/* * Copyright (C) 2013 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 "base/stringprintf.h" #include "sea_ir/ir/instruction_tools.h" #include "sea_ir/ir/sea.h" #include "sea_ir/code_gen/code_gen.h" #include "sea_ir/types/type_inference.h" #define MAX_REACHING_DEF_ITERERATIONS (10) // TODO: When development is done, this define should not // be needed, it is currently used as a cutoff // for cases where the iterative fixed point algorithm // does not reach a fixed point because of a bug. namespace sea_ir { int SeaNode::current_max_node_id_ = 0; void IRVisitor::Traverse(Region* region) { std::vector<PhiInstructionNode*>* phis = region->GetPhiNodes(); for (std::vector<PhiInstructionNode*>::const_iterator cit = phis->begin(); cit != phis->end(); cit++) { (*cit)->Accept(this); } std::vector<InstructionNode*>* instructions = region->GetInstructions(); for (std::vector<InstructionNode*>::const_iterator cit = instructions->begin(); cit != instructions->end(); cit++) { (*cit)->Accept(this); } } void IRVisitor::Traverse(SeaGraph* graph) { for (std::vector<Region*>::const_iterator cit = ordered_regions_.begin(); cit != ordered_regions_.end(); cit++ ) { (*cit)->Accept(this); } } SeaGraph* SeaGraph::GetGraph(const art::DexFile& dex_file) { return new SeaGraph(dex_file); } void SeaGraph::AddEdge(Region* src, Region* dst) const { src->AddSuccessor(dst); dst->AddPredecessor(src); } void SeaGraph::ComputeRPO(Region* current_region, int& current_rpo) { current_region->SetRPO(VISITING); std::vector<sea_ir::Region*>* succs = current_region->GetSuccessors(); for (std::vector<sea_ir::Region*>::iterator succ_it = succs->begin(); succ_it != succs->end(); ++succ_it) { if (NOT_VISITED == (*succ_it)->GetRPO()) { SeaGraph::ComputeRPO(*succ_it, current_rpo); } } current_region->SetRPO(current_rpo--); } void SeaGraph::ComputeIDominators() { bool changed = true; while (changed) { changed = false; // Entry node has itself as IDOM. std::vector<Region*>::iterator crt_it; std::set<Region*> processedNodes; // Find and mark the entry node(s). for (crt_it = regions_.begin(); crt_it != regions_.end(); ++crt_it) { if ((*crt_it)->GetPredecessors()->size() == 0) { processedNodes.insert(*crt_it); (*crt_it)->SetIDominator(*crt_it); } } for (crt_it = regions_.begin(); crt_it != regions_.end(); ++crt_it) { if ((*crt_it)->GetPredecessors()->size() == 0) { continue; } // NewIDom = first (processed) predecessor of b. Region* new_dom = NULL; std::vector<Region*>* preds = (*crt_it)->GetPredecessors(); DCHECK(NULL != preds); Region* root_pred = NULL; for (std::vector<Region*>::iterator pred_it = preds->begin(); pred_it != preds->end(); ++pred_it) { if (processedNodes.end() != processedNodes.find((*pred_it))) { root_pred = *pred_it; new_dom = root_pred; break; } } // For all other predecessors p of b, if idom is not set, // then NewIdom = Intersect(p, NewIdom) for (std::vector<Region*>::const_iterator pred_it = preds->begin(); pred_it != preds->end(); ++pred_it) { DCHECK(NULL != *pred_it); // if IDOMS[p] != UNDEFINED if ((*pred_it != root_pred) && (*pred_it)->GetIDominator() != NULL) { DCHECK(NULL != new_dom); new_dom = SeaGraph::Intersect(*pred_it, new_dom); } } DCHECK(NULL != *crt_it); if ((*crt_it)->GetIDominator() != new_dom) { (*crt_it)->SetIDominator(new_dom); changed = true; } processedNodes.insert(*crt_it); } } // For easily ordering of regions we need edges dominator->dominated. for (std::vector<Region*>::iterator region_it = regions_.begin(); region_it != regions_.end(); region_it++) { Region* idom = (*region_it)->GetIDominator(); if (idom != *region_it) { idom->AddToIDominatedSet(*region_it); } } } Region* SeaGraph::Intersect(Region* i, Region* j) { Region* finger1 = i; Region* finger2 = j; while (finger1 != finger2) { while (finger1->GetRPO() > finger2->GetRPO()) { DCHECK(NULL != finger1); finger1 = finger1->GetIDominator(); // should have: finger1 != NULL DCHECK(NULL != finger1); } while (finger1->GetRPO() < finger2->GetRPO()) { DCHECK(NULL != finger2); finger2 = finger2->GetIDominator(); // should have: finger1 != NULL DCHECK(NULL != finger2); } } return finger1; // finger1 should be equal to finger2 at this point. } void SeaGraph::ComputeDownExposedDefs() { for (std::vector<Region*>::iterator region_it = regions_.begin(); region_it != regions_.end(); region_it++) { (*region_it)->ComputeDownExposedDefs(); } } void SeaGraph::ComputeReachingDefs() { // Iterate until the reaching definitions set doesn't change anymore. // (See Cooper & Torczon, "Engineering a Compiler", second edition, page 487) bool changed = true; int iteration = 0; while (changed && (iteration < MAX_REACHING_DEF_ITERERATIONS)) { iteration++; changed = false; // TODO: optimize the ordering if this becomes performance bottleneck. for (std::vector<Region*>::iterator regions_it = regions_.begin(); regions_it != regions_.end(); regions_it++) { changed |= (*regions_it)->UpdateReachingDefs(); } } DCHECK(!changed) << "Reaching definitions computation did not reach a fixed point."; } void SeaGraph::InsertSignatureNodes(const art::DexFile::CodeItem* code_item, Region* r) { // Insert a fake SignatureNode for the first parameter. // TODO: Provide a register enum value for the fake parameter. SignatureNode* parameter_def_node = new sea_ir::SignatureNode(0, 0); AddParameterNode(parameter_def_node); r->AddChild(parameter_def_node); // Insert SignatureNodes for each Dalvik register parameter. for (unsigned int crt_offset = 0; crt_offset < code_item->ins_size_; crt_offset++) { int register_no = code_item->registers_size_ - crt_offset - 1; int position = crt_offset + 1; SignatureNode* parameter_def_node = new sea_ir::SignatureNode(register_no, position); AddParameterNode(parameter_def_node); r->AddChild(parameter_def_node); } } void SeaGraph::BuildMethodSeaGraph(const art::DexFile::CodeItem* code_item, const art::DexFile& dex_file, uint16_t class_def_idx, uint32_t method_idx, uint32_t method_access_flags) { code_item_ = code_item; class_def_idx_ = class_def_idx; method_idx_ = method_idx; method_access_flags_ = method_access_flags; const uint16_t* code = code_item->insns_; const size_t size_in_code_units = code_item->insns_size_in_code_units_; // This maps target instruction pointers to their corresponding region objects. std::map<const uint16_t*, Region*> target_regions; size_t i = 0; // Pass: Find the start instruction of basic blocks // by locating targets and flow-though instructions of branches. while (i < size_in_code_units) { const art::Instruction* inst = art::Instruction::At(&code[i]); if (inst->IsBranch() || inst->IsUnconditional()) { int32_t offset = inst->GetTargetOffset(); if (target_regions.end() == target_regions.find(&code[i + offset])) { Region* region = GetNewRegion(); target_regions.insert(std::pair<const uint16_t*, Region*>(&code[i + offset], region)); } if (inst->CanFlowThrough() && (target_regions.end() == target_regions.find(&code[i + inst->SizeInCodeUnits()]))) { Region* region = GetNewRegion(); target_regions.insert( std::pair<const uint16_t*, Region*>(&code[i + inst->SizeInCodeUnits()], region)); } } i += inst->SizeInCodeUnits(); } Region* r = GetNewRegion(); InsertSignatureNodes(code_item, r); // Pass: Assign instructions to region nodes and // assign branches their control flow successors. i = 0; sea_ir::InstructionNode* last_node = NULL; sea_ir::InstructionNode* node = NULL; while (i < size_in_code_units) { const art::Instruction* inst = art::Instruction::At(&code[i]); std::vector<InstructionNode*> sea_instructions_for_dalvik = sea_ir::InstructionNode::Create(inst); for (std::vector<InstructionNode*>::const_iterator cit = sea_instructions_for_dalvik.begin(); sea_instructions_for_dalvik.end() != cit; ++cit) { last_node = node; node = *cit; if (inst->IsBranch() || inst->IsUnconditional()) { int32_t offset = inst->GetTargetOffset(); std::map<const uint16_t*, Region*>::iterator it = target_regions.find(&code[i + offset]); DCHECK(it != target_regions.end()); AddEdge(r, it->second); // Add edge to branch target. } std::map<const uint16_t*, Region*>::iterator it = target_regions.find(&code[i]); if (target_regions.end() != it) { // Get the already created region because this is a branch target. Region* nextRegion = it->second; if (last_node->GetInstruction()->IsBranch() && last_node->GetInstruction()->CanFlowThrough()) { AddEdge(r, it->second); // Add flow-through edge. } r = nextRegion; } r->AddChild(node); } i += inst->SizeInCodeUnits(); } } void SeaGraph::ComputeRPO() { int rpo_id = regions_.size() - 1; for (std::vector<Region*>::const_iterator crt_it = regions_.begin(); crt_it != regions_.end(); ++crt_it) { if ((*crt_it)->GetPredecessors()->size() == 0) { ComputeRPO(*crt_it, rpo_id); } } } // Performs the renaming phase in traditional SSA transformations. // See: Cooper & Torczon, "Engineering a Compiler", second edition, page 505.) void SeaGraph::RenameAsSSA() { utils::ScopedHashtable<int, InstructionNode*> scoped_table; scoped_table.OpenScope(); for (std::vector<Region*>::iterator region_it = regions_.begin(); region_it != regions_.end(); region_it++) { if ((*region_it)->GetIDominator() == *region_it) { RenameAsSSA(*region_it, &scoped_table); } } scoped_table.CloseScope(); } void SeaGraph::ConvertToSSA() { // Pass: find global names. // The map @block maps registers to the blocks in which they are defined. std::map<int, std::set<Region*> > blocks; // The set @globals records registers whose use // is in a different block than the corresponding definition. std::set<int> globals; for (std::vector<Region*>::iterator region_it = regions_.begin(); region_it != regions_.end(); region_it++) { std::set<int> var_kill; std::vector<InstructionNode*>* instructions = (*region_it)->GetInstructions(); for (std::vector<InstructionNode*>::iterator inst_it = instructions->begin(); inst_it != instructions->end(); inst_it++) { std::vector<int> used_regs = (*inst_it)->GetUses(); for (std::size_t i = 0; i < used_regs.size(); i++) { int used_reg = used_regs[i]; if (var_kill.find(used_reg) == var_kill.end()) { globals.insert(used_reg); } } const int reg_def = (*inst_it)->GetResultRegister(); if (reg_def != NO_REGISTER) { var_kill.insert(reg_def); } blocks.insert(std::pair<int, std::set<Region*> >(reg_def, std::set<Region*>())); std::set<Region*>* reg_def_blocks = &(blocks.find(reg_def)->second); reg_def_blocks->insert(*region_it); } } // Pass: Actually add phi-nodes to regions. for (std::set<int>::const_iterator globals_it = globals.begin(); globals_it != globals.end(); globals_it++) { int global = *globals_it; // Copy the set, because we will modify the worklist as we go. std::set<Region*> worklist((*(blocks.find(global))).second); for (std::set<Region*>::const_iterator b_it = worklist.begin(); b_it != worklist.end(); b_it++) { std::set<Region*>* df = (*b_it)->GetDominanceFrontier(); for (std::set<Region*>::const_iterator df_it = df->begin(); df_it != df->end(); df_it++) { if ((*df_it)->InsertPhiFor(global)) { // Check that the dominance frontier element is in the worklist already // because we only want to break if the element is actually not there yet. if (worklist.find(*df_it) == worklist.end()) { worklist.insert(*df_it); b_it = worklist.begin(); break; } } } } } // Pass: Build edges to the definition corresponding to each use. // (This corresponds to the renaming phase in traditional SSA transformations. // See: Cooper & Torczon, "Engineering a Compiler", second edition, page 505.) RenameAsSSA(); } void SeaGraph::RenameAsSSA(Region* crt_region, utils::ScopedHashtable<int, InstructionNode*>* scoped_table) { scoped_table->OpenScope(); // Rename phi nodes defined in the current region. std::vector<PhiInstructionNode*>* phis = crt_region->GetPhiNodes(); for (std::vector<PhiInstructionNode*>::iterator phi_it = phis->begin(); phi_it != phis->end(); phi_it++) { int reg_no = (*phi_it)->GetRegisterNumber(); scoped_table->Add(reg_no, (*phi_it)); } // Rename operands of instructions from the current region. std::vector<InstructionNode*>* instructions = crt_region->GetInstructions(); for (std::vector<InstructionNode*>::const_iterator instructions_it = instructions->begin(); instructions_it != instructions->end(); instructions_it++) { InstructionNode* current_instruction = (*instructions_it); // Rename uses. std::vector<int> used_regs = current_instruction->GetUses(); for (std::vector<int>::const_iterator reg_it = used_regs.begin(); reg_it != used_regs.end(); reg_it++) { int current_used_reg = (*reg_it); InstructionNode* definition = scoped_table->Lookup(current_used_reg); current_instruction->RenameToSSA(current_used_reg, definition); } // Update scope table with latest definitions. std::vector<int> def_regs = current_instruction->GetDefinitions(); for (std::vector<int>::const_iterator reg_it = def_regs.begin(); reg_it != def_regs.end(); reg_it++) { int current_defined_reg = (*reg_it); scoped_table->Add(current_defined_reg, current_instruction); } } // Fill in uses of phi functions in CFG successor regions. const std::vector<Region*>* successors = crt_region->GetSuccessors(); for (std::vector<Region*>::const_iterator successors_it = successors->begin(); successors_it != successors->end(); successors_it++) { Region* successor = (*successors_it); successor->SetPhiDefinitionsForUses(scoped_table, crt_region); } // Rename all successors in the dominators tree. const std::set<Region*>* dominated_nodes = crt_region->GetIDominatedSet(); for (std::set<Region*>::const_iterator dominated_nodes_it = dominated_nodes->begin(); dominated_nodes_it != dominated_nodes->end(); dominated_nodes_it++) { Region* dominated_node = (*dominated_nodes_it); RenameAsSSA(dominated_node, scoped_table); } scoped_table->CloseScope(); } CodeGenData* SeaGraph::GenerateLLVM(const std::string& function_name, const art::DexFile& dex_file) { // Pass: Generate LLVM IR. CodeGenPrepassVisitor code_gen_prepass_visitor(function_name); std::cout << "Generating code..." << std::endl; Accept(&code_gen_prepass_visitor); CodeGenVisitor code_gen_visitor(code_gen_prepass_visitor.GetData(), dex_file); Accept(&code_gen_visitor); CodeGenPostpassVisitor code_gen_postpass_visitor(code_gen_visitor.GetData()); Accept(&code_gen_postpass_visitor); return code_gen_postpass_visitor.GetData(); } CodeGenData* SeaGraph::CompileMethod( const std::string& function_name, const art::DexFile::CodeItem* code_item, uint16_t class_def_idx, uint32_t method_idx, uint32_t method_access_flags, const art::DexFile& dex_file) { // Two passes: Builds the intermediate structure (non-SSA) of the sea-ir for the function. BuildMethodSeaGraph(code_item, dex_file, class_def_idx, method_idx, method_access_flags); // Pass: Compute reverse post-order of regions. ComputeRPO(); // Multiple passes: compute immediate dominators. ComputeIDominators(); // Pass: compute downward-exposed definitions. ComputeDownExposedDefs(); // Multiple Passes (iterative fixed-point algorithm): Compute reaching definitions ComputeReachingDefs(); // Pass (O(nlogN)): Compute the dominance frontier for region nodes. ComputeDominanceFrontier(); // Two Passes: Phi node insertion. ConvertToSSA(); // Pass: type inference ti_->ComputeTypes(this); // Pass: Generate LLVM IR. CodeGenData* cgd = GenerateLLVM(function_name, dex_file); return cgd; } void SeaGraph::ComputeDominanceFrontier() { for (std::vector<Region*>::iterator region_it = regions_.begin(); region_it != regions_.end(); region_it++) { std::vector<Region*>* preds = (*region_it)->GetPredecessors(); if (preds->size() > 1) { for (std::vector<Region*>::iterator pred_it = preds->begin(); pred_it != preds->end(); pred_it++) { Region* runner = *pred_it; while (runner != (*region_it)->GetIDominator()) { runner->AddToDominanceFrontier(*region_it); runner = runner->GetIDominator(); } } } } } Region* SeaGraph::GetNewRegion() { Region* new_region = new Region(); AddRegion(new_region); return new_region; } void SeaGraph::AddRegion(Region* r) { DCHECK(r) << "Tried to add NULL region to SEA graph."; regions_.push_back(r); } SeaGraph::SeaGraph(const art::DexFile& df) :ti_(new TypeInference()), class_def_idx_(0), method_idx_(0), method_access_flags_(), regions_(), parameters_(), dex_file_(df), code_item_(NULL) { } void Region::AddChild(sea_ir::InstructionNode* instruction) { DCHECK(instruction) << "Tried to add NULL instruction to region node."; instructions_.push_back(instruction); instruction->SetRegion(this); } SeaNode* Region::GetLastChild() const { if (instructions_.size() > 0) { return instructions_.back(); } return NULL; } void Region::ComputeDownExposedDefs() { for (std::vector<InstructionNode*>::const_iterator inst_it = instructions_.begin(); inst_it != instructions_.end(); inst_it++) { int reg_no = (*inst_it)->GetResultRegister(); std::map<int, InstructionNode*>::iterator res = de_defs_.find(reg_no); if ((reg_no != NO_REGISTER) && (res == de_defs_.end())) { de_defs_.insert(std::pair<int, InstructionNode*>(reg_no, *inst_it)); } else { res->second = *inst_it; } } for (std::map<int, sea_ir::InstructionNode*>::const_iterator cit = de_defs_.begin(); cit != de_defs_.end(); cit++) { (*cit).second->MarkAsDEDef(); } } const std::map<int, sea_ir::InstructionNode*>* Region::GetDownExposedDefs() const { return &de_defs_; } std::map<int, std::set<sea_ir::InstructionNode*>* >* Region::GetReachingDefs() { return &reaching_defs_; } bool Region::UpdateReachingDefs() { std::map<int, std::set<sea_ir::InstructionNode*>* > new_reaching; for (std::vector<Region*>::const_iterator pred_it = predecessors_.begin(); pred_it != predecessors_.end(); pred_it++) { // The reaching_defs variable will contain reaching defs __for current predecessor only__ std::map<int, std::set<sea_ir::InstructionNode*>* > reaching_defs; std::map<int, std::set<sea_ir::InstructionNode*>* >* pred_reaching = (*pred_it)->GetReachingDefs(); const std::map<int, InstructionNode*>* de_defs = (*pred_it)->GetDownExposedDefs(); // The definitions from the reaching set of the predecessor // may be shadowed by downward exposed definitions from the predecessor, // otherwise the defs from the reaching set are still good. for (std::map<int, InstructionNode*>::const_iterator de_def = de_defs->begin(); de_def != de_defs->end(); de_def++) { std::set<InstructionNode*>* solo_def; solo_def = new std::set<InstructionNode*>(); solo_def->insert(de_def->second); reaching_defs.insert( std::pair<int const, std::set<InstructionNode*>*>(de_def->first, solo_def)); } reaching_defs.insert(pred_reaching->begin(), pred_reaching->end()); // Now we combine the reaching map coming from the current predecessor (reaching_defs) // with the accumulated set from all predecessors so far (from new_reaching). std::map<int, std::set<sea_ir::InstructionNode*>*>::iterator reaching_it = reaching_defs.begin(); for (; reaching_it != reaching_defs.end(); reaching_it++) { std::map<int, std::set<sea_ir::InstructionNode*>*>::iterator crt_entry = new_reaching.find(reaching_it->first); if (new_reaching.end() != crt_entry) { crt_entry->second->insert(reaching_it->second->begin(), reaching_it->second->end()); } else { new_reaching.insert( std::pair<int, std::set<sea_ir::InstructionNode*>*>( reaching_it->first, reaching_it->second) ); } } } bool changed = false; // Because the sets are monotonically increasing, // we can compare sizes instead of using set comparison. // TODO: Find formal proof. int old_size = 0; if (-1 == reaching_defs_size_) { std::map<int, std::set<sea_ir::InstructionNode*>*>::iterator reaching_it = reaching_defs_.begin(); for (; reaching_it != reaching_defs_.end(); reaching_it++) { old_size += (*reaching_it).second->size(); } } else { old_size = reaching_defs_size_; } int new_size = 0; std::map<int, std::set<sea_ir::InstructionNode*>*>::iterator reaching_it = new_reaching.begin(); for (; reaching_it != new_reaching.end(); reaching_it++) { new_size += (*reaching_it).second->size(); } if (old_size != new_size) { changed = true; } if (changed) { reaching_defs_ = new_reaching; reaching_defs_size_ = new_size; } return changed; } bool Region::InsertPhiFor(int reg_no) { if (!ContainsPhiFor(reg_no)) { phi_set_.insert(reg_no); PhiInstructionNode* new_phi = new PhiInstructionNode(reg_no); new_phi->SetRegion(this); phi_instructions_.push_back(new_phi); return true; } return false; } void Region::SetPhiDefinitionsForUses( const utils::ScopedHashtable<int, InstructionNode*>* scoped_table, Region* predecessor) { int predecessor_id = -1; for (unsigned int crt_pred_id = 0; crt_pred_id < predecessors_.size(); crt_pred_id++) { if (predecessors_.at(crt_pred_id) == predecessor) { predecessor_id = crt_pred_id; } } DCHECK_NE(-1, predecessor_id); for (std::vector<PhiInstructionNode*>::iterator phi_it = phi_instructions_.begin(); phi_it != phi_instructions_.end(); phi_it++) { PhiInstructionNode* phi = (*phi_it); int reg_no = phi->GetRegisterNumber(); InstructionNode* definition = scoped_table->Lookup(reg_no); phi->RenameToSSA(reg_no, definition, predecessor_id); } } std::vector<InstructionNode*> InstructionNode::Create(const art::Instruction* in) { std::vector<InstructionNode*> sea_instructions; switch (in->Opcode()) { case art::Instruction::CONST_4: sea_instructions.push_back(new ConstInstructionNode(in)); break; case art::Instruction::RETURN: sea_instructions.push_back(new ReturnInstructionNode(in)); break; case art::Instruction::IF_NE: sea_instructions.push_back(new IfNeInstructionNode(in)); break; case art::Instruction::ADD_INT_LIT8: sea_instructions.push_back(new UnnamedConstInstructionNode(in, in->VRegC_22b())); sea_instructions.push_back(new AddIntLitInstructionNode(in)); break; case art::Instruction::MOVE_RESULT: sea_instructions.push_back(new MoveResultInstructionNode(in)); break; case art::Instruction::INVOKE_STATIC: sea_instructions.push_back(new InvokeStaticInstructionNode(in)); break; case art::Instruction::ADD_INT: sea_instructions.push_back(new AddIntInstructionNode(in)); break; case art::Instruction::GOTO: sea_instructions.push_back(new GotoInstructionNode(in)); break; case art::Instruction::IF_EQZ: sea_instructions.push_back(new IfEqzInstructionNode(in)); break; default: // Default, generic IR instruction node; default case should never be reached // when support for all instructions ahs been added. sea_instructions.push_back(new InstructionNode(in)); } return sea_instructions; } void InstructionNode::MarkAsDEDef() { de_def_ = true; } int InstructionNode::GetResultRegister() const { if (instruction_->HasVRegA() && InstructionTools::IsDefinition(instruction_)) { return instruction_->VRegA(); } return NO_REGISTER; } std::vector<int> InstructionNode::GetDefinitions() const { // TODO: Extend this to handle instructions defining more than one register (if any) // The return value should be changed to pointer to field then; for now it is an object // so that we avoid possible memory leaks from allocating objects dynamically. std::vector<int> definitions; int result = GetResultRegister(); if (NO_REGISTER != result) { definitions.push_back(result); } return definitions; } std::vector<int> InstructionNode::GetUses() const { std::vector<int> uses; // Using vector<> instead of set<> because order matters. if (!InstructionTools::IsDefinition(instruction_) && (instruction_->HasVRegA())) { int vA = instruction_->VRegA(); uses.push_back(vA); } if (instruction_->HasVRegB()) { int vB = instruction_->VRegB(); uses.push_back(vB); } if (instruction_->HasVRegC()) { int vC = instruction_->VRegC(); uses.push_back(vC); } return uses; } } // namespace sea_ir