blob: fe1361c93585f29471cc5d1e7ae3b197c3d2c288 [file] [log] [blame]
/*
* Copyright (C) 2014 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "dead_code_elimination.h"
#include "android-base/logging.h"
#include "base/array_ref.h"
#include "base/bit_vector-inl.h"
#include "base/logging.h"
#include "base/scoped_arena_allocator.h"
#include "base/scoped_arena_containers.h"
#include "base/stl_util.h"
#include "optimizing/nodes.h"
#include "optimizing/nodes_vector.h"
#include "ssa_phi_elimination.h"
namespace art HIDDEN {
static void MarkReachableBlocks(HGraph* graph, ArenaBitVector* visited) {
// Use local allocator for allocating memory.
ScopedArenaAllocator allocator(graph->GetArenaStack());
ScopedArenaVector<HBasicBlock*> worklist(allocator.Adapter(kArenaAllocDCE));
constexpr size_t kDefaultWorlistSize = 8;
worklist.reserve(kDefaultWorlistSize);
visited->SetBit(graph->GetEntryBlock()->GetBlockId());
worklist.push_back(graph->GetEntryBlock());
while (!worklist.empty()) {
HBasicBlock* block = worklist.back();
worklist.pop_back();
int block_id = block->GetBlockId();
DCHECK(visited->IsBitSet(block_id));
ArrayRef<HBasicBlock* const> live_successors(block->GetSuccessors());
HInstruction* last_instruction = block->GetLastInstruction();
if (last_instruction->IsIf()) {
HIf* if_instruction = last_instruction->AsIf();
HInstruction* condition = if_instruction->InputAt(0);
if (condition->IsIntConstant()) {
if (condition->AsIntConstant()->IsTrue()) {
live_successors = live_successors.SubArray(0u, 1u);
DCHECK_EQ(live_successors[0], if_instruction->IfTrueSuccessor());
} else {
DCHECK(condition->AsIntConstant()->IsFalse()) << condition->AsIntConstant()->GetValue();
live_successors = live_successors.SubArray(1u, 1u);
DCHECK_EQ(live_successors[0], if_instruction->IfFalseSuccessor());
}
}
} else if (last_instruction->IsPackedSwitch()) {
HPackedSwitch* switch_instruction = last_instruction->AsPackedSwitch();
HInstruction* switch_input = switch_instruction->InputAt(0);
if (switch_input->IsIntConstant()) {
int32_t switch_value = switch_input->AsIntConstant()->GetValue();
int32_t start_value = switch_instruction->GetStartValue();
// Note: Though the spec forbids packed-switch values to wrap around, we leave
// that task to the verifier and use unsigned arithmetic with it's "modulo 2^32"
// semantics to check if the value is in range, wrapped or not.
uint32_t switch_index =
static_cast<uint32_t>(switch_value) - static_cast<uint32_t>(start_value);
if (switch_index < switch_instruction->GetNumEntries()) {
live_successors = live_successors.SubArray(switch_index, 1u);
DCHECK_EQ(live_successors[0], block->GetSuccessors()[switch_index]);
} else {
live_successors = live_successors.SubArray(switch_instruction->GetNumEntries(), 1u);
DCHECK_EQ(live_successors[0], switch_instruction->GetDefaultBlock());
}
}
}
for (HBasicBlock* successor : live_successors) {
// Add only those successors that have not been visited yet.
if (!visited->IsBitSet(successor->GetBlockId())) {
visited->SetBit(successor->GetBlockId());
worklist.push_back(successor);
}
}
}
}
void HDeadCodeElimination::MaybeRecordDeadBlock(HBasicBlock* block) {
if (stats_ != nullptr) {
stats_->RecordStat(MethodCompilationStat::kRemovedDeadInstruction,
block->GetPhis().CountSize() + block->GetInstructions().CountSize());
}
}
void HDeadCodeElimination::MaybeRecordSimplifyIf() {
if (stats_ != nullptr) {
stats_->RecordStat(MethodCompilationStat::kSimplifyIf);
}
}
static bool HasInput(HCondition* instruction, HInstruction* input) {
return (instruction->InputAt(0) == input) ||
(instruction->InputAt(1) == input);
}
static bool HasEquality(IfCondition condition) {
switch (condition) {
case kCondEQ:
case kCondLE:
case kCondGE:
case kCondBE:
case kCondAE:
return true;
case kCondNE:
case kCondLT:
case kCondGT:
case kCondB:
case kCondA:
return false;
}
}
static HConstant* Evaluate(HCondition* condition, HInstruction* left, HInstruction* right) {
if (left == right && !DataType::IsFloatingPointType(left->GetType())) {
return condition->GetBlock()->GetGraph()->GetIntConstant(
HasEquality(condition->GetCondition()) ? 1 : 0);
}
if (!left->IsConstant() || !right->IsConstant()) {
return nullptr;
}
if (left->IsIntConstant()) {
return condition->Evaluate(left->AsIntConstant(), right->AsIntConstant());
} else if (left->IsNullConstant()) {
return condition->Evaluate(left->AsNullConstant(), right->AsNullConstant());
} else if (left->IsLongConstant()) {
return condition->Evaluate(left->AsLongConstant(), right->AsLongConstant());
} else if (left->IsFloatConstant()) {
return condition->Evaluate(left->AsFloatConstant(), right->AsFloatConstant());
} else {
DCHECK(left->IsDoubleConstant());
return condition->Evaluate(left->AsDoubleConstant(), right->AsDoubleConstant());
}
}
static bool RemoveNonNullControlDependences(HBasicBlock* block, HBasicBlock* throws) {
// Test for an if as last statement.
if (!block->EndsWithIf()) {
return false;
}
HIf* ifs = block->GetLastInstruction()->AsIf();
// Find either:
// if obj == null
// throws
// else
// not_throws
// or:
// if obj != null
// not_throws
// else
// throws
HInstruction* cond = ifs->InputAt(0);
HBasicBlock* not_throws = nullptr;
if (throws == ifs->IfTrueSuccessor() && cond->IsEqual()) {
not_throws = ifs->IfFalseSuccessor();
} else if (throws == ifs->IfFalseSuccessor() && cond->IsNotEqual()) {
not_throws = ifs->IfTrueSuccessor();
} else {
return false;
}
DCHECK(cond->IsEqual() || cond->IsNotEqual());
HInstruction* obj = cond->InputAt(1);
if (obj->IsNullConstant()) {
obj = cond->InputAt(0);
} else if (!cond->InputAt(0)->IsNullConstant()) {
return false;
}
// We can't create a BoundType for an object with an invalid RTI.
const ReferenceTypeInfo ti = obj->GetReferenceTypeInfo();
if (!ti.IsValid()) {
return false;
}
// Scan all uses of obj and find null check under control dependence.
HBoundType* bound = nullptr;
const HUseList<HInstruction*>& uses = obj->GetUses();
for (auto it = uses.begin(), end = uses.end(); it != end;) {
HInstruction* user = it->GetUser();
++it; // increment before possibly replacing
if (user->IsNullCheck()) {
HBasicBlock* user_block = user->GetBlock();
if (user_block != block &&
user_block != throws &&
block->Dominates(user_block)) {
if (bound == nullptr) {
bound = new (obj->GetBlock()->GetGraph()->GetAllocator()) HBoundType(obj);
bound->SetUpperBound(ti, /*can_be_null*/ false);
bound->SetReferenceTypeInfo(ti);
bound->SetCanBeNull(false);
not_throws->InsertInstructionBefore(bound, not_throws->GetFirstInstruction());
}
user->ReplaceWith(bound);
user_block->RemoveInstruction(user);
}
}
}
return bound != nullptr;
}
// Simplify the pattern:
//
// B1
// / \
// | instr_1
// | ...
// | instr_n
// | foo() // always throws
// | instr_n+2
// | ...
// | instr_n+m
// \ goto B2
// \ /
// B2
//
// Into:
//
// B1
// / \
// | instr_1
// | ...
// | instr_n
// | foo()
// | goto Exit
// | |
// B2 Exit
//
// Rationale:
// Removal of the never taken edge to B2 may expose other optimization opportunities, such as code
// sinking.
//
// Note: The example above is a simple one that uses a `goto` but we could end the block with an If,
// for example.
bool HDeadCodeElimination::SimplifyAlwaysThrows() {
HBasicBlock* exit = graph_->GetExitBlock();
if (!graph_->HasAlwaysThrowingInvokes() || exit == nullptr) {
return false;
}
bool rerun_dominance_and_loop_analysis = false;
// Order does not matter, just pick one.
for (HBasicBlock* block : graph_->GetReversePostOrder()) {
if (block->IsTryBlock()) {
// We don't want to perform the simplify always throws optimizations for throws inside of
// tries since those throws might not go to the exit block.
continue;
}
// We iterate to find the first instruction that always throws. If two instructions always
// throw, the first one will throw and the second one will never be reached.
HInstruction* throwing_invoke = nullptr;
for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
if (it.Current()->IsInvoke() && it.Current()->AsInvoke()->AlwaysThrows()) {
throwing_invoke = it.Current();
break;
}
}
if (throwing_invoke == nullptr) {
// No always-throwing instruction found. Continue with the rest of the blocks.
continue;
}
// If we are already pointing at the exit block we could still remove the instructions
// between the always throwing instruction, and the exit block. If we have no other
// instructions, just continue since there's nothing to do.
if (block->GetSuccessors().size() == 1 &&
block->GetSingleSuccessor() == exit &&
block->GetLastInstruction()->GetPrevious() == throwing_invoke) {
continue;
}
// We split the block at the throwing instruction, and the instructions after the throwing
// instructions will be disconnected from the graph after `block` points to the exit.
// `RemoveDeadBlocks` will take care of removing this new block and its instructions.
// Even though `SplitBefore` doesn't guarantee the graph to remain in SSA form, it is fine
// since we do not break it.
HBasicBlock* new_block = block->SplitBefore(throwing_invoke->GetNext(),
/* require_graph_not_in_ssa_form= */ false);
DCHECK_EQ(block->GetSingleSuccessor(), new_block);
block->ReplaceSuccessor(new_block, exit);
rerun_dominance_and_loop_analysis = true;
MaybeRecordStat(stats_, MethodCompilationStat::kSimplifyThrowingInvoke);
// Perform a quick follow up optimization on object != null control dependences
// that is much cheaper to perform now than in a later phase.
// If there are multiple predecessors, none may end with a HIf as required in
// RemoveNonNullControlDependences because we split critical edges.
if (block->GetPredecessors().size() == 1u &&
RemoveNonNullControlDependences(block->GetSinglePredecessor(), block)) {
MaybeRecordStat(stats_, MethodCompilationStat::kRemovedNullCheck);
}
}
// We need to re-analyze the graph in order to run DCE afterwards.
if (rerun_dominance_and_loop_analysis) {
graph_->RecomputeDominatorTree();
return true;
}
return false;
}
bool HDeadCodeElimination::SimplifyIfs() {
bool simplified_one_or_more_ifs = false;
bool rerun_dominance_and_loop_analysis = false;
// Iterating in PostOrder it's better for MaybeAddPhi as it can add a Phi for multiple If
// instructions in a chain without updating the dominator chain. The branch redirection itself can
// work in PostOrder or ReversePostOrder without issues.
for (HBasicBlock* block : graph_->GetPostOrder()) {
if (block->IsCatchBlock()) {
// This simplification cannot be applied to catch blocks, because exception handler edges do
// not represent normal control flow. Though in theory this could still apply to normal
// control flow going directly to a catch block, we cannot support it at the moment because
// the catch Phi's inputs do not correspond to the catch block's predecessors, so we cannot
// identify which predecessor corresponds to a given statically evaluated input.
continue;
}
HInstruction* last = block->GetLastInstruction();
if (!last->IsIf()) {
continue;
}
if (block->IsLoopHeader()) {
// We do not apply this optimization to loop headers as this could create irreducible loops.
continue;
}
// We will add a Phi which allows the simplification to take place in cases where it wouldn't.
MaybeAddPhi(block);
// TODO(solanes): Investigate support for multiple phis in `block`. We can potentially "push
// downwards" existing Phis into the true/false branches. For example, let's say we have another
// Phi: Phi(x1,x2,x3,x4,x5,x6). This could turn into Phi(x1,x2) in the true branch, Phi(x3,x4)
// in the false branch, and remain as Phi(x5,x6) in `block` (for edges that we couldn't
// redirect). We might even be able to remove some phis altogether as they will have only one
// value.
if (block->HasSinglePhi() &&
block->GetFirstPhi()->HasOnlyOneNonEnvironmentUse()) {
HInstruction* first = block->GetFirstInstruction();
bool has_only_phi_and_if = (last == first) && (last->InputAt(0) == block->GetFirstPhi());
bool has_only_phi_condition_and_if =
!has_only_phi_and_if &&
first->IsCondition() &&
HasInput(first->AsCondition(), block->GetFirstPhi()) &&
(first->GetNext() == last) &&
(last->InputAt(0) == first) &&
first->HasOnlyOneNonEnvironmentUse();
if (has_only_phi_and_if || has_only_phi_condition_and_if) {
HPhi* phi = block->GetFirstPhi()->AsPhi();
bool phi_input_is_left = (first->InputAt(0) == phi);
// Walk over all inputs of the phis and update the control flow of
// predecessors feeding constants to the phi.
// Note that phi->InputCount() may change inside the loop.
for (size_t i = 0; i < phi->InputCount();) {
HInstruction* input = phi->InputAt(i);
HInstruction* value_to_check = nullptr;
if (has_only_phi_and_if) {
if (input->IsIntConstant()) {
value_to_check = input;
}
} else {
DCHECK(has_only_phi_condition_and_if);
if (phi_input_is_left) {
value_to_check = Evaluate(first->AsCondition(), input, first->InputAt(1));
} else {
value_to_check = Evaluate(first->AsCondition(), first->InputAt(0), input);
}
}
if (value_to_check == nullptr) {
// Could not evaluate to a constant, continue iterating over the inputs.
++i;
} else {
HBasicBlock* predecessor_to_update = block->GetPredecessors()[i];
HBasicBlock* successor_to_update = nullptr;
if (value_to_check->AsIntConstant()->IsTrue()) {
successor_to_update = last->AsIf()->IfTrueSuccessor();
} else {
DCHECK(value_to_check->AsIntConstant()->IsFalse())
<< value_to_check->AsIntConstant()->GetValue();
successor_to_update = last->AsIf()->IfFalseSuccessor();
}
predecessor_to_update->ReplaceSuccessor(block, successor_to_update);
phi->RemoveInputAt(i);
simplified_one_or_more_ifs = true;
if (block->IsInLoop()) {
rerun_dominance_and_loop_analysis = true;
}
// For simplicity, don't create a dead block, let the dead code elimination
// pass deal with it.
if (phi->InputCount() == 1) {
break;
}
}
}
if (block->GetPredecessors().size() == 1) {
phi->ReplaceWith(phi->InputAt(0));
block->RemovePhi(phi);
if (has_only_phi_condition_and_if) {
// Evaluate here (and not wait for a constant folding pass) to open
// more opportunities for DCE.
HInstruction* result = first->AsCondition()->TryStaticEvaluation();
if (result != nullptr) {
first->ReplaceWith(result);
block->RemoveInstruction(first);
}
}
}
if (simplified_one_or_more_ifs) {
MaybeRecordSimplifyIf();
}
}
}
}
// We need to re-analyze the graph in order to run DCE afterwards.
if (simplified_one_or_more_ifs) {
if (rerun_dominance_and_loop_analysis) {
graph_->RecomputeDominatorTree();
} else {
graph_->ClearDominanceInformation();
// We have introduced critical edges, remove them.
graph_->SimplifyCFG();
graph_->ComputeDominanceInformation();
graph_->ComputeTryBlockInformation();
}
}
return simplified_one_or_more_ifs;
}
void HDeadCodeElimination::MaybeAddPhi(HBasicBlock* block) {
DCHECK(block->GetLastInstruction()->IsIf());
HIf* if_instruction = block->GetLastInstruction()->AsIf();
if (if_instruction->InputAt(0)->IsConstant()) {
// Constant values are handled in RemoveDeadBlocks.
return;
}
if (block->GetNumberOfPredecessors() < 2u) {
// Nothing to redirect.
return;
}
if (!block->GetPhis().IsEmpty()) {
// SimplifyIf doesn't currently work with multiple phis. Adding a phi here won't help that
// optimization.
return;
}
HBasicBlock* dominator = block->GetDominator();
if (!dominator->EndsWithIf()) {
return;
}
HInstruction* input = if_instruction->InputAt(0);
HInstruction* dominator_input = dominator->GetLastInstruction()->AsIf()->InputAt(0);
const bool same_input = dominator_input == input;
if (!same_input) {
// Try to see if the dominator has the opposite input (e.g. if(cond) and if(!cond)). If that's
// the case, we can perform the optimization with the false and true branches reversed.
if (!dominator_input->IsCondition() || !input->IsCondition()) {
return;
}
HCondition* block_cond = input->AsCondition();
HCondition* dominator_cond = dominator_input->AsCondition();
if (block_cond->GetLeft() != dominator_cond->GetLeft() ||
block_cond->GetRight() != dominator_cond->GetRight() ||
block_cond->GetOppositeCondition() != dominator_cond->GetCondition()) {
return;
}
}
if (kIsDebugBuild) {
// `block`'s successors should have only one predecessor. Otherwise, we have a critical edge in
// the graph.
for (HBasicBlock* succ : block->GetSuccessors()) {
DCHECK_EQ(succ->GetNumberOfPredecessors(), 1u);
}
}
const size_t pred_size = block->GetNumberOfPredecessors();
HPhi* new_phi = new (graph_->GetAllocator())
HPhi(graph_->GetAllocator(), kNoRegNumber, pred_size, DataType::Type::kInt32);
for (size_t index = 0; index < pred_size; index++) {
HBasicBlock* pred = block->GetPredecessors()[index];
const bool dominated_by_true =
dominator->GetLastInstruction()->AsIf()->IfTrueSuccessor()->Dominates(pred);
const bool dominated_by_false =
dominator->GetLastInstruction()->AsIf()->IfFalseSuccessor()->Dominates(pred);
if (dominated_by_true == dominated_by_false) {
// In this case, we can't know if we are coming from the true branch, or the false branch. It
// happens in cases like:
// 1 (outer if)
// / \
// 2 3 (inner if)
// | / \
// | 4 5
// \/ |
// 6 |
// \ |
// 7 (has the same if(cond) as 1)
// |
// 8
// `7` (which would be `block` in this example), and `6` will come from both the true path and
// the false path of `1`. We bumped into something similar in SelectGenerator. See
// HSelectGenerator::TryFixupDoubleDiamondPattern.
// TODO(solanes): Figure out if we can fix up the graph into a double diamond in a generic way
// so that DeadCodeElimination and SelectGenerator can take advantage of it.
if (!same_input) {
// `1` and `7` having the opposite condition is a case we are missing. We could potentially
// add a BooleanNot instruction to be able to add the Phi, but it seems like overkill since
// this case is not that common.
return;
}
// The Phi will have `0`, `1`, and `cond` as inputs. If SimplifyIf redirects 0s and 1s, we
// will end up with Phi(cond,...,cond) which will be replaced by `cond`. Effectively, we will
// redirect edges that we are able to redirect and the rest will remain as before (i.e. we
// won't have an extra Phi).
new_phi->SetRawInputAt(index, input);
} else {
// Redirect to either the true branch (1), or the false branch (0).
// Given that `dominated_by_true` is the exact opposite of `dominated_by_false`,
// `(same_input && dominated_by_true) || (!same_input && dominated_by_false)` is equivalent to
// `same_input == dominated_by_true`.
new_phi->SetRawInputAt(
index,
same_input == dominated_by_true ? graph_->GetIntConstant(1) : graph_->GetIntConstant(0));
}
}
block->AddPhi(new_phi);
if_instruction->ReplaceInput(new_phi, 0);
// Remove the old input now, if possible. This allows the branch redirection in SimplifyIf to
// work without waiting for another pass of DCE.
if (input->IsDeadAndRemovable()) {
DCHECK(!same_input)
<< " if both blocks have the same condition, it shouldn't be dead and removable since the "
<< "dominator block's If instruction would be using that condition.";
input->GetBlock()->RemoveInstruction(input);
}
MaybeRecordStat(stats_, MethodCompilationStat::kSimplifyIfAddedPhi);
}
void HDeadCodeElimination::ConnectSuccessiveBlocks() {
// Order does not matter. Skip the entry block by starting at index 1 in reverse post order.
for (size_t i = 1u, size = graph_->GetReversePostOrder().size(); i != size; ++i) {
HBasicBlock* block = graph_->GetReversePostOrder()[i];
DCHECK(!block->IsEntryBlock());
while (block->GetLastInstruction()->IsGoto()) {
HBasicBlock* successor = block->GetSingleSuccessor();
if (successor->IsExitBlock() || successor->GetPredecessors().size() != 1u) {
break;
}
DCHECK_LT(i, IndexOfElement(graph_->GetReversePostOrder(), successor));
block->MergeWith(successor);
--size;
DCHECK_EQ(size, graph_->GetReversePostOrder().size());
DCHECK_EQ(block, graph_->GetReversePostOrder()[i]);
// Reiterate on this block in case it can be merged with its new successor.
}
}
}
struct HDeadCodeElimination::TryBelongingInformation {
TryBelongingInformation(HGraph* graph, ScopedArenaAllocator* allocator)
: blocks_in_try(allocator, graph->GetBlocks().size(), /*expandable=*/false, kArenaAllocDCE),
coalesced_try_entries(
allocator, graph->GetBlocks().size(), /*expandable=*/false, kArenaAllocDCE) {
blocks_in_try.ClearAllBits();
coalesced_try_entries.ClearAllBits();
}
// Which blocks belong in the try.
ArenaBitVector blocks_in_try;
// Which other try entries are referencing this same try.
ArenaBitVector coalesced_try_entries;
};
bool HDeadCodeElimination::CanPerformTryRemoval(const TryBelongingInformation& try_belonging_info) {
const ArenaVector<HBasicBlock*>& blocks = graph_->GetBlocks();
for (uint32_t i : try_belonging_info.blocks_in_try.Indexes()) {
for (HInstructionIterator it(blocks[i]->GetInstructions()); !it.Done(); it.Advance()) {
if (it.Current()->CanThrow()) {
return false;
}
}
}
return true;
}
void HDeadCodeElimination::DisconnectHandlersAndUpdateTryBoundary(
HBasicBlock* block,
/* out */ bool* any_block_in_loop) {
if (block->IsInLoop()) {
*any_block_in_loop = true;
}
// Disconnect the handlers.
while (block->GetSuccessors().size() > 1) {
HBasicBlock* handler = block->GetSuccessors()[1];
DCHECK(handler->IsCatchBlock());
block->RemoveSuccessor(handler);
handler->RemovePredecessor(block);
if (handler->IsInLoop()) {
*any_block_in_loop = true;
}
}
// Change TryBoundary to Goto.
DCHECK(block->EndsWithTryBoundary());
HInstruction* last = block->GetLastInstruction();
block->RemoveInstruction(last);
block->AddInstruction(new (graph_->GetAllocator()) HGoto(last->GetDexPc()));
DCHECK_EQ(block->GetSuccessors().size(), 1u);
}
void HDeadCodeElimination::RemoveTry(HBasicBlock* try_entry,
const TryBelongingInformation& try_belonging_info,
/* out */ bool* any_block_in_loop) {
// Update all try entries.
DCHECK(try_entry->EndsWithTryBoundary());
DCHECK(try_entry->GetLastInstruction()->AsTryBoundary()->IsEntry());
DisconnectHandlersAndUpdateTryBoundary(try_entry, any_block_in_loop);
const ArenaVector<HBasicBlock*>& blocks = graph_->GetBlocks();
for (uint32_t i : try_belonging_info.coalesced_try_entries.Indexes()) {
HBasicBlock* other_try_entry = blocks[i];
DCHECK(other_try_entry->EndsWithTryBoundary());
DCHECK(other_try_entry->GetLastInstruction()->AsTryBoundary()->IsEntry());
DisconnectHandlersAndUpdateTryBoundary(other_try_entry, any_block_in_loop);
}
// Update the blocks in the try.
for (uint32_t i : try_belonging_info.blocks_in_try.Indexes()) {
HBasicBlock* block = blocks[i];
// Update the try catch information since now the try doesn't exist.
block->SetTryCatchInformation(nullptr);
if (block->IsInLoop()) {
*any_block_in_loop = true;
}
if (block->EndsWithTryBoundary()) {
// Try exits.
DCHECK(!block->GetLastInstruction()->AsTryBoundary()->IsEntry());
DisconnectHandlersAndUpdateTryBoundary(block, any_block_in_loop);
if (block->GetSingleSuccessor()->IsExitBlock()) {
// `block` used to be a single exit TryBoundary that got turned into a Goto. It
// is now pointing to the exit which we don't allow. To fix it, we disconnect
// `block` from its predecessor and RemoveDeadBlocks will remove it from the
// graph.
DCHECK(block->IsSingleGoto());
HBasicBlock* predecessor = block->GetSinglePredecessor();
predecessor->ReplaceSuccessor(block, graph_->GetExitBlock());
if (!block->GetDominatedBlocks().empty()) {
// Update domination tree if `block` dominates a block to keep the graph consistent.
DCHECK_EQ(block->GetDominatedBlocks().size(), 1u);
DCHECK_EQ(graph_->GetExitBlock()->GetDominator(), block);
predecessor->AddDominatedBlock(graph_->GetExitBlock());
graph_->GetExitBlock()->SetDominator(predecessor);
block->RemoveDominatedBlock(graph_->GetExitBlock());
}
}
}
}
}
bool HDeadCodeElimination::RemoveUnneededTries() {
if (!graph_->HasTryCatch()) {
return false;
}
// Use local allocator for allocating memory.
ScopedArenaAllocator allocator(graph_->GetArenaStack());
// Collect which blocks are part of which try.
ScopedArenaUnorderedMap<HBasicBlock*, TryBelongingInformation> tries(
allocator.Adapter(kArenaAllocDCE));
for (HBasicBlock* block : graph_->GetReversePostOrderSkipEntryBlock()) {
if (block->IsTryBlock()) {
HBasicBlock* key = block->GetTryCatchInformation()->GetTryEntry().GetBlock();
auto it = tries.find(key);
if (it == tries.end()) {
it = tries.insert({key, TryBelongingInformation(graph_, &allocator)}).first;
}
it->second.blocks_in_try.SetBit(block->GetBlockId());
}
}
// Deduplicate the tries which have different try entries but they are really the same try.
for (auto it = tries.begin(); it != tries.end(); it++) {
HBasicBlock* block = it->first;
DCHECK(block->EndsWithTryBoundary());
HTryBoundary* try_boundary = block->GetLastInstruction()->AsTryBoundary();
for (auto other_it = next(it); other_it != tries.end(); /*other_it++ in the loop*/) {
HBasicBlock* other_block = other_it->first;
DCHECK(other_block->EndsWithTryBoundary());
HTryBoundary* other_try_boundary = other_block->GetLastInstruction()->AsTryBoundary();
if (try_boundary->HasSameExceptionHandlersAs(*other_try_boundary)) {
// Merge the entries as they are really the same one.
// Block merging.
it->second.blocks_in_try.Union(&other_it->second.blocks_in_try);
// Add the coalesced try entry to update it too.
it->second.coalesced_try_entries.SetBit(other_block->GetBlockId());
// Erase the other entry.
other_it = tries.erase(other_it);
} else {
other_it++;
}
}
}
size_t removed_tries = 0;
bool any_block_in_loop = false;
// Check which tries contain throwing instructions.
for (const auto& entry : tries) {
if (CanPerformTryRemoval(entry.second)) {
++removed_tries;
RemoveTry(entry.first, entry.second, &any_block_in_loop);
}
}
if (removed_tries != 0) {
// We want to:
// 1) Update the dominance information
// 2) Remove catch block subtrees, if they are now unreachable.
// If we run the dominance recomputation without removing the code, those catch blocks will
// not be part of the post order and won't be removed. If we don't run the dominance
// recomputation, we risk RemoveDeadBlocks not running it and leaving the graph in an
// inconsistent state. So, what we can do is run RemoveDeadBlocks and force a recomputation.
// Note that we are not guaranteed to remove a catch block if we have nested try blocks:
//
// try {
// ... nothing can throw. TryBoundary A ...
// try {
// ... can throw. TryBoundary B...
// } catch (Error e) {}
// } catch (Exception e) {}
//
// In the example above, we can remove the TryBoundary A but the Exception catch cannot be
// removed as the TryBoundary B might still throw into that catch. TryBoundary A and B don't get
// coalesced since they have different catch handlers.
RemoveDeadBlocks(/* force_recomputation= */ true, any_block_in_loop);
MaybeRecordStat(stats_, MethodCompilationStat::kRemovedTry, removed_tries);
return true;
} else {
return false;
}
}
bool HDeadCodeElimination::RemoveEmptyIfs() {
bool did_opt = false;
for (HBasicBlock* block : graph_->GetPostOrder()) {
if (!block->EndsWithIf()) {
continue;
}
HIf* if_instr = block->GetLastInstruction()->AsIf();
HBasicBlock* true_block = if_instr->IfTrueSuccessor();
HBasicBlock* false_block = if_instr->IfFalseSuccessor();
// We can use `visited_blocks` to detect cases like
// 1
// / \
// 2 3
// \ /
// 4 ...
// | /
// 5
// where 2, 3, and 4 are single HGoto blocks, and block 5 has Phis.
ScopedArenaAllocator allocator(graph_->GetArenaStack());
ArenaBitVector visited_blocks(
&allocator, graph_->GetBlocks().size(), /*expandable=*/ false, kArenaAllocDCE);
visited_blocks.ClearAllBits();
HBasicBlock* merge_true = true_block;
visited_blocks.SetBit(merge_true->GetBlockId());
while (merge_true->IsSingleGoto()) {
merge_true = merge_true->GetSuccessors()[0];
visited_blocks.SetBit(merge_true->GetBlockId());
}
HBasicBlock* merge_false = false_block;
while (!visited_blocks.IsBitSet(merge_false->GetBlockId()) && merge_false->IsSingleGoto()) {
merge_false = merge_false->GetSuccessors()[0];
}
if (!visited_blocks.IsBitSet(merge_false->GetBlockId()) || !merge_false->GetPhis().IsEmpty()) {
// TODO(solanes): We could allow Phis iff both branches have the same value for all Phis. This
// may not be covered by SsaRedundantPhiElimination in cases like `HPhi[A,A,B]` where the Phi
// itself is not redundant for the general case but it is for a pair of branches.
continue;
}
// Data structures to help remove now-dead instructions.
ScopedArenaQueue<HInstruction*> maybe_remove(allocator.Adapter(kArenaAllocDCE));
ArenaBitVector visited(
&allocator, graph_->GetCurrentInstructionId(), /*expandable=*/ false, kArenaAllocDCE);
visited.ClearAllBits();
maybe_remove.push(if_instr->InputAt(0));
visited.SetBit(if_instr->GetId());
// Swap HIf with HGoto
block->ReplaceAndRemoveInstructionWith(
if_instr, new (graph_->GetAllocator()) HGoto(if_instr->GetDexPc()));
// Reconnect blocks
block->RemoveSuccessor(true_block);
block->RemoveSuccessor(false_block);
true_block->RemovePredecessor(block);
false_block->RemovePredecessor(block);
block->AddSuccessor(merge_false);
// Remove now dead instructions e.g. comparisons that are only used as input to the if
// instruction. This can allow for further removal of other empty ifs.
while (!maybe_remove.empty()) {
HInstruction* instr = maybe_remove.front();
maybe_remove.pop();
if (instr->IsDeadAndRemovable()) {
for (HInstruction* input : instr->GetInputs()) {
if (visited.IsBitSet(input->GetId())) {
continue;
}
visited.SetBit(input->GetId());
maybe_remove.push(input);
}
instr->GetBlock()->RemoveInstructionOrPhi(instr);
MaybeRecordStat(stats_, MethodCompilationStat::kRemovedDeadInstruction);
}
}
did_opt = true;
}
if (did_opt) {
graph_->RecomputeDominatorTree();
}
return did_opt;
}
bool HDeadCodeElimination::RemoveDeadBlocks(bool force_recomputation,
bool force_loop_recomputation) {
DCHECK_IMPLIES(force_loop_recomputation, force_recomputation);
// Use local allocator for allocating memory.
ScopedArenaAllocator allocator(graph_->GetArenaStack());
// Classify blocks as reachable/unreachable.
ArenaBitVector live_blocks(&allocator, graph_->GetBlocks().size(), false, kArenaAllocDCE);
live_blocks.ClearAllBits();
MarkReachableBlocks(graph_, &live_blocks);
bool removed_one_or_more_blocks = false;
bool rerun_dominance_and_loop_analysis = false;
// Remove all dead blocks. Iterate in post order because removal needs the
// block's chain of dominators and nested loops need to be updated from the
// inside out.
for (HBasicBlock* block : graph_->GetPostOrder()) {
int id = block->GetBlockId();
if (!live_blocks.IsBitSet(id)) {
MaybeRecordDeadBlock(block);
block->DisconnectAndDelete();
removed_one_or_more_blocks = true;
if (block->IsInLoop()) {
rerun_dominance_and_loop_analysis = true;
}
}
}
// If we removed at least one block, we need to recompute the full
// dominator tree and try block membership.
if (removed_one_or_more_blocks || force_recomputation) {
if (rerun_dominance_and_loop_analysis || force_loop_recomputation) {
graph_->RecomputeDominatorTree();
} else {
graph_->ClearDominanceInformation();
graph_->ComputeDominanceInformation();
graph_->ComputeTryBlockInformation();
}
}
return removed_one_or_more_blocks;
}
void HDeadCodeElimination::RemoveDeadInstructions() {
// Process basic blocks in post-order in the dominator tree, so that
// a dead instruction depending on another dead instruction is removed.
for (HBasicBlock* block : graph_->GetPostOrder()) {
// Traverse this block's instructions in backward order and remove
// the unused ones.
HBackwardInstructionIterator i(block->GetInstructions());
// Skip the first iteration, as the last instruction of a block is
// a branching instruction.
DCHECK(i.Current()->IsControlFlow());
for (i.Advance(); !i.Done(); i.Advance()) {
HInstruction* inst = i.Current();
DCHECK(!inst->IsControlFlow());
if (inst->IsDeadAndRemovable()) {
block->RemoveInstruction(inst);
MaybeRecordStat(stats_, MethodCompilationStat::kRemovedDeadInstruction);
}
}
// Same for Phis.
for (HBackwardInstructionIterator phi_it(block->GetPhis()); !phi_it.Done(); phi_it.Advance()) {
DCHECK(phi_it.Current()->IsPhi());
HPhi* phi = phi_it.Current()->AsPhi();
if (phi->IsPhiDeadAndRemovable()) {
block->RemovePhi(phi);
MaybeRecordStat(stats_, MethodCompilationStat::kRemovedDeadPhi);
}
}
}
}
void HDeadCodeElimination::UpdateGraphFlags() {
bool has_monitor_operations = false;
bool has_traditional_simd = false;
bool has_predicated_simd = false;
bool has_bounds_checks = false;
bool has_always_throwing_invokes = false;
for (HBasicBlock* block : graph_->GetReversePostOrder()) {
for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
HInstruction* instruction = it.Current();
if (instruction->IsMonitorOperation()) {
has_monitor_operations = true;
} else if (instruction->IsVecOperation()) {
HVecOperation* vec_instruction = instruction->AsVecOperation();
if (vec_instruction->IsPredicated()) {
has_predicated_simd = true;
} else {
has_traditional_simd = true;
}
} else if (instruction->IsBoundsCheck()) {
has_bounds_checks = true;
} else if (instruction->IsInvoke() && instruction->AsInvoke()->AlwaysThrows()) {
has_always_throwing_invokes = true;
}
}
}
graph_->SetHasMonitorOperations(has_monitor_operations);
graph_->SetHasTraditionalSIMD(has_traditional_simd);
graph_->SetHasPredicatedSIMD(has_predicated_simd);
graph_->SetHasBoundsChecks(has_bounds_checks);
graph_->SetHasAlwaysThrowingInvokes(has_always_throwing_invokes);
}
bool HDeadCodeElimination::Run() {
// Do not eliminate dead blocks if the graph has irreducible loops. We could
// support it, but that would require changes in our loop representation to handle
// multiple entry points. We decided it was not worth the complexity.
if (!graph_->HasIrreducibleLoops()) {
// Simplify graph to generate more dead block patterns.
ConnectSuccessiveBlocks();
bool did_any_simplification = false;
did_any_simplification |= SimplifyAlwaysThrows();
did_any_simplification |= SimplifyIfs();
did_any_simplification |= RemoveEmptyIfs();
did_any_simplification |= RemoveDeadBlocks();
// We call RemoveDeadBlocks before RemoveUnneededTries to remove the dead blocks from the
// previous optimizations. Otherwise, we might detect that a try has throwing instructions but
// they are actually dead code. RemoveUnneededTryBoundary will call RemoveDeadBlocks again if
// needed.
did_any_simplification |= RemoveUnneededTries();
if (did_any_simplification) {
// Connect successive blocks created by dead branches.
ConnectSuccessiveBlocks();
}
}
SsaRedundantPhiElimination(graph_).Run();
RemoveDeadInstructions();
UpdateGraphFlags();
return true;
}
} // namespace art