ALWAYS_INLINE HEnvironment(ArenaAllocator* arena, const HEnvironment& to_copy, HInstruction* holder)
: HEnvironment(arena,
to_copy.Size(),
to_copy.GetMethod(),
diff --git a/compiler/optimizing/ssa_liveness_analysis.h b/compiler/optimizing/ssa_liveness_analysis.h
index b62bf4e..a239bd5 100644
--- a/compiler/optimizing/ssa_liveness_analysis.h
+++ b/compiler/optimizing/ssa_liveness_analysis.h
@@ -331,7 +331,7 @@
instruction, /* environment */ nullptr, input_index, block->GetLifetimeEnd(), first_use_);
}
- void AddRange(size_t start, size_t end) {
+ ALWAYS_INLINE void AddRange(size_t start, size_t end) {
if (first_range_ == nullptr) {
first_range_ = last_range_ = range_search_start_ =
new (allocator_) LiveRange(start, end, first_range_);
diff --git a/runtime/base/mutex.h b/runtime/base/mutex.h
index c59664b..65d35ea 100644
--- a/runtime/base/mutex.h
+++ b/runtime/base/mutex.h
@@ -516,12 +516,12 @@
// construction and releases it upon destruction.
class SCOPED_CAPABILITY ReaderMutexLock {
public:
- ReaderMutexLock(Thread* self, ReaderWriterMutex& mu) ACQUIRE(mu) :
+ ReaderMutexLock(Thread* self, ReaderWriterMutex& mu) ACQUIRE(mu) ALWAYS_INLINE :
self_(self), mu_(mu) {
mu_.SharedLock(self_);
}
- ~ReaderMutexLock() RELEASE() {
+ ~ReaderMutexLock() RELEASE() ALWAYS_INLINE {
mu_.SharedUnlock(self_);
}
diff --git a/runtime/scoped_thread_state_change.h b/runtime/scoped_thread_state_change.h
index a3286ac..5f03741 100644
--- a/runtime/scoped_thread_state_change.h
+++ b/runtime/scoped_thread_state_change.h
@@ -141,6 +141,8 @@
ALWAYS_INLINE explicit ScopedObjectAccessUnchecked(Thread* self)
REQUIRES(!Locks::thread_suspend_count_lock_);
+ ALWAYS_INLINE ~ScopedObjectAccessUnchecked() REQUIRES(!Locks::thread_suspend_count_lock_) {}
+
// Used when we want a scoped JNI thread state but have no thread/JNIEnv. Consequently doesn't
// change into Runnable or acquire a share on the mutator_lock_.
explicit ScopedObjectAccessUnchecked(JavaVM* vm) ALWAYS_INLINE
diff --git a/runtime/stack_map.h b/runtime/stack_map.h
index 67f0b57..d936ce9 100644
--- a/runtime/stack_map.h
+++ b/runtime/stack_map.h
@@ -571,7 +571,7 @@
}
}
- bool IsDexRegisterLive(uint16_t dex_register_number) const {
+ ALWAYS_INLINE bool IsDexRegisterLive(uint16_t dex_register_number) const {
size_t live_bit_mask_offset_in_bits = GetLiveBitMaskOffset() * kBitsPerByte;
return region_.LoadBit(live_bit_mask_offset_in_bits + dex_register_number);
}
