Optimizing: Fix register allocator validation memory usage.
Also attribute ArenaBitVector allocations to appropriate
passes. This was used to track down the source of the
excessive memory alloactions.
Bug: 27690481
Change-Id: Ib895984cb7c04e24cbc7abbd8322079bab8ab100
diff --git a/compiler/optimizing/builder.cc b/compiler/optimizing/builder.cc
index 57660c2..124afbc 100644
--- a/compiler/optimizing/builder.cc
+++ b/compiler/optimizing/builder.cc
@@ -367,7 +367,8 @@
ArenaBitVector* native_debug_info_locations;
if (native_debuggable) {
const uint32_t num_instructions = code_item.insns_size_in_code_units_;
- native_debug_info_locations = new (arena_) ArenaBitVector (arena_, num_instructions, false);
+ native_debug_info_locations =
+ ArenaBitVector::Create(arena_, num_instructions, false, kArenaAllocGraphBuilder);
FindNativeDebugInfoLocations(code_item, native_debug_info_locations);
}
diff --git a/compiler/optimizing/code_generator.cc b/compiler/optimizing/code_generator.cc
index d64a786..32869ec 100644
--- a/compiler/optimizing/code_generator.cc
+++ b/compiler/optimizing/code_generator.cc
@@ -856,7 +856,8 @@
uint32_t register_mask = 0; // Not used.
// The stack mask is not used, so we leave it empty.
- ArenaBitVector* stack_mask = new (arena) ArenaBitVector(arena, 0, /* expandable */ true);
+ ArenaBitVector* stack_mask =
+ ArenaBitVector::Create(arena, 0, /* expandable */ true, kArenaAllocCodeGenerator);
stack_map_stream_.BeginStackMapEntry(dex_pc,
native_pc,
diff --git a/compiler/optimizing/dead_code_elimination.cc b/compiler/optimizing/dead_code_elimination.cc
index e170e37..d7bf16e 100644
--- a/compiler/optimizing/dead_code_elimination.cc
+++ b/compiler/optimizing/dead_code_elimination.cc
@@ -97,7 +97,7 @@
}
// Classify blocks as reachable/unreachable.
ArenaAllocator* allocator = graph_->GetArena();
- ArenaBitVector live_blocks(allocator, graph_->GetBlocks().size(), false);
+ ArenaBitVector live_blocks(allocator, graph_->GetBlocks().size(), false, kArenaAllocDCE);
MarkReachableBlocks(graph_, &live_blocks);
bool removed_one_or_more_blocks = false;
diff --git a/compiler/optimizing/graph_checker.cc b/compiler/optimizing/graph_checker.cc
index 11e3689..9491ef6 100644
--- a/compiler/optimizing/graph_checker.cc
+++ b/compiler/optimizing/graph_checker.cc
@@ -812,7 +812,10 @@
phi->GetRegNumber(),
type_str.str().c_str()));
} else {
- ArenaBitVector visited(GetGraph()->GetArena(), 0, /* expandable */ true);
+ ArenaBitVector visited(GetGraph()->GetArena(),
+ 0,
+ /* expandable */ true,
+ kArenaAllocGraphChecker);
if (!IsConstantEquivalent(phi, other_phi, &visited)) {
AddError(StringPrintf("Two phis (%d and %d) found for VReg %d but they "
"are not equivalents of constants.",
diff --git a/compiler/optimizing/graph_checker.h b/compiler/optimizing/graph_checker.h
index 52252cd..8da8457 100644
--- a/compiler/optimizing/graph_checker.h
+++ b/compiler/optimizing/graph_checker.h
@@ -30,7 +30,10 @@
: HGraphDelegateVisitor(graph),
errors_(graph->GetArena()->Adapter(kArenaAllocGraphChecker)),
dump_prefix_(dump_prefix),
- seen_ids_(graph->GetArena(), graph->GetCurrentInstructionId(), false) {}
+ seen_ids_(graph->GetArena(),
+ graph->GetCurrentInstructionId(),
+ false,
+ kArenaAllocGraphChecker) {}
// Check the whole graph (in reverse post-order).
void Run() {
diff --git a/compiler/optimizing/gvn.cc b/compiler/optimizing/gvn.cc
index b492278..f7eb2ad 100644
--- a/compiler/optimizing/gvn.cc
+++ b/compiler/optimizing/gvn.cc
@@ -40,7 +40,7 @@
: allocator_(allocator),
num_buckets_(kMinimumNumberOfBuckets),
buckets_(allocator->AllocArray<Node*>(num_buckets_, kArenaAllocGvn)),
- buckets_owned_(allocator, num_buckets_, false),
+ buckets_owned_(allocator, num_buckets_, false, kArenaAllocGvn),
num_entries_(0) {
// ArenaAllocator returns zeroed memory, so no need to set buckets to null.
DCHECK(IsPowerOfTwo(num_buckets_));
@@ -53,7 +53,7 @@
: allocator_(allocator),
num_buckets_(to_copy.IdealBucketCount()),
buckets_(allocator->AllocArray<Node*>(num_buckets_, kArenaAllocGvn)),
- buckets_owned_(allocator, num_buckets_, false),
+ buckets_owned_(allocator, num_buckets_, false, kArenaAllocGvn),
num_entries_(to_copy.num_entries_) {
// ArenaAllocator returns zeroed memory, so entries of buckets_ and
// buckets_owned_ are initialized to null and false, respectively.
diff --git a/compiler/optimizing/licm.cc b/compiler/optimizing/licm.cc
index 33bb2e8..7a1e06b 100644
--- a/compiler/optimizing/licm.cc
+++ b/compiler/optimizing/licm.cc
@@ -80,7 +80,7 @@
void LICM::Run() {
DCHECK(side_effects_.HasRun());
// Only used during debug.
- ArenaBitVector visited(graph_->GetArena(), graph_->GetBlocks().size(), false);
+ ArenaBitVector visited(graph_->GetArena(), graph_->GetBlocks().size(), false, kArenaAllocLICM);
// Post order visit to visit inner loops before outer loops.
for (HPostOrderIterator it(*graph_); !it.Done(); it.Advance()) {
diff --git a/compiler/optimizing/load_store_elimination.cc b/compiler/optimizing/load_store_elimination.cc
index 9601b06..e1977b1 100644
--- a/compiler/optimizing/load_store_elimination.cc
+++ b/compiler/optimizing/load_store_elimination.cc
@@ -186,7 +186,10 @@
: HGraphVisitor(graph),
ref_info_array_(graph->GetArena()->Adapter(kArenaAllocLSE)),
heap_locations_(graph->GetArena()->Adapter(kArenaAllocLSE)),
- aliasing_matrix_(graph->GetArena(), kInitialAliasingMatrixBitVectorSize, true),
+ aliasing_matrix_(graph->GetArena(),
+ kInitialAliasingMatrixBitVectorSize,
+ true,
+ kArenaAllocLSE),
has_heap_stores_(false),
has_volatile_(false),
has_monitor_operations_(false),
diff --git a/compiler/optimizing/locations.cc b/compiler/optimizing/locations.cc
index 1ab206f..83596da 100644
--- a/compiler/optimizing/locations.cc
+++ b/compiler/optimizing/locations.cc
@@ -37,7 +37,7 @@
if (NeedsSafepoint()) {
ArenaAllocator* arena = instruction->GetBlock()->GetGraph()->GetArena();
- stack_mask_ = new (arena) ArenaBitVector(arena, 0, true);
+ stack_mask_ = ArenaBitVector::Create(arena, 0, true, kArenaAllocLocationSummary);
}
}
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index 98766a3..c83340b 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -54,7 +54,7 @@
DCHECK_EQ(visited->GetHighestBitSet(), -1);
// Nodes that we're currently visiting, indexed by block id.
- ArenaBitVector visiting(arena_, blocks_.size(), false);
+ ArenaBitVector visiting(arena_, blocks_.size(), false, kArenaAllocGraphBuilder);
// Number of successors visited from a given node, indexed by block id.
ArenaVector<size_t> successors_visited(blocks_.size(), 0u, arena_->Adapter());
// Stack of nodes that we're currently visiting (same as marked in "visiting" above).
@@ -140,7 +140,7 @@
// collect both normal- and exceptional-flow values at the same time.
SimplifyCatchBlocks();
- ArenaBitVector visited(arena_, blocks_.size(), false);
+ ArenaBitVector visited(arena_, blocks_.size(), false, kArenaAllocGraphBuilder);
// (2) Find the back edges in the graph doing a DFS traversal.
FindBackEdges(&visited);
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 46377ee..6736319 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -645,7 +645,7 @@
irreducible_(false),
back_edges_(graph->GetArena()->Adapter(kArenaAllocLoopInfoBackEdges)),
// Make bit vector growable, as the number of blocks may change.
- blocks_(graph->GetArena(), graph->GetBlocks().size(), true) {
+ blocks_(graph->GetArena(), graph->GetBlocks().size(), true, kArenaAllocLoopInfoBackEdges) {
back_edges_.reserve(kDefaultNumberOfBackEdges);
}
diff --git a/compiler/optimizing/register_allocator.cc b/compiler/optimizing/register_allocator.cc
index b8d76b9..34d9af1 100644
--- a/compiler/optimizing/register_allocator.cc
+++ b/compiler/optimizing/register_allocator.cc
@@ -445,7 +445,7 @@
bool RegisterAllocator::ValidateInternal(bool log_fatal_on_failure) const {
// To simplify unit testing, we eagerly create the array of intervals, and
// call the helper method.
- ArenaVector<LiveInterval*> intervals(allocator_->Adapter(kArenaAllocRegisterAllocator));
+ ArenaVector<LiveInterval*> intervals(allocator_->Adapter(kArenaAllocRegisterAllocatorValidate));
for (size_t i = 0; i < liveness_.GetNumberOfSsaValues(); ++i) {
HInstruction* instruction = liveness_.GetInstructionFromSsaIndex(i);
if (ShouldProcess(processing_core_registers_, instruction->GetLiveInterval())) {
@@ -483,13 +483,21 @@
? codegen.GetNumberOfCoreRegisters()
: codegen.GetNumberOfFloatingPointRegisters();
ArenaVector<ArenaBitVector*> liveness_of_values(
- allocator->Adapter(kArenaAllocRegisterAllocator));
+ allocator->Adapter(kArenaAllocRegisterAllocatorValidate));
liveness_of_values.reserve(number_of_registers + number_of_spill_slots);
+ size_t max_end = 0u;
+ for (LiveInterval* start_interval : intervals) {
+ for (AllRangesIterator it(start_interval); !it.Done(); it.Advance()) {
+ max_end = std::max(max_end, it.CurrentRange()->GetEnd());
+ }
+ }
+
// Allocate a bit vector per register. A live interval that has a register
// allocated will populate the associated bit vector based on its live ranges.
for (size_t i = 0; i < number_of_registers + number_of_spill_slots; ++i) {
- liveness_of_values.push_back(new (allocator) ArenaBitVector(allocator, 0, true));
+ liveness_of_values.push_back(
+ ArenaBitVector::Create(allocator, max_end, false, kArenaAllocRegisterAllocatorValidate));
}
for (LiveInterval* start_interval : intervals) {
diff --git a/compiler/optimizing/ssa_liveness_analysis.h b/compiler/optimizing/ssa_liveness_analysis.h
index a78aedc..97f2aee 100644
--- a/compiler/optimizing/ssa_liveness_analysis.h
+++ b/compiler/optimizing/ssa_liveness_analysis.h
@@ -31,9 +31,9 @@
public:
BlockInfo(ArenaAllocator* allocator, const HBasicBlock& block, size_t number_of_ssa_values)
: block_(block),
- live_in_(allocator, number_of_ssa_values, false),
- live_out_(allocator, number_of_ssa_values, false),
- kill_(allocator, number_of_ssa_values, false) {
+ live_in_(allocator, number_of_ssa_values, false, kArenaAllocSsaLiveness),
+ live_out_(allocator, number_of_ssa_values, false, kArenaAllocSsaLiveness),
+ kill_(allocator, number_of_ssa_values, false, kArenaAllocSsaLiveness) {
UNUSED(block_);
live_in_.ClearAllBits();
live_out_.ClearAllBits();
diff --git a/compiler/optimizing/stack_map_stream.cc b/compiler/optimizing/stack_map_stream.cc
index 54cbdf8..3f41e35 100644
--- a/compiler/optimizing/stack_map_stream.cc
+++ b/compiler/optimizing/stack_map_stream.cc
@@ -37,7 +37,7 @@
current_entry_.same_dex_register_map_as_ = kNoSameDexMapFound;
if (num_dex_registers != 0) {
current_entry_.live_dex_registers_mask =
- new (allocator_) ArenaBitVector(allocator_, num_dex_registers, true);
+ ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream);
} else {
current_entry_.live_dex_registers_mask = nullptr;
}
@@ -111,7 +111,7 @@
current_inline_info_.dex_register_locations_start_index = dex_register_locations_.size();
if (num_dex_registers != 0) {
current_inline_info_.live_dex_registers_mask =
- new (allocator_) ArenaBitVector(allocator_, num_dex_registers, true);
+ ArenaBitVector::Create(allocator_, num_dex_registers, true, kArenaAllocStackMapStream);
} else {
current_inline_info_.live_dex_registers_mask = nullptr;
}
@@ -256,7 +256,7 @@
// Ensure we reached the end of the Dex registers location_catalog.
DCHECK_EQ(location_catalog_offset, dex_register_location_catalog_region.size());
- ArenaBitVector empty_bitmask(allocator_, 0, /* expandable */ false);
+ ArenaBitVector empty_bitmask(allocator_, 0, /* expandable */ false, kArenaAllocStackMapStream);
uintptr_t next_dex_register_map_offset = 0;
uintptr_t next_inline_info_offset = 0;
for (size_t i = 0, e = stack_maps_.size(); i < e; ++i) {