summaryrefslogtreecommitdiff
path: root/compiler/optimizing
diff options
context:
space:
mode:
Diffstat (limited to 'compiler/optimizing')
-rw-r--r--compiler/optimizing/code_generator_arm.cc81
-rw-r--r--compiler/optimizing/inliner.cc20
-rw-r--r--compiler/optimizing/load_store_elimination.cc2
-rw-r--r--compiler/optimizing/nodes.cc29
-rw-r--r--compiler/optimizing/nodes.h3
-rw-r--r--compiler/optimizing/optimizing_compiler.cc4
6 files changed, 98 insertions, 41 deletions
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 8d9794bd79..3dc3b7fba0 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -56,6 +56,8 @@ static constexpr SRegister kFpuCalleeSaves[] =
// S registers. Therefore there is no need to block it.
static constexpr DRegister DTMP = D31;
+static constexpr uint32_t kPackedSwitchJumpTableThreshold = 6;
+
#define __ down_cast<ArmAssembler*>(codegen->GetAssembler())->
#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmWordSize, x).Int32Value()
@@ -513,17 +515,6 @@ void CodeGeneratorARM::Finalize(CodeAllocator* allocator) {
uint32_t new_position = __ GetAdjustedPosition(old_position);
stack_map_stream_.SetStackMapNativePcOffset(i, new_position);
}
- // Adjust native pc offsets of block labels.
- for (HBasicBlock* block : *block_order_) {
- // Get the label directly from block_labels_ rather than through GetLabelOf() to avoid
- // FirstNonEmptyBlock() which could lead to adjusting a label more than once.
- DCHECK_LT(block->GetBlockId(), GetGraph()->GetBlocks().size());
- Label* block_label = &block_labels_[block->GetBlockId()];
- DCHECK_EQ(block_label->IsBound(), !block->IsSingleJump());
- if (block_label->IsBound()) {
- __ AdjustLabelPosition(block_label);
- }
- }
// Adjust pc offsets for the disassembly information.
if (disasm_info_ != nullptr) {
GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval();
@@ -538,10 +529,6 @@ void CodeGeneratorARM::Finalize(CodeAllocator* allocator) {
it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end);
}
}
- // Adjust pc offsets for relative call patches.
- for (MethodPatchInfo<Label>& info : relative_call_patches_) {
- __ AdjustLabelPosition(&info.label);
- }
CodeGenerator::Finalize(allocator);
}
@@ -732,7 +719,8 @@ void CodeGeneratorARM::GenerateFrameExit() {
}
void CodeGeneratorARM::Bind(HBasicBlock* block) {
- __ Bind(GetLabelOf(block));
+ Label* label = GetLabelOf(block);
+ __ BindTrackedLabel(label);
}
Location CodeGeneratorARM::GetStackLocation(HLoadLocal* load) const {
@@ -5255,7 +5243,7 @@ void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke,
break;
case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative:
relative_call_patches_.emplace_back(invoke->GetTargetMethod());
- __ Bind(&relative_call_patches_.back().label);
+ __ BindTrackedLabel(&relative_call_patches_.back().label);
// Arbitrarily branch to the BL itself, override at link time.
__ bl(&relative_call_patches_.back().label);
break;
@@ -5378,25 +5366,64 @@ void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) {
LocationSummary* locations =
new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall);
locations->SetInAt(0, Location::RequiresRegister());
+ if (switch_instr->GetNumEntries() >= kPackedSwitchJumpTableThreshold &&
+ codegen_->GetAssembler()->IsThumb()) {
+ locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base.
+ if (switch_instr->GetStartValue() != 0) {
+ locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias.
+ }
+ }
}
void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) {
int32_t lower_bound = switch_instr->GetStartValue();
- int32_t num_entries = switch_instr->GetNumEntries();
+ uint32_t num_entries = switch_instr->GetNumEntries();
LocationSummary* locations = switch_instr->GetLocations();
Register value_reg = locations->InAt(0).AsRegister<Register>();
HBasicBlock* default_block = switch_instr->GetDefaultBlock();
- // Create a series of compare/jumps.
- const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors();
- for (int32_t i = 0; i < num_entries; i++) {
- GenerateCompareWithImmediate(value_reg, lower_bound + i);
- __ b(codegen_->GetLabelOf(successors[i]), EQ);
- }
+ if (num_entries < kPackedSwitchJumpTableThreshold || !codegen_->GetAssembler()->IsThumb()) {
+ // Create a series of compare/jumps.
+ const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors();
+ for (uint32_t i = 0; i < num_entries; i++) {
+ GenerateCompareWithImmediate(value_reg, lower_bound + i);
+ __ b(codegen_->GetLabelOf(successors[i]), EQ);
+ }
+
+ // And the default for any other value.
+ if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) {
+ __ b(codegen_->GetLabelOf(default_block));
+ }
+ } else {
+ // Create a table lookup.
+ Register temp_reg = locations->GetTemp(0).AsRegister<Register>();
+
+ // Materialize a pointer to the switch table
+ std::vector<Label*> labels(num_entries);
+ const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors();
+ for (uint32_t i = 0; i < num_entries; i++) {
+ labels[i] = codegen_->GetLabelOf(successors[i]);
+ }
+ JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg);
+
+ // Remove the bias.
+ Register key_reg;
+ if (lower_bound != 0) {
+ key_reg = locations->GetTemp(1).AsRegister<Register>();
+ __ AddConstant(key_reg, value_reg, -lower_bound);
+ } else {
+ key_reg = value_reg;
+ }
+
+ // Check whether the value is in the table, jump to default block if not.
+ __ CmpConstant(key_reg, num_entries - 1);
+ __ b(codegen_->GetLabelOf(default_block), Condition::HI);
+
+ // Load the displacement from the table.
+ __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2));
- // And the default for any other value.
- if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) {
- __ b(codegen_->GetLabelOf(default_block));
+ // Dispatch is a direct add to the PC (for Thumb2).
+ __ EmitJumpTableDispatch(table, temp_reg);
}
}
diff --git a/compiler/optimizing/inliner.cc b/compiler/optimizing/inliner.cc
index 0aaa6b3f2c..353881e47a 100644
--- a/compiler/optimizing/inliner.cc
+++ b/compiler/optimizing/inliner.cc
@@ -494,6 +494,26 @@ bool HInliner::TryBuildAndInline(ArtMethod* resolved_method,
<< " it is in a different dex file and requires access to the dex cache";
return false;
}
+
+ if (current->IsNewInstance() &&
+ (current->AsNewInstance()->GetEntrypoint() == kQuickAllocObjectWithAccessCheck)) {
+ // Allocation entrypoint does not handle inlined frames.
+ return false;
+ }
+
+ if (current->IsNewArray() &&
+ (current->AsNewArray()->GetEntrypoint() == kQuickAllocArrayWithAccessCheck)) {
+ // Allocation entrypoint does not handle inlined frames.
+ return false;
+ }
+
+ if (current->IsUnresolvedStaticFieldGet() ||
+ current->IsUnresolvedInstanceFieldGet() ||
+ current->IsUnresolvedStaticFieldSet() ||
+ current->IsUnresolvedInstanceFieldSet()) {
+ // Entrypoint for unresolved fields does not handle inlined frames.
+ return false;
+ }
}
}
number_of_inlined_instructions_ += number_of_instructions;
diff --git a/compiler/optimizing/load_store_elimination.cc b/compiler/optimizing/load_store_elimination.cc
index 90f28e511e..6fbb6823d6 100644
--- a/compiler/optimizing/load_store_elimination.cc
+++ b/compiler/optimizing/load_store_elimination.cc
@@ -59,7 +59,7 @@ class ReferenceInfo : public ArenaObject<kArenaAllocMisc> {
(use->IsInstanceFieldSet() && (reference_ == use->InputAt(1))) ||
(use->IsUnresolvedInstanceFieldSet() && (reference_ == use->InputAt(1))) ||
(use->IsStaticFieldSet() && (reference_ == use->InputAt(1))) ||
- (use->IsUnresolvedStaticFieldSet() && (reference_ == use->InputAt(1))) ||
+ (use->IsUnresolvedStaticFieldSet() && (reference_ == use->InputAt(0))) ||
(use->IsArraySet() && (reference_ == use->InputAt(2)))) {
// reference_ is merged to a phi, passed to a callee, or stored to heap.
// reference_ isn't the only name that can refer to its value anymore.
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index 8b28ff91d4..68fb0acf7f 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -1652,7 +1652,8 @@ HInstruction* HGraph::InlineInto(HGraph* outer_graph, HInvoke* invoke) {
// Update the meta information surrounding blocks:
// (1) the graph they are now in,
// (2) the reverse post order of that graph,
- // (3) the potential loop information they are now in.
+ // (3) the potential loop information they are now in,
+ // (4) try block membership.
// We don't add the entry block, the exit block, and the first block, which
// has been merged with `at`.
@@ -1668,41 +1669,47 @@ HInstruction* HGraph::InlineInto(HGraph* outer_graph, HInvoke* invoke) {
size_t index_of_at = IndexOfElement(outer_graph->reverse_post_order_, at);
MakeRoomFor(&outer_graph->reverse_post_order_, blocks_added, index_of_at);
- // Do a reverse post order of the blocks in the callee and do (1), (2),
- // and (3) to the blocks that apply.
- HLoopInformation* info = at->GetLoopInformation();
+ HLoopInformation* loop_info = at->GetLoopInformation();
+ // Copy TryCatchInformation if `at` is a try block, not if it is a catch block.
+ TryCatchInformation* try_catch_info = at->IsTryBlock() ? at->GetTryCatchInformation() : nullptr;
+
+ // Do a reverse post order of the blocks in the callee and do (1), (2), (3)
+ // and (4) to the blocks that apply.
for (HReversePostOrderIterator it(*this); !it.Done(); it.Advance()) {
HBasicBlock* current = it.Current();
if (current != exit_block_ && current != entry_block_ && current != first) {
DCHECK(!current->IsInLoop());
+ DCHECK(current->GetTryCatchInformation() == nullptr);
DCHECK(current->GetGraph() == this);
current->SetGraph(outer_graph);
outer_graph->AddBlock(current);
outer_graph->reverse_post_order_[++index_of_at] = current;
- if (info != nullptr) {
- current->SetLoopInformation(info);
+ if (loop_info != nullptr) {
+ current->SetLoopInformation(loop_info);
for (HLoopInformationOutwardIterator loop_it(*at); !loop_it.Done(); loop_it.Advance()) {
loop_it.Current()->Add(current);
}
}
+ current->SetTryCatchInformation(try_catch_info);
}
}
- // Do (1), (2), and (3) to `to`.
+ // Do (1), (2), (3) and (4) to `to`.
to->SetGraph(outer_graph);
outer_graph->AddBlock(to);
outer_graph->reverse_post_order_[++index_of_at] = to;
- if (info != nullptr) {
- to->SetLoopInformation(info);
+ if (loop_info != nullptr) {
+ to->SetLoopInformation(loop_info);
for (HLoopInformationOutwardIterator loop_it(*at); !loop_it.Done(); loop_it.Advance()) {
loop_it.Current()->Add(to);
}
- if (info->IsBackEdge(*at)) {
+ if (loop_info->IsBackEdge(*at)) {
// Only `to` can become a back edge, as the inlined blocks
// are predecessors of `to`.
- info->ReplaceBackEdge(at, to);
+ loop_info->ReplaceBackEdge(at, to);
}
}
+ to->SetTryCatchInformation(try_catch_info);
}
// Update the next instruction id of the outer graph, so that instructions
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 7df586692b..0f2c1cffee 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -4750,6 +4750,9 @@ class HLoadClass : public HExpression<1> {
return generate_clinit_check_;
}
void SetMustGenerateClinitCheck(bool generate_clinit_check) {
+ // The entrypoint the code generator is going to call does not do
+ // clinit of the class.
+ DCHECK(!NeedsAccessCheck());
generate_clinit_check_ = generate_clinit_check;
}
diff --git a/compiler/optimizing/optimizing_compiler.cc b/compiler/optimizing/optimizing_compiler.cc
index 6632f95ebe..8cb2cfc816 100644
--- a/compiler/optimizing/optimizing_compiler.cc
+++ b/compiler/optimizing/optimizing_compiler.cc
@@ -492,6 +492,8 @@ static void RunOptimizations(HGraph* graph,
RunOptimizations(optimizations1, arraysize(optimizations1), pass_observer);
+ MaybeRunInliner(graph, codegen, driver, stats, dex_compilation_unit, pass_observer, handles);
+
// TODO: Update passes incompatible with try/catch so we have the same
// pipeline for all methods.
if (graph->HasTryCatch()) {
@@ -507,8 +509,6 @@ static void RunOptimizations(HGraph* graph,
RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer);
} else {
- MaybeRunInliner(graph, codegen, driver, stats, dex_compilation_unit, pass_observer, handles);
-
HOptimization* optimizations2[] = {
// BooleanSimplifier depends on the InstructionSimplifier removing
// redundant suspend checks to recognize empty blocks.
generated by cgit v1.2.3-59-g8ed1b (git 2.39.0) at 2025-10-25 01:55:59 +0000