mir_optimization.cc revision 1fd4821f6b3ac57a44c2ce91025686da4641d197
1/* 2 * Copyright (C) 2011 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include "compiler_internals.h" 18#include "global_value_numbering.h" 19#include "local_value_numbering.h" 20#include "dataflow_iterator-inl.h" 21#include "dex/global_value_numbering.h" 22#include "dex/quick/dex_file_method_inliner.h" 23#include "dex/quick/dex_file_to_method_inliner_map.h" 24#include "utils/scoped_arena_containers.h" 25 26namespace art { 27 28static unsigned int Predecessors(BasicBlock* bb) { 29 return bb->predecessors->Size(); 30} 31 32/* Setup a constant value for opcodes thare have the DF_SETS_CONST attribute */ 33void MIRGraph::SetConstant(int32_t ssa_reg, int value) { 34 is_constant_v_->SetBit(ssa_reg); 35 constant_values_[ssa_reg] = value; 36} 37 38void MIRGraph::SetConstantWide(int ssa_reg, int64_t value) { 39 is_constant_v_->SetBit(ssa_reg); 40 constant_values_[ssa_reg] = Low32Bits(value); 41 constant_values_[ssa_reg + 1] = High32Bits(value); 42} 43 44void MIRGraph::DoConstantPropagation(BasicBlock* bb) { 45 MIR* mir; 46 47 for (mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 48 // Skip pass if BB has MIR without SSA representation. 49 if (mir->ssa_rep == nullptr) { 50 return; 51 } 52 53 uint64_t df_attributes = GetDataFlowAttributes(mir); 54 55 MIR::DecodedInstruction* d_insn = &mir->dalvikInsn; 56 57 if (!(df_attributes & DF_HAS_DEFS)) continue; 58 59 /* Handle instructions that set up constants directly */ 60 if (df_attributes & DF_SETS_CONST) { 61 if (df_attributes & DF_DA) { 62 int32_t vB = static_cast<int32_t>(d_insn->vB); 63 switch (d_insn->opcode) { 64 case Instruction::CONST_4: 65 case Instruction::CONST_16: 66 case Instruction::CONST: 67 SetConstant(mir->ssa_rep->defs[0], vB); 68 break; 69 case Instruction::CONST_HIGH16: 70 SetConstant(mir->ssa_rep->defs[0], vB << 16); 71 break; 72 case Instruction::CONST_WIDE_16: 73 case Instruction::CONST_WIDE_32: 74 SetConstantWide(mir->ssa_rep->defs[0], static_cast<int64_t>(vB)); 75 break; 76 case Instruction::CONST_WIDE: 77 SetConstantWide(mir->ssa_rep->defs[0], d_insn->vB_wide); 78 break; 79 case Instruction::CONST_WIDE_HIGH16: 80 SetConstantWide(mir->ssa_rep->defs[0], static_cast<int64_t>(vB) << 48); 81 break; 82 default: 83 break; 84 } 85 } 86 /* Handle instructions that set up constants directly */ 87 } else if (df_attributes & DF_IS_MOVE) { 88 int i; 89 90 for (i = 0; i < mir->ssa_rep->num_uses; i++) { 91 if (!is_constant_v_->IsBitSet(mir->ssa_rep->uses[i])) break; 92 } 93 /* Move a register holding a constant to another register */ 94 if (i == mir->ssa_rep->num_uses) { 95 SetConstant(mir->ssa_rep->defs[0], constant_values_[mir->ssa_rep->uses[0]]); 96 if (df_attributes & DF_A_WIDE) { 97 SetConstant(mir->ssa_rep->defs[1], constant_values_[mir->ssa_rep->uses[1]]); 98 } 99 } 100 } 101 } 102 /* TODO: implement code to handle arithmetic operations */ 103} 104 105/* Advance to next strictly dominated MIR node in an extended basic block */ 106MIR* MIRGraph::AdvanceMIR(BasicBlock** p_bb, MIR* mir) { 107 BasicBlock* bb = *p_bb; 108 if (mir != NULL) { 109 mir = mir->next; 110 if (mir == NULL) { 111 bb = GetBasicBlock(bb->fall_through); 112 if ((bb == NULL) || Predecessors(bb) != 1) { 113 mir = NULL; 114 } else { 115 *p_bb = bb; 116 mir = bb->first_mir_insn; 117 } 118 } 119 } 120 return mir; 121} 122 123/* 124 * To be used at an invoke mir. If the logically next mir node represents 125 * a move-result, return it. Else, return NULL. If a move-result exists, 126 * it is required to immediately follow the invoke with no intervening 127 * opcodes or incoming arcs. However, if the result of the invoke is not 128 * used, a move-result may not be present. 129 */ 130MIR* MIRGraph::FindMoveResult(BasicBlock* bb, MIR* mir) { 131 BasicBlock* tbb = bb; 132 mir = AdvanceMIR(&tbb, mir); 133 while (mir != NULL) { 134 if ((mir->dalvikInsn.opcode == Instruction::MOVE_RESULT) || 135 (mir->dalvikInsn.opcode == Instruction::MOVE_RESULT_OBJECT) || 136 (mir->dalvikInsn.opcode == Instruction::MOVE_RESULT_WIDE)) { 137 break; 138 } 139 // Keep going if pseudo op, otherwise terminate 140 if (MIR::DecodedInstruction::IsPseudoMirOp(mir->dalvikInsn.opcode)) { 141 mir = AdvanceMIR(&tbb, mir); 142 } else { 143 mir = NULL; 144 } 145 } 146 return mir; 147} 148 149BasicBlock* MIRGraph::NextDominatedBlock(BasicBlock* bb) { 150 if (bb->block_type == kDead) { 151 return NULL; 152 } 153 DCHECK((bb->block_type == kEntryBlock) || (bb->block_type == kDalvikByteCode) 154 || (bb->block_type == kExitBlock)); 155 BasicBlock* bb_taken = GetBasicBlock(bb->taken); 156 BasicBlock* bb_fall_through = GetBasicBlock(bb->fall_through); 157 if (((bb_fall_through == NULL) && (bb_taken != NULL)) && 158 ((bb_taken->block_type == kDalvikByteCode) || (bb_taken->block_type == kExitBlock))) { 159 // Follow simple unconditional branches. 160 bb = bb_taken; 161 } else { 162 // Follow simple fallthrough 163 bb = (bb_taken != NULL) ? NULL : bb_fall_through; 164 } 165 if (bb == NULL || (Predecessors(bb) != 1)) { 166 return NULL; 167 } 168 DCHECK((bb->block_type == kDalvikByteCode) || (bb->block_type == kExitBlock)); 169 return bb; 170} 171 172static MIR* FindPhi(BasicBlock* bb, int ssa_name) { 173 for (MIR* mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 174 if (static_cast<int>(mir->dalvikInsn.opcode) == kMirOpPhi) { 175 for (int i = 0; i < mir->ssa_rep->num_uses; i++) { 176 if (mir->ssa_rep->uses[i] == ssa_name) { 177 return mir; 178 } 179 } 180 } 181 } 182 return NULL; 183} 184 185static SelectInstructionKind SelectKind(MIR* mir) { 186 switch (mir->dalvikInsn.opcode) { 187 case Instruction::MOVE: 188 case Instruction::MOVE_OBJECT: 189 case Instruction::MOVE_16: 190 case Instruction::MOVE_OBJECT_16: 191 case Instruction::MOVE_FROM16: 192 case Instruction::MOVE_OBJECT_FROM16: 193 return kSelectMove; 194 case Instruction::CONST: 195 case Instruction::CONST_4: 196 case Instruction::CONST_16: 197 return kSelectConst; 198 case Instruction::GOTO: 199 case Instruction::GOTO_16: 200 case Instruction::GOTO_32: 201 return kSelectGoto; 202 default: 203 return kSelectNone; 204 } 205} 206 207static constexpr ConditionCode kIfCcZConditionCodes[] = { 208 kCondEq, kCondNe, kCondLt, kCondGe, kCondGt, kCondLe 209}; 210 211COMPILE_ASSERT(arraysize(kIfCcZConditionCodes) == Instruction::IF_LEZ - Instruction::IF_EQZ + 1, 212 if_ccz_ccodes_size1); 213 214static constexpr bool IsInstructionIfCcZ(Instruction::Code opcode) { 215 return Instruction::IF_EQZ <= opcode && opcode <= Instruction::IF_LEZ; 216} 217 218static constexpr ConditionCode ConditionCodeForIfCcZ(Instruction::Code opcode) { 219 return kIfCcZConditionCodes[opcode - Instruction::IF_EQZ]; 220} 221 222COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_EQZ) == kCondEq, check_if_eqz_ccode); 223COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_NEZ) == kCondNe, check_if_nez_ccode); 224COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_LTZ) == kCondLt, check_if_ltz_ccode); 225COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_GEZ) == kCondGe, check_if_gez_ccode); 226COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_GTZ) == kCondGt, check_if_gtz_ccode); 227COMPILE_ASSERT(ConditionCodeForIfCcZ(Instruction::IF_LEZ) == kCondLe, check_if_lez_ccode); 228 229int MIRGraph::GetSSAUseCount(int s_reg) { 230 return raw_use_counts_.Get(s_reg); 231} 232 233size_t MIRGraph::GetNumAvailableNonSpecialCompilerTemps() { 234 if (num_non_special_compiler_temps_ >= max_available_non_special_compiler_temps_) { 235 return 0; 236 } else { 237 return max_available_non_special_compiler_temps_ - num_non_special_compiler_temps_; 238 } 239} 240 241 242// FIXME - will probably need to revisit all uses of this, as type not defined. 243static const RegLocation temp_loc = {kLocCompilerTemp, 244 0, 1 /*defined*/, 0, 0, 0, 0, 0, 1 /*home*/, 245 RegStorage(), INVALID_SREG, INVALID_SREG}; 246 247CompilerTemp* MIRGraph::GetNewCompilerTemp(CompilerTempType ct_type, bool wide) { 248 // There is a limit to the number of non-special temps so check to make sure it wasn't exceeded. 249 if (ct_type == kCompilerTempVR) { 250 size_t available_temps = GetNumAvailableNonSpecialCompilerTemps(); 251 if (available_temps <= 0 || (available_temps <= 1 && wide)) { 252 return 0; 253 } 254 } 255 256 CompilerTemp *compiler_temp = static_cast<CompilerTemp *>(arena_->Alloc(sizeof(CompilerTemp), 257 kArenaAllocRegAlloc)); 258 259 // Create the type of temp requested. Special temps need special handling because 260 // they have a specific virtual register assignment. 261 if (ct_type == kCompilerTempSpecialMethodPtr) { 262 DCHECK_EQ(wide, false); 263 compiler_temp->v_reg = static_cast<int>(kVRegMethodPtrBaseReg); 264 compiler_temp->s_reg_low = AddNewSReg(compiler_temp->v_reg); 265 266 // The MIR graph keeps track of the sreg for method pointer specially, so record that now. 267 method_sreg_ = compiler_temp->s_reg_low; 268 } else { 269 DCHECK_EQ(ct_type, kCompilerTempVR); 270 271 // The new non-special compiler temp must receive a unique v_reg with a negative value. 272 compiler_temp->v_reg = static_cast<int>(kVRegNonSpecialTempBaseReg) - 273 num_non_special_compiler_temps_; 274 compiler_temp->s_reg_low = AddNewSReg(compiler_temp->v_reg); 275 num_non_special_compiler_temps_++; 276 277 if (wide) { 278 // Create a new CompilerTemp for the high part. 279 CompilerTemp *compiler_temp_high = 280 static_cast<CompilerTemp *>(arena_->Alloc(sizeof(CompilerTemp), kArenaAllocRegAlloc)); 281 compiler_temp_high->v_reg = compiler_temp->v_reg; 282 compiler_temp_high->s_reg_low = compiler_temp->s_reg_low; 283 compiler_temps_.Insert(compiler_temp_high); 284 285 // Ensure that the two registers are consecutive. Since the virtual registers used for temps 286 // grow in a negative fashion, we need the smaller to refer to the low part. Thus, we 287 // redefine the v_reg and s_reg_low. 288 compiler_temp->v_reg--; 289 int ssa_reg_high = compiler_temp->s_reg_low; 290 compiler_temp->s_reg_low = AddNewSReg(compiler_temp->v_reg); 291 int ssa_reg_low = compiler_temp->s_reg_low; 292 293 // If needed initialize the register location for the high part. 294 // The low part is handled later in this method on a common path. 295 if (reg_location_ != nullptr) { 296 reg_location_[ssa_reg_high] = temp_loc; 297 reg_location_[ssa_reg_high].high_word = 1; 298 reg_location_[ssa_reg_high].s_reg_low = ssa_reg_low; 299 reg_location_[ssa_reg_high].wide = true; 300 } 301 302 num_non_special_compiler_temps_++; 303 } 304 } 305 306 // Have we already allocated the register locations? 307 if (reg_location_ != nullptr) { 308 int ssa_reg_low = compiler_temp->s_reg_low; 309 reg_location_[ssa_reg_low] = temp_loc; 310 reg_location_[ssa_reg_low].s_reg_low = ssa_reg_low; 311 reg_location_[ssa_reg_low].wide = wide; 312 } 313 314 compiler_temps_.Insert(compiler_temp); 315 return compiler_temp; 316} 317 318/* Do some MIR-level extended basic block optimizations */ 319bool MIRGraph::BasicBlockOpt(BasicBlock* bb) { 320 if (bb->block_type == kDead) { 321 return true; 322 } 323 // Don't do a separate LVN if we did the GVN. 324 bool use_lvn = bb->use_lvn && (cu_->disable_opt & (1u << kGlobalValueNumbering)) != 0u; 325 std::unique_ptr<ScopedArenaAllocator> allocator; 326 std::unique_ptr<GlobalValueNumbering> global_valnum; 327 std::unique_ptr<LocalValueNumbering> local_valnum; 328 if (use_lvn) { 329 allocator.reset(ScopedArenaAllocator::Create(&cu_->arena_stack)); 330 global_valnum.reset(new (allocator.get()) GlobalValueNumbering(cu_, allocator.get())); 331 local_valnum.reset(new (allocator.get()) LocalValueNumbering(global_valnum.get(), bb->id)); 332 } 333 while (bb != NULL) { 334 for (MIR* mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 335 // TUNING: use the returned value number for CSE. 336 if (use_lvn) { 337 local_valnum->GetValueNumber(mir); 338 } 339 // Look for interesting opcodes, skip otherwise 340 Instruction::Code opcode = mir->dalvikInsn.opcode; 341 switch (opcode) { 342 case Instruction::CMPL_FLOAT: 343 case Instruction::CMPL_DOUBLE: 344 case Instruction::CMPG_FLOAT: 345 case Instruction::CMPG_DOUBLE: 346 case Instruction::CMP_LONG: 347 if ((cu_->disable_opt & (1 << kBranchFusing)) != 0) { 348 // Bitcode doesn't allow this optimization. 349 break; 350 } 351 if (mir->next != NULL) { 352 MIR* mir_next = mir->next; 353 // Make sure result of cmp is used by next insn and nowhere else 354 if (IsInstructionIfCcZ(mir_next->dalvikInsn.opcode) && 355 (mir->ssa_rep->defs[0] == mir_next->ssa_rep->uses[0]) && 356 (GetSSAUseCount(mir->ssa_rep->defs[0]) == 1)) { 357 mir_next->meta.ccode = ConditionCodeForIfCcZ(mir_next->dalvikInsn.opcode); 358 switch (opcode) { 359 case Instruction::CMPL_FLOAT: 360 mir_next->dalvikInsn.opcode = 361 static_cast<Instruction::Code>(kMirOpFusedCmplFloat); 362 break; 363 case Instruction::CMPL_DOUBLE: 364 mir_next->dalvikInsn.opcode = 365 static_cast<Instruction::Code>(kMirOpFusedCmplDouble); 366 break; 367 case Instruction::CMPG_FLOAT: 368 mir_next->dalvikInsn.opcode = 369 static_cast<Instruction::Code>(kMirOpFusedCmpgFloat); 370 break; 371 case Instruction::CMPG_DOUBLE: 372 mir_next->dalvikInsn.opcode = 373 static_cast<Instruction::Code>(kMirOpFusedCmpgDouble); 374 break; 375 case Instruction::CMP_LONG: 376 mir_next->dalvikInsn.opcode = 377 static_cast<Instruction::Code>(kMirOpFusedCmpLong); 378 break; 379 default: LOG(ERROR) << "Unexpected opcode: " << opcode; 380 } 381 mir->dalvikInsn.opcode = static_cast<Instruction::Code>(kMirOpNop); 382 // Copy the SSA information that is relevant. 383 mir_next->ssa_rep->num_uses = mir->ssa_rep->num_uses; 384 mir_next->ssa_rep->uses = mir->ssa_rep->uses; 385 mir_next->ssa_rep->fp_use = mir->ssa_rep->fp_use; 386 mir_next->ssa_rep->num_defs = 0; 387 mir->ssa_rep->num_uses = 0; 388 mir->ssa_rep->num_defs = 0; 389 // Copy in the decoded instruction information for potential SSA re-creation. 390 mir_next->dalvikInsn.vA = mir->dalvikInsn.vB; 391 mir_next->dalvikInsn.vB = mir->dalvikInsn.vC; 392 } 393 } 394 break; 395 case Instruction::GOTO: 396 case Instruction::GOTO_16: 397 case Instruction::GOTO_32: 398 case Instruction::IF_EQ: 399 case Instruction::IF_NE: 400 case Instruction::IF_LT: 401 case Instruction::IF_GE: 402 case Instruction::IF_GT: 403 case Instruction::IF_LE: 404 case Instruction::IF_EQZ: 405 case Instruction::IF_NEZ: 406 case Instruction::IF_LTZ: 407 case Instruction::IF_GEZ: 408 case Instruction::IF_GTZ: 409 case Instruction::IF_LEZ: 410 // If we've got a backwards branch to return, no need to suspend check. 411 if ((IsBackedge(bb, bb->taken) && GetBasicBlock(bb->taken)->dominates_return) || 412 (IsBackedge(bb, bb->fall_through) && 413 GetBasicBlock(bb->fall_through)->dominates_return)) { 414 mir->optimization_flags |= MIR_IGNORE_SUSPEND_CHECK; 415 if (cu_->verbose) { 416 LOG(INFO) << "Suppressed suspend check on branch to return at 0x" << std::hex 417 << mir->offset; 418 } 419 } 420 break; 421 default: 422 break; 423 } 424 // Is this the select pattern? 425 // TODO: flesh out support for Mips. NOTE: llvm's select op doesn't quite work here. 426 // TUNING: expand to support IF_xx compare & branches 427 if (!cu_->compiler->IsPortable() && 428 (cu_->instruction_set == kArm64 || cu_->instruction_set == kThumb2 || 429 cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) && 430 IsInstructionIfCcZ(mir->dalvikInsn.opcode)) { 431 BasicBlock* ft = GetBasicBlock(bb->fall_through); 432 DCHECK(ft != NULL); 433 BasicBlock* ft_ft = GetBasicBlock(ft->fall_through); 434 BasicBlock* ft_tk = GetBasicBlock(ft->taken); 435 436 BasicBlock* tk = GetBasicBlock(bb->taken); 437 DCHECK(tk != NULL); 438 BasicBlock* tk_ft = GetBasicBlock(tk->fall_through); 439 BasicBlock* tk_tk = GetBasicBlock(tk->taken); 440 441 /* 442 * In the select pattern, the taken edge goes to a block that unconditionally 443 * transfers to the rejoin block and the fall_though edge goes to a block that 444 * unconditionally falls through to the rejoin block. 445 */ 446 if ((tk_ft == NULL) && (ft_tk == NULL) && (tk_tk == ft_ft) && 447 (Predecessors(tk) == 1) && (Predecessors(ft) == 1)) { 448 /* 449 * Okay - we have the basic diamond shape. At the very least, we can eliminate the 450 * suspend check on the taken-taken branch back to the join point. 451 */ 452 if (SelectKind(tk->last_mir_insn) == kSelectGoto) { 453 tk->last_mir_insn->optimization_flags |= (MIR_IGNORE_SUSPEND_CHECK); 454 } 455 456 // TODO: Add logic for LONG. 457 // Are the block bodies something we can handle? 458 if ((ft->first_mir_insn == ft->last_mir_insn) && 459 (tk->first_mir_insn != tk->last_mir_insn) && 460 (tk->first_mir_insn->next == tk->last_mir_insn) && 461 ((SelectKind(ft->first_mir_insn) == kSelectMove) || 462 (SelectKind(ft->first_mir_insn) == kSelectConst)) && 463 (SelectKind(ft->first_mir_insn) == SelectKind(tk->first_mir_insn)) && 464 (SelectKind(tk->last_mir_insn) == kSelectGoto)) { 465 // Almost there. Are the instructions targeting the same vreg? 466 MIR* if_true = tk->first_mir_insn; 467 MIR* if_false = ft->first_mir_insn; 468 // It's possible that the target of the select isn't used - skip those (rare) cases. 469 MIR* phi = FindPhi(tk_tk, if_true->ssa_rep->defs[0]); 470 if ((phi != NULL) && (if_true->dalvikInsn.vA == if_false->dalvikInsn.vA)) { 471 /* 472 * We'll convert the IF_EQZ/IF_NEZ to a SELECT. We need to find the 473 * Phi node in the merge block and delete it (while using the SSA name 474 * of the merge as the target of the SELECT. Delete both taken and 475 * fallthrough blocks, and set fallthrough to merge block. 476 * NOTE: not updating other dataflow info (no longer used at this point). 477 * If this changes, need to update i_dom, etc. here (and in CombineBlocks). 478 */ 479 mir->meta.ccode = ConditionCodeForIfCcZ(mir->dalvikInsn.opcode); 480 mir->dalvikInsn.opcode = static_cast<Instruction::Code>(kMirOpSelect); 481 bool const_form = (SelectKind(if_true) == kSelectConst); 482 if ((SelectKind(if_true) == kSelectMove)) { 483 if (IsConst(if_true->ssa_rep->uses[0]) && 484 IsConst(if_false->ssa_rep->uses[0])) { 485 const_form = true; 486 if_true->dalvikInsn.vB = ConstantValue(if_true->ssa_rep->uses[0]); 487 if_false->dalvikInsn.vB = ConstantValue(if_false->ssa_rep->uses[0]); 488 } 489 } 490 if (const_form) { 491 /* 492 * TODO: If both constants are the same value, then instead of generating 493 * a select, we should simply generate a const bytecode. This should be 494 * considered after inlining which can lead to CFG of this form. 495 */ 496 // "true" set val in vB 497 mir->dalvikInsn.vB = if_true->dalvikInsn.vB; 498 // "false" set val in vC 499 mir->dalvikInsn.vC = if_false->dalvikInsn.vB; 500 } else { 501 DCHECK_EQ(SelectKind(if_true), kSelectMove); 502 DCHECK_EQ(SelectKind(if_false), kSelectMove); 503 int* src_ssa = 504 static_cast<int*>(arena_->Alloc(sizeof(int) * 3, kArenaAllocDFInfo)); 505 src_ssa[0] = mir->ssa_rep->uses[0]; 506 src_ssa[1] = if_true->ssa_rep->uses[0]; 507 src_ssa[2] = if_false->ssa_rep->uses[0]; 508 mir->ssa_rep->uses = src_ssa; 509 mir->ssa_rep->num_uses = 3; 510 } 511 mir->ssa_rep->num_defs = 1; 512 mir->ssa_rep->defs = 513 static_cast<int*>(arena_->Alloc(sizeof(int) * 1, kArenaAllocDFInfo)); 514 mir->ssa_rep->fp_def = 515 static_cast<bool*>(arena_->Alloc(sizeof(bool) * 1, kArenaAllocDFInfo)); 516 mir->ssa_rep->fp_def[0] = if_true->ssa_rep->fp_def[0]; 517 // Match type of uses to def. 518 mir->ssa_rep->fp_use = 519 static_cast<bool*>(arena_->Alloc(sizeof(bool) * mir->ssa_rep->num_uses, 520 kArenaAllocDFInfo)); 521 for (int i = 0; i < mir->ssa_rep->num_uses; i++) { 522 mir->ssa_rep->fp_use[i] = mir->ssa_rep->fp_def[0]; 523 } 524 /* 525 * There is usually a Phi node in the join block for our two cases. If the 526 * Phi node only contains our two cases as input, we will use the result 527 * SSA name of the Phi node as our select result and delete the Phi. If 528 * the Phi node has more than two operands, we will arbitrarily use the SSA 529 * name of the "true" path, delete the SSA name of the "false" path from the 530 * Phi node (and fix up the incoming arc list). 531 */ 532 if (phi->ssa_rep->num_uses == 2) { 533 mir->ssa_rep->defs[0] = phi->ssa_rep->defs[0]; 534 phi->dalvikInsn.opcode = static_cast<Instruction::Code>(kMirOpNop); 535 } else { 536 int dead_def = if_false->ssa_rep->defs[0]; 537 int live_def = if_true->ssa_rep->defs[0]; 538 mir->ssa_rep->defs[0] = live_def; 539 BasicBlockId* incoming = phi->meta.phi_incoming; 540 for (int i = 0; i < phi->ssa_rep->num_uses; i++) { 541 if (phi->ssa_rep->uses[i] == live_def) { 542 incoming[i] = bb->id; 543 } 544 } 545 for (int i = 0; i < phi->ssa_rep->num_uses; i++) { 546 if (phi->ssa_rep->uses[i] == dead_def) { 547 int last_slot = phi->ssa_rep->num_uses - 1; 548 phi->ssa_rep->uses[i] = phi->ssa_rep->uses[last_slot]; 549 incoming[i] = incoming[last_slot]; 550 } 551 } 552 } 553 phi->ssa_rep->num_uses--; 554 bb->taken = NullBasicBlockId; 555 tk->block_type = kDead; 556 for (MIR* tmir = ft->first_mir_insn; tmir != NULL; tmir = tmir->next) { 557 tmir->dalvikInsn.opcode = static_cast<Instruction::Code>(kMirOpNop); 558 } 559 } 560 } 561 } 562 } 563 } 564 bb = ((cu_->disable_opt & (1 << kSuppressExceptionEdges)) != 0) ? NextDominatedBlock(bb) : NULL; 565 } 566 if (use_lvn && UNLIKELY(!global_valnum->Good())) { 567 LOG(WARNING) << "LVN overflow in " << PrettyMethod(cu_->method_idx, *cu_->dex_file); 568 } 569 570 return true; 571} 572 573/* Collect stats on number of checks removed */ 574void MIRGraph::CountChecks(struct BasicBlock* bb) { 575 if (bb->data_flow_info != NULL) { 576 for (MIR* mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 577 if (mir->ssa_rep == NULL) { 578 continue; 579 } 580 uint64_t df_attributes = GetDataFlowAttributes(mir); 581 if (df_attributes & DF_HAS_NULL_CHKS) { 582 checkstats_->null_checks++; 583 if (mir->optimization_flags & MIR_IGNORE_NULL_CHECK) { 584 checkstats_->null_checks_eliminated++; 585 } 586 } 587 if (df_attributes & DF_HAS_RANGE_CHKS) { 588 checkstats_->range_checks++; 589 if (mir->optimization_flags & MIR_IGNORE_RANGE_CHECK) { 590 checkstats_->range_checks_eliminated++; 591 } 592 } 593 } 594 } 595} 596 597/* Try to make common case the fallthrough path */ 598bool MIRGraph::LayoutBlocks(BasicBlock* bb) { 599 // TODO: For now, just looking for direct throws. Consider generalizing for profile feedback 600 if (!bb->explicit_throw) { 601 return false; 602 } 603 BasicBlock* walker = bb; 604 while (true) { 605 // Check termination conditions 606 if ((walker->block_type == kEntryBlock) || (Predecessors(walker) != 1)) { 607 break; 608 } 609 BasicBlock* prev = GetBasicBlock(walker->predecessors->Get(0)); 610 if (prev->conditional_branch) { 611 if (GetBasicBlock(prev->fall_through) == walker) { 612 // Already done - return 613 break; 614 } 615 DCHECK_EQ(walker, GetBasicBlock(prev->taken)); 616 // Got one. Flip it and exit 617 Instruction::Code opcode = prev->last_mir_insn->dalvikInsn.opcode; 618 switch (opcode) { 619 case Instruction::IF_EQ: opcode = Instruction::IF_NE; break; 620 case Instruction::IF_NE: opcode = Instruction::IF_EQ; break; 621 case Instruction::IF_LT: opcode = Instruction::IF_GE; break; 622 case Instruction::IF_GE: opcode = Instruction::IF_LT; break; 623 case Instruction::IF_GT: opcode = Instruction::IF_LE; break; 624 case Instruction::IF_LE: opcode = Instruction::IF_GT; break; 625 case Instruction::IF_EQZ: opcode = Instruction::IF_NEZ; break; 626 case Instruction::IF_NEZ: opcode = Instruction::IF_EQZ; break; 627 case Instruction::IF_LTZ: opcode = Instruction::IF_GEZ; break; 628 case Instruction::IF_GEZ: opcode = Instruction::IF_LTZ; break; 629 case Instruction::IF_GTZ: opcode = Instruction::IF_LEZ; break; 630 case Instruction::IF_LEZ: opcode = Instruction::IF_GTZ; break; 631 default: LOG(FATAL) << "Unexpected opcode " << opcode; 632 } 633 prev->last_mir_insn->dalvikInsn.opcode = opcode; 634 BasicBlockId t_bb = prev->taken; 635 prev->taken = prev->fall_through; 636 prev->fall_through = t_bb; 637 break; 638 } 639 walker = prev; 640 } 641 return false; 642} 643 644/* Combine any basic blocks terminated by instructions that we now know can't throw */ 645void MIRGraph::CombineBlocks(struct BasicBlock* bb) { 646 // Loop here to allow combining a sequence of blocks 647 while (true) { 648 // Check termination conditions 649 if ((bb->first_mir_insn == NULL) 650 || (bb->data_flow_info == NULL) 651 || (bb->block_type == kExceptionHandling) 652 || (bb->block_type == kExitBlock) 653 || (bb->block_type == kDead) 654 || (bb->taken == NullBasicBlockId) 655 || (GetBasicBlock(bb->taken)->block_type != kExceptionHandling) 656 || (bb->successor_block_list_type != kNotUsed) 657 || (static_cast<int>(bb->last_mir_insn->dalvikInsn.opcode) != kMirOpCheck)) { 658 break; 659 } 660 661 // Test the kMirOpCheck instruction 662 MIR* mir = bb->last_mir_insn; 663 // Grab the attributes from the paired opcode 664 MIR* throw_insn = mir->meta.throw_insn; 665 uint64_t df_attributes = GetDataFlowAttributes(throw_insn); 666 bool can_combine = true; 667 if (df_attributes & DF_HAS_NULL_CHKS) { 668 can_combine &= ((throw_insn->optimization_flags & MIR_IGNORE_NULL_CHECK) != 0); 669 } 670 if (df_attributes & DF_HAS_RANGE_CHKS) { 671 can_combine &= ((throw_insn->optimization_flags & MIR_IGNORE_RANGE_CHECK) != 0); 672 } 673 if (!can_combine) { 674 break; 675 } 676 // OK - got one. Combine 677 BasicBlock* bb_next = GetBasicBlock(bb->fall_through); 678 DCHECK(!bb_next->catch_entry); 679 DCHECK_EQ(Predecessors(bb_next), 1U); 680 // Overwrite the kOpCheck insn with the paired opcode 681 DCHECK_EQ(bb_next->first_mir_insn, throw_insn); 682 *bb->last_mir_insn = *throw_insn; 683 // Use the successor info from the next block 684 bb->successor_block_list_type = bb_next->successor_block_list_type; 685 bb->successor_blocks = bb_next->successor_blocks; 686 // Use the ending block linkage from the next block 687 bb->fall_through = bb_next->fall_through; 688 GetBasicBlock(bb->taken)->block_type = kDead; // Kill the unused exception block 689 bb->taken = bb_next->taken; 690 // Include the rest of the instructions 691 bb->last_mir_insn = bb_next->last_mir_insn; 692 /* 693 * If lower-half of pair of blocks to combine contained a return, move the flag 694 * to the newly combined block. 695 */ 696 bb->terminated_by_return = bb_next->terminated_by_return; 697 698 /* 699 * NOTE: we aren't updating all dataflow info here. Should either make sure this pass 700 * happens after uses of i_dominated, dom_frontier or update the dataflow info here. 701 */ 702 703 // Kill bb_next and remap now-dead id to parent 704 bb_next->block_type = kDead; 705 block_id_map_.Overwrite(bb_next->id, bb->id); 706 707 // Now, loop back and see if we can keep going 708 } 709} 710 711void MIRGraph::EliminateNullChecksAndInferTypesStart() { 712 if ((cu_->disable_opt & (1 << kNullCheckElimination)) == 0) { 713 if (kIsDebugBuild) { 714 AllNodesIterator iter(this); 715 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 716 CHECK(bb->data_flow_info == nullptr || bb->data_flow_info->ending_check_v == nullptr); 717 } 718 } 719 720 DCHECK(temp_scoped_alloc_.get() == nullptr); 721 temp_scoped_alloc_.reset(ScopedArenaAllocator::Create(&cu_->arena_stack)); 722 temp_bit_vector_size_ = GetNumSSARegs(); 723 temp_bit_vector_ = new (temp_scoped_alloc_.get()) ArenaBitVector( 724 temp_scoped_alloc_.get(), temp_bit_vector_size_, false, kBitMapTempSSARegisterV); 725 } 726} 727 728/* 729 * Eliminate unnecessary null checks for a basic block. Also, while we're doing 730 * an iterative walk go ahead and perform type and size inference. 731 */ 732bool MIRGraph::EliminateNullChecksAndInferTypes(BasicBlock* bb) { 733 if (bb->data_flow_info == NULL) return false; 734 bool infer_changed = false; 735 bool do_nce = ((cu_->disable_opt & (1 << kNullCheckElimination)) == 0); 736 737 ArenaBitVector* ssa_regs_to_check = temp_bit_vector_; 738 if (do_nce) { 739 /* 740 * Set initial state. Be conservative with catch 741 * blocks and start with no assumptions about null check 742 * status (except for "this"). 743 */ 744 if ((bb->block_type == kEntryBlock) | bb->catch_entry) { 745 ssa_regs_to_check->ClearAllBits(); 746 // Assume all ins are objects. 747 for (uint16_t in_reg = cu_->num_dalvik_registers - cu_->num_ins; 748 in_reg < cu_->num_dalvik_registers; in_reg++) { 749 ssa_regs_to_check->SetBit(in_reg); 750 } 751 if ((cu_->access_flags & kAccStatic) == 0) { 752 // If non-static method, mark "this" as non-null 753 int this_reg = cu_->num_dalvik_registers - cu_->num_ins; 754 ssa_regs_to_check->ClearBit(this_reg); 755 } 756 } else if (bb->predecessors->Size() == 1) { 757 BasicBlock* pred_bb = GetBasicBlock(bb->predecessors->Get(0)); 758 // pred_bb must have already been processed at least once. 759 DCHECK(pred_bb->data_flow_info->ending_check_v != nullptr); 760 ssa_regs_to_check->Copy(pred_bb->data_flow_info->ending_check_v); 761 if (pred_bb->block_type == kDalvikByteCode) { 762 // Check to see if predecessor had an explicit null-check. 763 MIR* last_insn = pred_bb->last_mir_insn; 764 if (last_insn != nullptr) { 765 Instruction::Code last_opcode = last_insn->dalvikInsn.opcode; 766 if (last_opcode == Instruction::IF_EQZ) { 767 if (pred_bb->fall_through == bb->id) { 768 // The fall-through of a block following a IF_EQZ, set the vA of the IF_EQZ to show that 769 // it can't be null. 770 ssa_regs_to_check->ClearBit(last_insn->ssa_rep->uses[0]); 771 } 772 } else if (last_opcode == Instruction::IF_NEZ) { 773 if (pred_bb->taken == bb->id) { 774 // The taken block following a IF_NEZ, set the vA of the IF_NEZ to show that it can't be 775 // null. 776 ssa_regs_to_check->ClearBit(last_insn->ssa_rep->uses[0]); 777 } 778 } 779 } 780 } 781 } else { 782 // Starting state is union of all incoming arcs 783 GrowableArray<BasicBlockId>::Iterator iter(bb->predecessors); 784 BasicBlock* pred_bb = GetBasicBlock(iter.Next()); 785 CHECK(pred_bb != NULL); 786 while (pred_bb->data_flow_info->ending_check_v == nullptr) { 787 pred_bb = GetBasicBlock(iter.Next()); 788 // At least one predecessor must have been processed before this bb. 789 DCHECK(pred_bb != nullptr); 790 DCHECK(pred_bb->data_flow_info != nullptr); 791 } 792 ssa_regs_to_check->Copy(pred_bb->data_flow_info->ending_check_v); 793 while (true) { 794 pred_bb = GetBasicBlock(iter.Next()); 795 if (!pred_bb) break; 796 DCHECK(pred_bb->data_flow_info != nullptr); 797 if (pred_bb->data_flow_info->ending_check_v == nullptr) { 798 continue; 799 } 800 ssa_regs_to_check->Union(pred_bb->data_flow_info->ending_check_v); 801 } 802 } 803 // At this point, ssa_regs_to_check shows which sregs have an object definition with 804 // no intervening uses. 805 } 806 807 // Walk through the instruction in the block, updating as necessary 808 for (MIR* mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 809 if (mir->ssa_rep == NULL) { 810 continue; 811 } 812 813 // Propagate type info. 814 infer_changed = InferTypeAndSize(bb, mir, infer_changed); 815 if (!do_nce) { 816 continue; 817 } 818 819 uint64_t df_attributes = GetDataFlowAttributes(mir); 820 821 // Might need a null check? 822 if (df_attributes & DF_HAS_NULL_CHKS) { 823 int src_idx; 824 if (df_attributes & DF_NULL_CHK_1) { 825 src_idx = 1; 826 } else if (df_attributes & DF_NULL_CHK_2) { 827 src_idx = 2; 828 } else { 829 src_idx = 0; 830 } 831 int src_sreg = mir->ssa_rep->uses[src_idx]; 832 if (!ssa_regs_to_check->IsBitSet(src_sreg)) { 833 // Eliminate the null check. 834 mir->optimization_flags |= MIR_IGNORE_NULL_CHECK; 835 } else { 836 // Do the null check. 837 mir->optimization_flags &= ~MIR_IGNORE_NULL_CHECK; 838 // Mark s_reg as null-checked 839 ssa_regs_to_check->ClearBit(src_sreg); 840 } 841 } 842 843 if ((df_attributes & DF_A_WIDE) || 844 (df_attributes & (DF_REF_A | DF_SETS_CONST | DF_NULL_TRANSFER)) == 0) { 845 continue; 846 } 847 848 /* 849 * First, mark all object definitions as requiring null check. 850 * Note: we can't tell if a CONST definition might be used as an object, so treat 851 * them all as object definitions. 852 */ 853 if (((df_attributes & (DF_DA | DF_REF_A)) == (DF_DA | DF_REF_A)) || 854 (df_attributes & DF_SETS_CONST)) { 855 ssa_regs_to_check->SetBit(mir->ssa_rep->defs[0]); 856 } 857 858 // Now, remove mark from all object definitions we know are non-null. 859 if (df_attributes & DF_NON_NULL_DST) { 860 // Mark target of NEW* as non-null 861 ssa_regs_to_check->ClearBit(mir->ssa_rep->defs[0]); 862 } 863 864 // Mark non-null returns from invoke-style NEW* 865 if (df_attributes & DF_NON_NULL_RET) { 866 MIR* next_mir = mir->next; 867 // Next should be an MOVE_RESULT_OBJECT 868 if (next_mir && 869 next_mir->dalvikInsn.opcode == Instruction::MOVE_RESULT_OBJECT) { 870 // Mark as null checked 871 ssa_regs_to_check->ClearBit(next_mir->ssa_rep->defs[0]); 872 } else { 873 if (next_mir) { 874 LOG(WARNING) << "Unexpected opcode following new: " << next_mir->dalvikInsn.opcode; 875 } else if (bb->fall_through != NullBasicBlockId) { 876 // Look in next basic block 877 struct BasicBlock* next_bb = GetBasicBlock(bb->fall_through); 878 for (MIR* tmir = next_bb->first_mir_insn; tmir != NULL; 879 tmir =tmir->next) { 880 if (MIR::DecodedInstruction::IsPseudoMirOp(tmir->dalvikInsn.opcode)) { 881 continue; 882 } 883 // First non-pseudo should be MOVE_RESULT_OBJECT 884 if (tmir->dalvikInsn.opcode == Instruction::MOVE_RESULT_OBJECT) { 885 // Mark as null checked 886 ssa_regs_to_check->ClearBit(tmir->ssa_rep->defs[0]); 887 } else { 888 LOG(WARNING) << "Unexpected op after new: " << tmir->dalvikInsn.opcode; 889 } 890 break; 891 } 892 } 893 } 894 } 895 896 /* 897 * Propagate nullcheck state on register copies (including 898 * Phi pseudo copies. For the latter, nullcheck state is 899 * the "or" of all the Phi's operands. 900 */ 901 if (df_attributes & (DF_NULL_TRANSFER_0 | DF_NULL_TRANSFER_N)) { 902 int tgt_sreg = mir->ssa_rep->defs[0]; 903 int operands = (df_attributes & DF_NULL_TRANSFER_0) ? 1 : 904 mir->ssa_rep->num_uses; 905 bool needs_null_check = false; 906 for (int i = 0; i < operands; i++) { 907 needs_null_check |= ssa_regs_to_check->IsBitSet(mir->ssa_rep->uses[i]); 908 } 909 if (needs_null_check) { 910 ssa_regs_to_check->SetBit(tgt_sreg); 911 } else { 912 ssa_regs_to_check->ClearBit(tgt_sreg); 913 } 914 } 915 } 916 917 // Did anything change? 918 bool nce_changed = false; 919 if (do_nce) { 920 if (bb->data_flow_info->ending_check_v == nullptr) { 921 DCHECK(temp_scoped_alloc_.get() != nullptr); 922 bb->data_flow_info->ending_check_v = new (temp_scoped_alloc_.get()) ArenaBitVector( 923 temp_scoped_alloc_.get(), temp_bit_vector_size_, false, kBitMapNullCheck); 924 nce_changed = ssa_regs_to_check->GetHighestBitSet() != -1; 925 bb->data_flow_info->ending_check_v->Copy(ssa_regs_to_check); 926 } else if (!ssa_regs_to_check->SameBitsSet(bb->data_flow_info->ending_check_v)) { 927 nce_changed = true; 928 bb->data_flow_info->ending_check_v->Copy(ssa_regs_to_check); 929 } 930 } 931 return infer_changed | nce_changed; 932} 933 934void MIRGraph::EliminateNullChecksAndInferTypesEnd() { 935 if ((cu_->disable_opt & (1 << kNullCheckElimination)) == 0) { 936 // Clean up temporaries. 937 temp_bit_vector_size_ = 0u; 938 temp_bit_vector_ = nullptr; 939 AllNodesIterator iter(this); 940 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 941 if (bb->data_flow_info != nullptr) { 942 bb->data_flow_info->ending_check_v = nullptr; 943 } 944 } 945 DCHECK(temp_scoped_alloc_.get() != nullptr); 946 temp_scoped_alloc_.reset(); 947 } 948} 949 950bool MIRGraph::EliminateClassInitChecksGate() { 951 if ((cu_->disable_opt & (1 << kClassInitCheckElimination)) != 0 || 952 !cu_->mir_graph->HasStaticFieldAccess()) { 953 return false; 954 } 955 956 if (kIsDebugBuild) { 957 AllNodesIterator iter(this); 958 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 959 CHECK(bb->data_flow_info == nullptr || bb->data_flow_info->ending_check_v == nullptr); 960 } 961 } 962 963 DCHECK(temp_scoped_alloc_.get() == nullptr); 964 temp_scoped_alloc_.reset(ScopedArenaAllocator::Create(&cu_->arena_stack)); 965 966 // Each insn we use here has at least 2 code units, offset/2 will be a unique index. 967 const size_t end = (cu_->code_item->insns_size_in_code_units_ + 1u) / 2u; 968 temp_insn_data_ = static_cast<uint16_t*>( 969 temp_scoped_alloc_->Alloc(end * sizeof(*temp_insn_data_), kArenaAllocGrowableArray)); 970 971 uint32_t unique_class_count = 0u; 972 { 973 // Get unique_class_count and store indexes in temp_insn_data_ using a map on a nested 974 // ScopedArenaAllocator. 975 976 // Embed the map value in the entry to save space. 977 struct MapEntry { 978 // Map key: the class identified by the declaring dex file and type index. 979 const DexFile* declaring_dex_file; 980 uint16_t declaring_class_idx; 981 // Map value: index into bit vectors of classes requiring initialization checks. 982 uint16_t index; 983 }; 984 struct MapEntryComparator { 985 bool operator()(const MapEntry& lhs, const MapEntry& rhs) const { 986 if (lhs.declaring_class_idx != rhs.declaring_class_idx) { 987 return lhs.declaring_class_idx < rhs.declaring_class_idx; 988 } 989 return lhs.declaring_dex_file < rhs.declaring_dex_file; 990 } 991 }; 992 993 ScopedArenaAllocator allocator(&cu_->arena_stack); 994 ScopedArenaSet<MapEntry, MapEntryComparator> class_to_index_map(MapEntryComparator(), 995 allocator.Adapter()); 996 997 // First, find all SGET/SPUTs that may need class initialization checks, record INVOKE_STATICs. 998 AllNodesIterator iter(this); 999 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 1000 for (MIR* mir = bb->first_mir_insn; mir != nullptr; mir = mir->next) { 1001 DCHECK(bb->data_flow_info != nullptr); 1002 if (mir->dalvikInsn.opcode >= Instruction::SGET && 1003 mir->dalvikInsn.opcode <= Instruction::SPUT_SHORT) { 1004 const MirSFieldLoweringInfo& field_info = GetSFieldLoweringInfo(mir); 1005 uint16_t index = 0xffffu; 1006 if (!field_info.IsInitialized()) { 1007 DCHECK_LT(class_to_index_map.size(), 0xffffu); 1008 MapEntry entry = { 1009 // Treat unresolved fields as if each had its own class. 1010 field_info.IsResolved() ? field_info.DeclaringDexFile() 1011 : nullptr, 1012 field_info.IsResolved() ? field_info.DeclaringClassIndex() 1013 : field_info.FieldIndex(), 1014 static_cast<uint16_t>(class_to_index_map.size()) 1015 }; 1016 index = class_to_index_map.insert(entry).first->index; 1017 } 1018 // Using offset/2 for index into temp_insn_data_. 1019 temp_insn_data_[mir->offset / 2u] = index; 1020 } 1021 } 1022 } 1023 unique_class_count = static_cast<uint32_t>(class_to_index_map.size()); 1024 } 1025 1026 if (unique_class_count == 0u) { 1027 // All SGET/SPUTs refer to initialized classes. Nothing to do. 1028 temp_insn_data_ = nullptr; 1029 temp_scoped_alloc_.reset(); 1030 return false; 1031 } 1032 1033 temp_bit_vector_size_ = unique_class_count; 1034 temp_bit_vector_ = new (temp_scoped_alloc_.get()) ArenaBitVector( 1035 temp_scoped_alloc_.get(), temp_bit_vector_size_, false, kBitMapClInitCheck); 1036 DCHECK_GT(temp_bit_vector_size_, 0u); 1037 return true; 1038} 1039 1040/* 1041 * Eliminate unnecessary class initialization checks for a basic block. 1042 */ 1043bool MIRGraph::EliminateClassInitChecks(BasicBlock* bb) { 1044 DCHECK_EQ((cu_->disable_opt & (1 << kClassInitCheckElimination)), 0u); 1045 if (bb->data_flow_info == NULL) { 1046 return false; 1047 } 1048 1049 /* 1050 * Set initial state. Be conservative with catch 1051 * blocks and start with no assumptions about class init check status. 1052 */ 1053 ArenaBitVector* classes_to_check = temp_bit_vector_; 1054 DCHECK(classes_to_check != nullptr); 1055 if ((bb->block_type == kEntryBlock) | bb->catch_entry) { 1056 classes_to_check->SetInitialBits(temp_bit_vector_size_); 1057 } else if (bb->predecessors->Size() == 1) { 1058 BasicBlock* pred_bb = GetBasicBlock(bb->predecessors->Get(0)); 1059 // pred_bb must have already been processed at least once. 1060 DCHECK(pred_bb != nullptr); 1061 DCHECK(pred_bb->data_flow_info != nullptr); 1062 DCHECK(pred_bb->data_flow_info->ending_check_v != nullptr); 1063 classes_to_check->Copy(pred_bb->data_flow_info->ending_check_v); 1064 } else { 1065 // Starting state is union of all incoming arcs 1066 GrowableArray<BasicBlockId>::Iterator iter(bb->predecessors); 1067 BasicBlock* pred_bb = GetBasicBlock(iter.Next()); 1068 DCHECK(pred_bb != NULL); 1069 DCHECK(pred_bb->data_flow_info != NULL); 1070 while (pred_bb->data_flow_info->ending_check_v == nullptr) { 1071 pred_bb = GetBasicBlock(iter.Next()); 1072 // At least one predecessor must have been processed before this bb. 1073 DCHECK(pred_bb != nullptr); 1074 DCHECK(pred_bb->data_flow_info != nullptr); 1075 } 1076 classes_to_check->Copy(pred_bb->data_flow_info->ending_check_v); 1077 while (true) { 1078 pred_bb = GetBasicBlock(iter.Next()); 1079 if (!pred_bb) break; 1080 DCHECK(pred_bb->data_flow_info != nullptr); 1081 if (pred_bb->data_flow_info->ending_check_v == nullptr) { 1082 continue; 1083 } 1084 classes_to_check->Union(pred_bb->data_flow_info->ending_check_v); 1085 } 1086 } 1087 // At this point, classes_to_check shows which classes need clinit checks. 1088 1089 // Walk through the instruction in the block, updating as necessary 1090 for (MIR* mir = bb->first_mir_insn; mir != nullptr; mir = mir->next) { 1091 if (mir->dalvikInsn.opcode >= Instruction::SGET && 1092 mir->dalvikInsn.opcode <= Instruction::SPUT_SHORT) { 1093 uint16_t index = temp_insn_data_[mir->offset / 2u]; 1094 if (index != 0xffffu) { 1095 if (mir->dalvikInsn.opcode >= Instruction::SGET && 1096 mir->dalvikInsn.opcode <= Instruction::SPUT_SHORT) { 1097 if (!classes_to_check->IsBitSet(index)) { 1098 // Eliminate the class init check. 1099 mir->optimization_flags |= MIR_IGNORE_CLINIT_CHECK; 1100 } else { 1101 // Do the class init check. 1102 mir->optimization_flags &= ~MIR_IGNORE_CLINIT_CHECK; 1103 } 1104 } 1105 // Mark the class as initialized. 1106 classes_to_check->ClearBit(index); 1107 } 1108 } 1109 } 1110 1111 // Did anything change? 1112 bool changed = false; 1113 if (bb->data_flow_info->ending_check_v == nullptr) { 1114 DCHECK(temp_scoped_alloc_.get() != nullptr); 1115 DCHECK(bb->data_flow_info != nullptr); 1116 bb->data_flow_info->ending_check_v = new (temp_scoped_alloc_.get()) ArenaBitVector( 1117 temp_scoped_alloc_.get(), temp_bit_vector_size_, false, kBitMapClInitCheck); 1118 changed = classes_to_check->GetHighestBitSet() != -1; 1119 bb->data_flow_info->ending_check_v->Copy(classes_to_check); 1120 } else if (!classes_to_check->Equal(bb->data_flow_info->ending_check_v)) { 1121 changed = true; 1122 bb->data_flow_info->ending_check_v->Copy(classes_to_check); 1123 } 1124 return changed; 1125} 1126 1127void MIRGraph::EliminateClassInitChecksEnd() { 1128 // Clean up temporaries. 1129 temp_bit_vector_size_ = 0u; 1130 temp_bit_vector_ = nullptr; 1131 AllNodesIterator iter(this); 1132 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 1133 if (bb->data_flow_info != nullptr) { 1134 bb->data_flow_info->ending_check_v = nullptr; 1135 } 1136 } 1137 1138 DCHECK(temp_insn_data_ != nullptr); 1139 temp_insn_data_ = nullptr; 1140 DCHECK(temp_scoped_alloc_.get() != nullptr); 1141 temp_scoped_alloc_.reset(); 1142} 1143 1144bool MIRGraph::ApplyGlobalValueNumberingGate() { 1145 if ((cu_->disable_opt & (1u << kGlobalValueNumbering)) != 0u) { 1146 return false; 1147 } 1148 1149 DCHECK(temp_scoped_alloc_ == nullptr); 1150 temp_scoped_alloc_.reset(ScopedArenaAllocator::Create(&cu_->arena_stack)); 1151 DCHECK(temp_gvn_ == nullptr); 1152 temp_gvn_.reset( 1153 new (temp_scoped_alloc_.get()) GlobalValueNumbering(cu_, temp_scoped_alloc_.get())); 1154 return true; 1155} 1156 1157bool MIRGraph::ApplyGlobalValueNumbering(BasicBlock* bb) { 1158 DCHECK(temp_gvn_ != nullptr); 1159 LocalValueNumbering* lvn = temp_gvn_->PrepareBasicBlock(bb); 1160 if (lvn != nullptr) { 1161 for (MIR* mir = bb->first_mir_insn; mir != nullptr; mir = mir->next) { 1162 lvn->GetValueNumber(mir); 1163 } 1164 } 1165 bool change = (lvn != nullptr) && temp_gvn_->FinishBasicBlock(bb); 1166 return change; 1167} 1168 1169void MIRGraph::ApplyGlobalValueNumberingEnd() { 1170 // Perform modifications. 1171 if (temp_gvn_->Good()) { 1172 temp_gvn_->AllowModifications(); 1173 PreOrderDfsIterator iter(this); 1174 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) { 1175 LocalValueNumbering* lvn = temp_gvn_->PrepareBasicBlock(bb); 1176 if (lvn != nullptr) { 1177 for (MIR* mir = bb->first_mir_insn; mir != nullptr; mir = mir->next) { 1178 lvn->GetValueNumber(mir); 1179 } 1180 bool change = temp_gvn_->FinishBasicBlock(bb); 1181 DCHECK(!change) << PrettyMethod(cu_->method_idx, *cu_->dex_file); 1182 } 1183 } 1184 } else { 1185 LOG(WARNING) << "GVN failed for " << PrettyMethod(cu_->method_idx, *cu_->dex_file); 1186 } 1187 1188 DCHECK(temp_gvn_ != nullptr); 1189 temp_gvn_.reset(); 1190 DCHECK(temp_scoped_alloc_ != nullptr); 1191 temp_scoped_alloc_.reset(); 1192} 1193 1194void MIRGraph::ComputeInlineIFieldLoweringInfo(uint16_t field_idx, MIR* invoke, MIR* iget_or_iput) { 1195 uint32_t method_index = invoke->meta.method_lowering_info; 1196 if (temp_bit_vector_->IsBitSet(method_index)) { 1197 iget_or_iput->meta.ifield_lowering_info = temp_insn_data_[method_index]; 1198 DCHECK_EQ(field_idx, GetIFieldLoweringInfo(iget_or_iput).FieldIndex()); 1199 return; 1200 } 1201 1202 const MirMethodLoweringInfo& method_info = GetMethodLoweringInfo(invoke); 1203 MethodReference target = method_info.GetTargetMethod(); 1204 DexCompilationUnit inlined_unit( 1205 cu_, cu_->class_loader, cu_->class_linker, *target.dex_file, 1206 nullptr /* code_item not used */, 0u /* class_def_idx not used */, target.dex_method_index, 1207 0u /* access_flags not used */, nullptr /* verified_method not used */); 1208 MirIFieldLoweringInfo inlined_field_info(field_idx); 1209 MirIFieldLoweringInfo::Resolve(cu_->compiler_driver, &inlined_unit, &inlined_field_info, 1u); 1210 DCHECK(inlined_field_info.IsResolved()); 1211 1212 uint32_t field_info_index = ifield_lowering_infos_.Size(); 1213 ifield_lowering_infos_.Insert(inlined_field_info); 1214 temp_bit_vector_->SetBit(method_index); 1215 temp_insn_data_[method_index] = field_info_index; 1216 iget_or_iput->meta.ifield_lowering_info = field_info_index; 1217} 1218 1219bool MIRGraph::InlineSpecialMethodsGate() { 1220 if ((cu_->disable_opt & (1 << kSuppressMethodInlining)) != 0 || 1221 method_lowering_infos_.Size() == 0u) { 1222 return false; 1223 } 1224 if (cu_->compiler_driver->GetMethodInlinerMap() == nullptr) { 1225 // This isn't the Quick compiler. 1226 return false; 1227 } 1228 return true; 1229} 1230 1231void MIRGraph::InlineSpecialMethodsStart() { 1232 // Prepare for inlining getters/setters. Since we're inlining at most 1 IGET/IPUT from 1233 // each INVOKE, we can index the data by the MIR::meta::method_lowering_info index. 1234 1235 DCHECK(temp_scoped_alloc_.get() == nullptr); 1236 temp_scoped_alloc_.reset(ScopedArenaAllocator::Create(&cu_->arena_stack)); 1237 temp_bit_vector_size_ = method_lowering_infos_.Size(); 1238 temp_bit_vector_ = new (temp_scoped_alloc_.get()) ArenaBitVector( 1239 temp_scoped_alloc_.get(), temp_bit_vector_size_, false, kBitMapMisc); 1240 temp_bit_vector_->ClearAllBits(); 1241 temp_insn_data_ = static_cast<uint16_t*>(temp_scoped_alloc_->Alloc( 1242 temp_bit_vector_size_ * sizeof(*temp_insn_data_), kArenaAllocGrowableArray)); 1243} 1244 1245void MIRGraph::InlineSpecialMethods(BasicBlock* bb) { 1246 if (bb->block_type != kDalvikByteCode) { 1247 return; 1248 } 1249 for (MIR* mir = bb->first_mir_insn; mir != NULL; mir = mir->next) { 1250 if (MIR::DecodedInstruction::IsPseudoMirOp(mir->dalvikInsn.opcode)) { 1251 continue; 1252 } 1253 if (!(Instruction::FlagsOf(mir->dalvikInsn.opcode) & Instruction::kInvoke)) { 1254 continue; 1255 } 1256 const MirMethodLoweringInfo& method_info = GetMethodLoweringInfo(mir); 1257 if (!method_info.FastPath()) { 1258 continue; 1259 } 1260 InvokeType sharp_type = method_info.GetSharpType(); 1261 if ((sharp_type != kDirect) && 1262 (sharp_type != kStatic || method_info.NeedsClassInitialization())) { 1263 continue; 1264 } 1265 DCHECK(cu_->compiler_driver->GetMethodInlinerMap() != nullptr); 1266 MethodReference target = method_info.GetTargetMethod(); 1267 if (cu_->compiler_driver->GetMethodInlinerMap()->GetMethodInliner(target.dex_file) 1268 ->GenInline(this, bb, mir, target.dex_method_index)) { 1269 if (cu_->verbose || cu_->print_pass) { 1270 LOG(INFO) << "SpecialMethodInliner: Inlined " << method_info.GetInvokeType() << " (" 1271 << sharp_type << ") call to \"" << PrettyMethod(target.dex_method_index, *target.dex_file) 1272 << "\" from \"" << PrettyMethod(cu_->method_idx, *cu_->dex_file) 1273 << "\" @0x" << std::hex << mir->offset; 1274 } 1275 } 1276 } 1277} 1278 1279void MIRGraph::InlineSpecialMethodsEnd() { 1280 DCHECK(temp_insn_data_ != nullptr); 1281 temp_insn_data_ = nullptr; 1282 DCHECK(temp_bit_vector_ != nullptr); 1283 temp_bit_vector_ = nullptr; 1284 DCHECK(temp_scoped_alloc_.get() != nullptr); 1285 temp_scoped_alloc_.reset(); 1286} 1287 1288void MIRGraph::DumpCheckStats() { 1289 Checkstats* stats = 1290 static_cast<Checkstats*>(arena_->Alloc(sizeof(Checkstats), kArenaAllocDFInfo)); 1291 checkstats_ = stats; 1292 AllNodesIterator iter(this); 1293 for (BasicBlock* bb = iter.Next(); bb != NULL; bb = iter.Next()) { 1294 CountChecks(bb); 1295 } 1296 if (stats->null_checks > 0) { 1297 float eliminated = static_cast<float>(stats->null_checks_eliminated); 1298 float checks = static_cast<float>(stats->null_checks); 1299 LOG(INFO) << "Null Checks: " << PrettyMethod(cu_->method_idx, *cu_->dex_file) << " " 1300 << stats->null_checks_eliminated << " of " << stats->null_checks << " -> " 1301 << (eliminated/checks) * 100.0 << "%"; 1302 } 1303 if (stats->range_checks > 0) { 1304 float eliminated = static_cast<float>(stats->range_checks_eliminated); 1305 float checks = static_cast<float>(stats->range_checks); 1306 LOG(INFO) << "Range Checks: " << PrettyMethod(cu_->method_idx, *cu_->dex_file) << " " 1307 << stats->range_checks_eliminated << " of " << stats->range_checks << " -> " 1308 << (eliminated/checks) * 100.0 << "%"; 1309 } 1310} 1311 1312bool MIRGraph::BuildExtendedBBList(struct BasicBlock* bb) { 1313 if (bb->visited) return false; 1314 if (!((bb->block_type == kEntryBlock) || (bb->block_type == kDalvikByteCode) 1315 || (bb->block_type == kExitBlock))) { 1316 // Ignore special blocks 1317 bb->visited = true; 1318 return false; 1319 } 1320 // Must be head of extended basic block. 1321 BasicBlock* start_bb = bb; 1322 extended_basic_blocks_.push_back(bb->id); 1323 bool terminated_by_return = false; 1324 bool do_local_value_numbering = false; 1325 // Visit blocks strictly dominated by this head. 1326 while (bb != NULL) { 1327 bb->visited = true; 1328 terminated_by_return |= bb->terminated_by_return; 1329 do_local_value_numbering |= bb->use_lvn; 1330 bb = NextDominatedBlock(bb); 1331 } 1332 if (terminated_by_return || do_local_value_numbering) { 1333 // Do lvn for all blocks in this extended set. 1334 bb = start_bb; 1335 while (bb != NULL) { 1336 bb->use_lvn = do_local_value_numbering; 1337 bb->dominates_return = terminated_by_return; 1338 bb = NextDominatedBlock(bb); 1339 } 1340 } 1341 return false; // Not iterative - return value will be ignored 1342} 1343 1344void MIRGraph::BasicBlockOptimization() { 1345 if ((cu_->disable_opt & (1 << kSuppressExceptionEdges)) != 0) { 1346 ClearAllVisitedFlags(); 1347 PreOrderDfsIterator iter2(this); 1348 for (BasicBlock* bb = iter2.Next(); bb != NULL; bb = iter2.Next()) { 1349 BuildExtendedBBList(bb); 1350 } 1351 // Perform extended basic block optimizations. 1352 for (unsigned int i = 0; i < extended_basic_blocks_.size(); i++) { 1353 BasicBlockOpt(GetBasicBlock(extended_basic_blocks_[i])); 1354 } 1355 } else { 1356 PreOrderDfsIterator iter(this); 1357 for (BasicBlock* bb = iter.Next(); bb != NULL; bb = iter.Next()) { 1358 BasicBlockOpt(bb); 1359 } 1360 } 1361} 1362 1363} // namespace art 1364