Merge "Enable allocation elimination as part of LSE"

3 files changed, 25 insertions, 23 deletions

diff --git a/compiler/optimizing/instruction_builder.cc b/compiler/optimizing/instruction_builder.cc
index f5e49c2235..12cb826395 100644
--- a/compiler/optimizing/instruction_builder.cc
+++ b/compiler/optimizing/instruction_builder.cc
@@ -897,12 +897,12 @@ bool HInstructionBuilder::BuildNewInstance(uint16_t type_index, uint32_t dex_pc)
   Handle<mirror::DexCache> outer_dex_cache = outer_compilation_unit_->GetDexCache();
 
   bool finalizable;
-  bool can_throw = NeedsAccessCheck(type_index, dex_cache, &finalizable);
+  bool needs_access_check = NeedsAccessCheck(type_index, dex_cache, &finalizable);
 
   // Only the non-resolved entrypoint handles the finalizable class case. If we
   // need access checks, then we haven't resolved the method and the class may
   // again be finalizable.
-  QuickEntrypointEnum entrypoint = (finalizable || can_throw)
+  QuickEntrypointEnum entrypoint = (finalizable || needs_access_check)
       ? kQuickAllocObject
       : kQuickAllocObjectInitialized;
 
@@ -917,7 +917,7 @@ bool HInstructionBuilder::BuildNewInstance(uint16_t type_index, uint32_t dex_pc)
       outer_dex_file,
       IsOutermostCompilingClass(type_index),
       dex_pc,
-      /*needs_access_check*/ can_throw,
+      needs_access_check,
       compiler_driver_->CanAssumeTypeIsPresentInDexCache(outer_dex_cache, type_index));
 
   AppendInstruction(load_class);
@@ -933,7 +933,7 @@ bool HInstructionBuilder::BuildNewInstance(uint16_t type_index, uint32_t dex_pc)
       dex_pc,
       type_index,
       *dex_compilation_unit_->GetDexFile(),
-      can_throw,
+      needs_access_check,
       finalizable,
       entrypoint));
   return true;
diff --git a/compiler/optimizing/load_store_elimination.cc b/compiler/optimizing/load_store_elimination.cc
index e1977b1798..ac7ed86d1d 100644
--- a/compiler/optimizing/load_store_elimination.cc
+++ b/compiler/optimizing/load_store_elimination.cc
@@ -480,7 +480,7 @@ class HeapLocationCollector : public HGraphVisitor {
                             // alias analysis and won't be as effective.
   bool has_volatile_;       // If there are volatile field accesses.
   bool has_monitor_operations_;    // If there are monitor operations.
-  bool may_deoptimize_;
+  bool may_deoptimize_;     // Only true for HDeoptimize with single-frame deoptimization.
 
   DISALLOW_COPY_AND_ASSIGN(HeapLocationCollector);
 };
@@ -551,19 +551,20 @@ class LSEVisitor : public HGraphVisitor {
     }
 
     // At this point, stores in possibly_removed_stores_ can be safely removed.
-    size = possibly_removed_stores_.size();
-    for (size_t i = 0; i < size; i++) {
+    for (size_t i = 0, e = possibly_removed_stores_.size(); i < e; i++) {
       HInstruction* store = possibly_removed_stores_[i];
       DCHECK(store->IsInstanceFieldSet() || store->IsStaticFieldSet() || store->IsArraySet());
       store->GetBlock()->RemoveInstruction(store);
     }
 
-    // TODO: remove unnecessary allocations.
-    // Eliminate instructions in singleton_new_instances_ that:
-    // - don't have uses,
-    // - don't have finalizers,
-    // - are instantiable and accessible,
-    // - have no/separate clinit check.
+    // Eliminate allocations that are not used.
+    for (size_t i = 0, e = singleton_new_instances_.size(); i < e; i++) {
+      HInstruction* new_instance = singleton_new_instances_[i];
+      if (!new_instance->HasNonEnvironmentUses()) {
+        new_instance->RemoveEnvironmentUsers();
+        new_instance->GetBlock()->RemoveInstruction(new_instance);
+      }
+    }
   }
 
  private:
@@ -969,8 +970,8 @@ class LSEVisitor : public HGraphVisitor {
     if (!heap_location_collector_.MayDeoptimize() &&
         ref_info->IsSingletonAndNotReturned() &&
         !new_instance->IsFinalizable() &&
-        !new_instance->CanThrow()) {
-      // TODO: add new_instance to singleton_new_instances_ and enable allocation elimination.
+        !new_instance->NeedsAccessCheck()) {
+      singleton_new_instances_.push_back(new_instance);
     }
     ArenaVector<HInstruction*>& heap_values =
         heap_values_for_[new_instance->GetBlock()->GetBlockId()];
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index dc5a8fa9cb..69f0b514f9 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -3652,14 +3652,14 @@ class HNewInstance : public HExpression<2> {
                uint32_t dex_pc,
                uint16_t type_index,
                const DexFile& dex_file,
-               bool can_throw,
+               bool needs_access_check,
                bool finalizable,
                QuickEntrypointEnum entrypoint)
       : HExpression(Primitive::kPrimNot, SideEffects::CanTriggerGC(), dex_pc),
         type_index_(type_index),
         dex_file_(dex_file),
         entrypoint_(entrypoint) {
-    SetPackedFlag<kFlagCanThrow>(can_throw);
+    SetPackedFlag<kFlagNeedsAccessCheck>(needs_access_check);
     SetPackedFlag<kFlagFinalizable>(finalizable);
     SetRawInputAt(0, cls);
     SetRawInputAt(1, current_method);
@@ -3671,10 +3671,11 @@ class HNewInstance : public HExpression<2> {
   // Calls runtime so needs an environment.
   bool NeedsEnvironment() const OVERRIDE { return true; }
 
-  // It may throw when called on type that's not instantiable/accessible.
-  // It can throw OOME.
-  // TODO: distinguish between the two cases so we can for example allow allocation elimination.
-  bool CanThrow() const OVERRIDE { return GetPackedFlag<kFlagCanThrow>() || true; }
+  // Can throw errors when out-of-memory or if it's not instantiable/accessible.
+  bool CanThrow() const OVERRIDE { return true; }
+
+  // Needs to call into runtime to make sure it's instantiable/accessible.
+  bool NeedsAccessCheck() const { return GetPackedFlag<kFlagNeedsAccessCheck>(); }
 
   bool IsFinalizable() const { return GetPackedFlag<kFlagFinalizable>(); }
 
@@ -3691,8 +3692,8 @@ class HNewInstance : public HExpression<2> {
   DECLARE_INSTRUCTION(NewInstance);
 
  private:
-  static constexpr size_t kFlagCanThrow = kNumberOfExpressionPackedBits;
-  static constexpr size_t kFlagFinalizable = kFlagCanThrow + 1;
+  static constexpr size_t kFlagNeedsAccessCheck = kNumberOfExpressionPackedBits;
+  static constexpr size_t kFlagFinalizable = kFlagNeedsAccessCheck + 1;
   static constexpr size_t kNumberOfNewInstancePackedBits = kFlagFinalizable + 1;
   static_assert(kNumberOfNewInstancePackedBits <= kMaxNumberOfPackedBits,
                 "Too many packed fields.");