1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass munges the code in the input function to better prepare it for 11 // SelectionDAG-based code generation. This works around limitations in it's 12 // basic-block-at-a-time approach. It should eventually be removed. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #define DEBUG_TYPE "codegenprepare" 17 #include "llvm/Transforms/Scalar.h" 18 #include "llvm/Constants.h" 19 #include "llvm/DerivedTypes.h" 20 #include "llvm/Function.h" 21 #include "llvm/IRBuilder.h" 22 #include "llvm/InlineAsm.h" 23 #include "llvm/Instructions.h" 24 #include "llvm/IntrinsicInst.h" 25 #include "llvm/Pass.h" 26 #include "llvm/ADT/DenseMap.h" 27 #include "llvm/ADT/SmallSet.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/Analysis/Dominators.h" 30 #include "llvm/Analysis/InstructionSimplify.h" 31 #include "llvm/Analysis/ProfileInfo.h" 32 #include "llvm/Assembly/Writer.h" 33 #include "llvm/Support/CallSite.h" 34 #include "llvm/Support/CommandLine.h" 35 #include "llvm/Support/Debug.h" 36 #include "llvm/Support/GetElementPtrTypeIterator.h" 37 #include "llvm/Support/PatternMatch.h" 38 #include "llvm/Support/ValueHandle.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include "llvm/Target/TargetData.h" 41 #include "llvm/Target/TargetLibraryInfo.h" 42 #include "llvm/Target/TargetLowering.h" 43 #include "llvm/Transforms/Utils/AddrModeMatcher.h" 44 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 45 #include "llvm/Transforms/Utils/BuildLibCalls.h" 46 #include "llvm/Transforms/Utils/BypassSlowDivision.h" 47 #include "llvm/Transforms/Utils/Local.h" 48 using namespace llvm; 49 using namespace llvm::PatternMatch; 50 51 STATISTIC(NumBlocksElim, "Number of blocks eliminated"); 52 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated"); 53 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts"); 54 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of " 55 "sunken Cmps"); 56 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses " 57 "of sunken Casts"); 58 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address " 59 "computations were sunk"); 60 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads"); 61 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized"); 62 STATISTIC(NumRetsDup, "Number of return instructions duplicated"); 63 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved"); 64 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches"); 65 66 static cl::opt<bool> DisableBranchOpts( 67 "disable-cgp-branch-opts", cl::Hidden, cl::init(false), 68 cl::desc("Disable branch optimizations in CodeGenPrepare")); 69 70 static cl::opt<bool> DisableSelectToBranch( 71 "disable-cgp-select2branch", cl::Hidden, cl::init(false), 72 cl::desc("Disable select to branch conversion.")); 73 74 namespace { 75 class CodeGenPrepare : public FunctionPass { 76 /// TLI - Keep a pointer of a TargetLowering to consult for determining 77 /// transformation profitability. 78 const TargetLowering *TLI; 79 const TargetLibraryInfo *TLInfo; 80 DominatorTree *DT; 81 ProfileInfo *PFI; 82 83 /// CurInstIterator - As we scan instructions optimizing them, this is the 84 /// next instruction to optimize. Xforms that can invalidate this should 85 /// update it. 86 BasicBlock::iterator CurInstIterator; 87 88 /// Keeps track of non-local addresses that have been sunk into a block. 89 /// This allows us to avoid inserting duplicate code for blocks with 90 /// multiple load/stores of the same address. 91 DenseMap<Value*, Value*> SunkAddrs; 92 93 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to 94 /// be updated. 95 bool ModifiedDT; 96 97 /// OptSize - True if optimizing for size. 98 bool OptSize; 99 100 public: 101 static char ID; // Pass identification, replacement for typeid 102 explicit CodeGenPrepare(const TargetLowering *tli = 0) 103 : FunctionPass(ID), TLI(tli) { 104 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); 105 } 106 bool runOnFunction(Function &F); 107 108 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 109 AU.addPreserved<DominatorTree>(); 110 AU.addPreserved<ProfileInfo>(); 111 AU.addRequired<TargetLibraryInfo>(); 112 } 113 114 private: 115 bool EliminateFallThrough(Function &F); 116 bool EliminateMostlyEmptyBlocks(Function &F); 117 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 118 void EliminateMostlyEmptyBlock(BasicBlock *BB); 119 bool OptimizeBlock(BasicBlock &BB); 120 bool OptimizeInst(Instruction *I); 121 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy); 122 bool OptimizeInlineAsmInst(CallInst *CS); 123 bool OptimizeCallInst(CallInst *CI); 124 bool MoveExtToFormExtLoad(Instruction *I); 125 bool OptimizeExtUses(Instruction *I); 126 bool OptimizeSelectInst(SelectInst *SI); 127 bool DupRetToEnableTailCallOpts(ReturnInst *RI); 128 bool PlaceDbgValues(Function &F); 129 }; 130 } 131 132 char CodeGenPrepare::ID = 0; 133 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare", 134 "Optimize for code generation", false, false) 135 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 136 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare", 137 "Optimize for code generation", false, false) 138 139 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 140 return new CodeGenPrepare(TLI); 141 } 142 143 bool CodeGenPrepare::runOnFunction(Function &F) { 144 bool EverMadeChange = false; 145 146 ModifiedDT = false; 147 TLInfo = &getAnalysis<TargetLibraryInfo>(); 148 DT = getAnalysisIfAvailable<DominatorTree>(); 149 PFI = getAnalysisIfAvailable<ProfileInfo>(); 150 OptSize = F.hasFnAttr(Attribute::OptimizeForSize); 151 152 /// This optimization identifies DIV instructions that can be 153 /// profitably bypassed and carried out with a shorter, faster divide. 154 if (TLI && TLI->isSlowDivBypassed()) { 155 const DenseMap<Type *, Type *> &BypassTypeMap = TLI->getBypassSlowDivTypes(); 156 157 for (Function::iterator I = F.begin(); I != F.end(); I++) { 158 EverMadeChange |= bypassSlowDivision(F, 159 I, 160 BypassTypeMap); 161 } 162 } 163 164 // Eliminate blocks that contain only PHI nodes and an 165 // unconditional branch. 166 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 167 168 // llvm.dbg.value is far away from the value then iSel may not be able 169 // handle it properly. iSel will drop llvm.dbg.value if it can not 170 // find a node corresponding to the value. 171 EverMadeChange |= PlaceDbgValues(F); 172 173 bool MadeChange = true; 174 while (MadeChange) { 175 MadeChange = false; 176 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) { 177 BasicBlock *BB = I++; 178 MadeChange |= OptimizeBlock(*BB); 179 } 180 EverMadeChange |= MadeChange; 181 } 182 183 SunkAddrs.clear(); 184 185 if (!DisableBranchOpts) { 186 MadeChange = false; 187 SmallPtrSet<BasicBlock*, 8> WorkList; 188 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 189 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB)); 190 MadeChange |= ConstantFoldTerminator(BB, true); 191 if (!MadeChange) continue; 192 193 for (SmallVectorImpl<BasicBlock*>::iterator 194 II = Successors.begin(), IE = Successors.end(); II != IE; ++II) 195 if (pred_begin(*II) == pred_end(*II)) 196 WorkList.insert(*II); 197 } 198 199 for (SmallPtrSet<BasicBlock*, 8>::iterator 200 I = WorkList.begin(), E = WorkList.end(); I != E; ++I) 201 DeleteDeadBlock(*I); 202 203 // Merge pairs of basic blocks with unconditional branches, connected by 204 // a single edge. 205 if (EverMadeChange || MadeChange) 206 MadeChange |= EliminateFallThrough(F); 207 208 if (MadeChange) 209 ModifiedDT = true; 210 EverMadeChange |= MadeChange; 211 } 212 213 if (ModifiedDT && DT) 214 DT->DT->recalculate(F); 215 216 return EverMadeChange; 217 } 218 219 /// EliminateFallThrough - Merge basic blocks which are connected 220 /// by a single edge, where one of the basic blocks has a single successor 221 /// pointing to the other basic block, which has a single predecessor. 222 bool CodeGenPrepare::EliminateFallThrough(Function &F) { 223 bool Changed = false; 224 // Scan all of the blocks in the function, except for the entry block. 225 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 226 BasicBlock *BB = I++; 227 // If the destination block has a single pred, then this is a trivial 228 // edge, just collapse it. 229 BasicBlock *SinglePred = BB->getSinglePredecessor(); 230 231 if (!SinglePred || SinglePred == BB) continue; 232 233 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); 234 if (Term && !Term->isConditional()) { 235 Changed = true; 236 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n"); 237 // Remember if SinglePred was the entry block of the function. 238 // If so, we will need to move BB back to the entry position. 239 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 240 MergeBasicBlockIntoOnlyPred(BB, this); 241 242 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 243 BB->moveBefore(&BB->getParent()->getEntryBlock()); 244 245 // We have erased a block. Update the iterator. 246 I = BB; 247 } 248 } 249 return Changed; 250 } 251 252 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 253 /// debug info directives, and an unconditional branch. Passes before isel 254 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 255 /// isel. Start by eliminating these blocks so we can split them the way we 256 /// want them. 257 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 258 bool MadeChange = false; 259 // Note that this intentionally skips the entry block. 260 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 261 BasicBlock *BB = I++; 262 263 // If this block doesn't end with an uncond branch, ignore it. 264 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 265 if (!BI || !BI->isUnconditional()) 266 continue; 267 268 // If the instruction before the branch (skipping debug info) isn't a phi 269 // node, then other stuff is happening here. 270 BasicBlock::iterator BBI = BI; 271 if (BBI != BB->begin()) { 272 --BBI; 273 while (isa<DbgInfoIntrinsic>(BBI)) { 274 if (BBI == BB->begin()) 275 break; 276 --BBI; 277 } 278 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 279 continue; 280 } 281 282 // Do not break infinite loops. 283 BasicBlock *DestBB = BI->getSuccessor(0); 284 if (DestBB == BB) 285 continue; 286 287 if (!CanMergeBlocks(BB, DestBB)) 288 continue; 289 290 EliminateMostlyEmptyBlock(BB); 291 MadeChange = true; 292 } 293 return MadeChange; 294 } 295 296 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 297 /// single uncond branch between them, and BB contains no other non-phi 298 /// instructions. 299 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 300 const BasicBlock *DestBB) const { 301 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 302 // the successor. If there are more complex condition (e.g. preheaders), 303 // don't mess around with them. 304 BasicBlock::const_iterator BBI = BB->begin(); 305 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 306 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end(); 307 UI != E; ++UI) { 308 const Instruction *User = cast<Instruction>(*UI); 309 if (User->getParent() != DestBB || !isa<PHINode>(User)) 310 return false; 311 // If User is inside DestBB block and it is a PHINode then check 312 // incoming value. If incoming value is not from BB then this is 313 // a complex condition (e.g. preheaders) we want to avoid here. 314 if (User->getParent() == DestBB) { 315 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 316 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 317 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 318 if (Insn && Insn->getParent() == BB && 319 Insn->getParent() != UPN->getIncomingBlock(I)) 320 return false; 321 } 322 } 323 } 324 } 325 326 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 327 // and DestBB may have conflicting incoming values for the block. If so, we 328 // can't merge the block. 329 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 330 if (!DestBBPN) return true; // no conflict. 331 332 // Collect the preds of BB. 333 SmallPtrSet<const BasicBlock*, 16> BBPreds; 334 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 335 // It is faster to get preds from a PHI than with pred_iterator. 336 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 337 BBPreds.insert(BBPN->getIncomingBlock(i)); 338 } else { 339 BBPreds.insert(pred_begin(BB), pred_end(BB)); 340 } 341 342 // Walk the preds of DestBB. 343 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 344 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 345 if (BBPreds.count(Pred)) { // Common predecessor? 346 BBI = DestBB->begin(); 347 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 348 const Value *V1 = PN->getIncomingValueForBlock(Pred); 349 const Value *V2 = PN->getIncomingValueForBlock(BB); 350 351 // If V2 is a phi node in BB, look up what the mapped value will be. 352 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 353 if (V2PN->getParent() == BB) 354 V2 = V2PN->getIncomingValueForBlock(Pred); 355 356 // If there is a conflict, bail out. 357 if (V1 != V2) return false; 358 } 359 } 360 } 361 362 return true; 363 } 364 365 366 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 367 /// an unconditional branch in it. 368 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 369 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 370 BasicBlock *DestBB = BI->getSuccessor(0); 371 372 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB); 373 374 // If the destination block has a single pred, then this is a trivial edge, 375 // just collapse it. 376 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 377 if (SinglePred != DestBB) { 378 // Remember if SinglePred was the entry block of the function. If so, we 379 // will need to move BB back to the entry position. 380 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 381 MergeBasicBlockIntoOnlyPred(DestBB, this); 382 383 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 384 BB->moveBefore(&BB->getParent()->getEntryBlock()); 385 386 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 387 return; 388 } 389 } 390 391 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 392 // to handle the new incoming edges it is about to have. 393 PHINode *PN; 394 for (BasicBlock::iterator BBI = DestBB->begin(); 395 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 396 // Remove the incoming value for BB, and remember it. 397 Value *InVal = PN->removeIncomingValue(BB, false); 398 399 // Two options: either the InVal is a phi node defined in BB or it is some 400 // value that dominates BB. 401 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 402 if (InValPhi && InValPhi->getParent() == BB) { 403 // Add all of the input values of the input PHI as inputs of this phi. 404 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 405 PN->addIncoming(InValPhi->getIncomingValue(i), 406 InValPhi->getIncomingBlock(i)); 407 } else { 408 // Otherwise, add one instance of the dominating value for each edge that 409 // we will be adding. 410 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 411 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 412 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 413 } else { 414 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 415 PN->addIncoming(InVal, *PI); 416 } 417 } 418 } 419 420 // The PHIs are now updated, change everything that refers to BB to use 421 // DestBB and remove BB. 422 BB->replaceAllUsesWith(DestBB); 423 if (DT && !ModifiedDT) { 424 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock(); 425 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock(); 426 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom); 427 DT->changeImmediateDominator(DestBB, NewIDom); 428 DT->eraseNode(BB); 429 } 430 if (PFI) { 431 PFI->replaceAllUses(BB, DestBB); 432 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB)); 433 } 434 BB->eraseFromParent(); 435 ++NumBlocksElim; 436 437 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 438 } 439 440 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 441 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), 442 /// sink it into user blocks to reduce the number of virtual 443 /// registers that must be created and coalesced. 444 /// 445 /// Return true if any changes are made. 446 /// 447 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 448 // If this is a noop copy, 449 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 450 EVT DstVT = TLI.getValueType(CI->getType()); 451 452 // This is an fp<->int conversion? 453 if (SrcVT.isInteger() != DstVT.isInteger()) 454 return false; 455 456 // If this is an extension, it will be a zero or sign extension, which 457 // isn't a noop. 458 if (SrcVT.bitsLT(DstVT)) return false; 459 460 // If these values will be promoted, find out what they will be promoted 461 // to. This helps us consider truncates on PPC as noop copies when they 462 // are. 463 if (TLI.getTypeAction(CI->getContext(), SrcVT) == 464 TargetLowering::TypePromoteInteger) 465 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); 466 if (TLI.getTypeAction(CI->getContext(), DstVT) == 467 TargetLowering::TypePromoteInteger) 468 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); 469 470 // If, after promotion, these are the same types, this is a noop copy. 471 if (SrcVT != DstVT) 472 return false; 473 474 BasicBlock *DefBB = CI->getParent(); 475 476 /// InsertedCasts - Only insert a cast in each block once. 477 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 478 479 bool MadeChange = false; 480 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 481 UI != E; ) { 482 Use &TheUse = UI.getUse(); 483 Instruction *User = cast<Instruction>(*UI); 484 485 // Figure out which BB this cast is used in. For PHI's this is the 486 // appropriate predecessor block. 487 BasicBlock *UserBB = User->getParent(); 488 if (PHINode *PN = dyn_cast<PHINode>(User)) { 489 UserBB = PN->getIncomingBlock(UI); 490 } 491 492 // Preincrement use iterator so we don't invalidate it. 493 ++UI; 494 495 // If this user is in the same block as the cast, don't change the cast. 496 if (UserBB == DefBB) continue; 497 498 // If we have already inserted a cast into this block, use it. 499 CastInst *&InsertedCast = InsertedCasts[UserBB]; 500 501 if (!InsertedCast) { 502 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 503 InsertedCast = 504 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 505 InsertPt); 506 MadeChange = true; 507 } 508 509 // Replace a use of the cast with a use of the new cast. 510 TheUse = InsertedCast; 511 ++NumCastUses; 512 } 513 514 // If we removed all uses, nuke the cast. 515 if (CI->use_empty()) { 516 CI->eraseFromParent(); 517 MadeChange = true; 518 } 519 520 return MadeChange; 521 } 522 523 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 524 /// the number of virtual registers that must be created and coalesced. This is 525 /// a clear win except on targets with multiple condition code registers 526 /// (PowerPC), where it might lose; some adjustment may be wanted there. 527 /// 528 /// Return true if any changes are made. 529 static bool OptimizeCmpExpression(CmpInst *CI) { 530 BasicBlock *DefBB = CI->getParent(); 531 532 /// InsertedCmp - Only insert a cmp in each block once. 533 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 534 535 bool MadeChange = false; 536 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 537 UI != E; ) { 538 Use &TheUse = UI.getUse(); 539 Instruction *User = cast<Instruction>(*UI); 540 541 // Preincrement use iterator so we don't invalidate it. 542 ++UI; 543 544 // Don't bother for PHI nodes. 545 if (isa<PHINode>(User)) 546 continue; 547 548 // Figure out which BB this cmp is used in. 549 BasicBlock *UserBB = User->getParent(); 550 551 // If this user is in the same block as the cmp, don't change the cmp. 552 if (UserBB == DefBB) continue; 553 554 // If we have already inserted a cmp into this block, use it. 555 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 556 557 if (!InsertedCmp) { 558 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 559 InsertedCmp = 560 CmpInst::Create(CI->getOpcode(), 561 CI->getPredicate(), CI->getOperand(0), 562 CI->getOperand(1), "", InsertPt); 563 MadeChange = true; 564 } 565 566 // Replace a use of the cmp with a use of the new cmp. 567 TheUse = InsertedCmp; 568 ++NumCmpUses; 569 } 570 571 // If we removed all uses, nuke the cmp. 572 if (CI->use_empty()) 573 CI->eraseFromParent(); 574 575 return MadeChange; 576 } 577 578 namespace { 579 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls { 580 protected: 581 void replaceCall(Value *With) { 582 CI->replaceAllUsesWith(With); 583 CI->eraseFromParent(); 584 } 585 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const { 586 if (ConstantInt *SizeCI = 587 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) 588 return SizeCI->isAllOnesValue(); 589 return false; 590 } 591 }; 592 } // end anonymous namespace 593 594 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { 595 BasicBlock *BB = CI->getParent(); 596 597 // Lower inline assembly if we can. 598 // If we found an inline asm expession, and if the target knows how to 599 // lower it to normal LLVM code, do so now. 600 if (TLI && isa<InlineAsm>(CI->getCalledValue())) { 601 if (TLI->ExpandInlineAsm(CI)) { 602 // Avoid invalidating the iterator. 603 CurInstIterator = BB->begin(); 604 // Avoid processing instructions out of order, which could cause 605 // reuse before a value is defined. 606 SunkAddrs.clear(); 607 return true; 608 } 609 // Sink address computing for memory operands into the block. 610 if (OptimizeInlineAsmInst(CI)) 611 return true; 612 } 613 614 // Lower all uses of llvm.objectsize.* 615 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); 616 if (II && II->getIntrinsicID() == Intrinsic::objectsize) { 617 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1); 618 Type *ReturnTy = CI->getType(); 619 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 620 621 // Substituting this can cause recursive simplifications, which can 622 // invalidate our iterator. Use a WeakVH to hold onto it in case this 623 // happens. 624 WeakVH IterHandle(CurInstIterator); 625 626 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getTargetData() : 0, 627 TLInfo, ModifiedDT ? 0 : DT); 628 629 // If the iterator instruction was recursively deleted, start over at the 630 // start of the block. 631 if (IterHandle != CurInstIterator) { 632 CurInstIterator = BB->begin(); 633 SunkAddrs.clear(); 634 } 635 return true; 636 } 637 638 if (II && TLI) { 639 SmallVector<Value*, 2> PtrOps; 640 Type *AccessTy; 641 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy)) 642 while (!PtrOps.empty()) 643 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy)) 644 return true; 645 } 646 647 // From here on out we're working with named functions. 648 if (CI->getCalledFunction() == 0) return false; 649 650 // We'll need TargetData from here on out. 651 const TargetData *TD = TLI ? TLI->getTargetData() : 0; 652 if (!TD) return false; 653 654 // Lower all default uses of _chk calls. This is very similar 655 // to what InstCombineCalls does, but here we are only lowering calls 656 // that have the default "don't know" as the objectsize. Anything else 657 // should be left alone. 658 CodeGenPrepareFortifiedLibCalls Simplifier; 659 return Simplifier.fold(CI, TD, TLInfo); 660 } 661 662 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return 663 /// instructions to the predecessor to enable tail call optimizations. The 664 /// case it is currently looking for is: 665 /// bb0: 666 /// %tmp0 = tail call i32 @f0() 667 /// br label %return 668 /// bb1: 669 /// %tmp1 = tail call i32 @f1() 670 /// br label %return 671 /// bb2: 672 /// %tmp2 = tail call i32 @f2() 673 /// br label %return 674 /// return: 675 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] 676 /// ret i32 %retval 677 /// 678 /// => 679 /// 680 /// bb0: 681 /// %tmp0 = tail call i32 @f0() 682 /// ret i32 %tmp0 683 /// bb1: 684 /// %tmp1 = tail call i32 @f1() 685 /// ret i32 %tmp1 686 /// bb2: 687 /// %tmp2 = tail call i32 @f2() 688 /// ret i32 %tmp2 689 /// 690 bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) { 691 if (!TLI) 692 return false; 693 694 PHINode *PN = 0; 695 BitCastInst *BCI = 0; 696 Value *V = RI->getReturnValue(); 697 if (V) { 698 BCI = dyn_cast<BitCastInst>(V); 699 if (BCI) 700 V = BCI->getOperand(0); 701 702 PN = dyn_cast<PHINode>(V); 703 if (!PN) 704 return false; 705 } 706 707 BasicBlock *BB = RI->getParent(); 708 if (PN && PN->getParent() != BB) 709 return false; 710 711 // It's not safe to eliminate the sign / zero extension of the return value. 712 // See llvm::isInTailCallPosition(). 713 const Function *F = BB->getParent(); 714 Attributes CallerRetAttr = F->getAttributes().getRetAttributes(); 715 if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt)) 716 return false; 717 718 // Make sure there are no instructions between the PHI and return, or that the 719 // return is the first instruction in the block. 720 if (PN) { 721 BasicBlock::iterator BI = BB->begin(); 722 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI)); 723 if (&*BI == BCI) 724 // Also skip over the bitcast. 725 ++BI; 726 if (&*BI != RI) 727 return false; 728 } else { 729 BasicBlock::iterator BI = BB->begin(); 730 while (isa<DbgInfoIntrinsic>(BI)) ++BI; 731 if (&*BI != RI) 732 return false; 733 } 734 735 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail 736 /// call. 737 SmallVector<CallInst*, 4> TailCalls; 738 if (PN) { 739 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) { 740 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I)); 741 // Make sure the phi value is indeed produced by the tail call. 742 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) && 743 TLI->mayBeEmittedAsTailCall(CI)) 744 TailCalls.push_back(CI); 745 } 746 } else { 747 SmallPtrSet<BasicBlock*, 4> VisitedBBs; 748 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) { 749 if (!VisitedBBs.insert(*PI)) 750 continue; 751 752 BasicBlock::InstListType &InstList = (*PI)->getInstList(); 753 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin(); 754 BasicBlock::InstListType::reverse_iterator RE = InstList.rend(); 755 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI)); 756 if (RI == RE) 757 continue; 758 759 CallInst *CI = dyn_cast<CallInst>(&*RI); 760 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI)) 761 TailCalls.push_back(CI); 762 } 763 } 764 765 bool Changed = false; 766 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) { 767 CallInst *CI = TailCalls[i]; 768 CallSite CS(CI); 769 770 // Conservatively require the attributes of the call to match those of the 771 // return. Ignore noalias because it doesn't affect the call sequence. 772 Attributes CalleeRetAttr = CS.getAttributes().getRetAttributes(); 773 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias) 774 continue; 775 776 // Make sure the call instruction is followed by an unconditional branch to 777 // the return block. 778 BasicBlock *CallBB = CI->getParent(); 779 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator()); 780 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB) 781 continue; 782 783 // Duplicate the return into CallBB. 784 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB); 785 ModifiedDT = Changed = true; 786 ++NumRetsDup; 787 } 788 789 // If we eliminated all predecessors of the block, delete the block now. 790 if (Changed && pred_begin(BB) == pred_end(BB)) 791 BB->eraseFromParent(); 792 793 return Changed; 794 } 795 796 //===----------------------------------------------------------------------===// 797 // Memory Optimization 798 //===----------------------------------------------------------------------===// 799 800 /// IsNonLocalValue - Return true if the specified values are defined in a 801 /// different basic block than BB. 802 static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 803 if (Instruction *I = dyn_cast<Instruction>(V)) 804 return I->getParent() != BB; 805 return false; 806 } 807 808 /// OptimizeMemoryInst - Load and Store Instructions often have 809 /// addressing modes that can do significant amounts of computation. As such, 810 /// instruction selection will try to get the load or store to do as much 811 /// computation as possible for the program. The problem is that isel can only 812 /// see within a single block. As such, we sink as much legal addressing mode 813 /// stuff into the block as possible. 814 /// 815 /// This method is used to optimize both load/store and inline asms with memory 816 /// operands. 817 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 818 Type *AccessTy) { 819 Value *Repl = Addr; 820 821 // Try to collapse single-value PHI nodes. This is necessary to undo 822 // unprofitable PRE transformations. 823 SmallVector<Value*, 8> worklist; 824 SmallPtrSet<Value*, 16> Visited; 825 worklist.push_back(Addr); 826 827 // Use a worklist to iteratively look through PHI nodes, and ensure that 828 // the addressing mode obtained from the non-PHI roots of the graph 829 // are equivalent. 830 Value *Consensus = 0; 831 unsigned NumUsesConsensus = 0; 832 bool IsNumUsesConsensusValid = false; 833 SmallVector<Instruction*, 16> AddrModeInsts; 834 ExtAddrMode AddrMode; 835 while (!worklist.empty()) { 836 Value *V = worklist.back(); 837 worklist.pop_back(); 838 839 // Break use-def graph loops. 840 if (!Visited.insert(V)) { 841 Consensus = 0; 842 break; 843 } 844 845 // For a PHI node, push all of its incoming values. 846 if (PHINode *P = dyn_cast<PHINode>(V)) { 847 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) 848 worklist.push_back(P->getIncomingValue(i)); 849 continue; 850 } 851 852 // For non-PHIs, determine the addressing mode being computed. 853 SmallVector<Instruction*, 16> NewAddrModeInsts; 854 ExtAddrMode NewAddrMode = 855 AddressingModeMatcher::Match(V, AccessTy, MemoryInst, 856 NewAddrModeInsts, *TLI); 857 858 // This check is broken into two cases with very similar code to avoid using 859 // getNumUses() as much as possible. Some values have a lot of uses, so 860 // calling getNumUses() unconditionally caused a significant compile-time 861 // regression. 862 if (!Consensus) { 863 Consensus = V; 864 AddrMode = NewAddrMode; 865 AddrModeInsts = NewAddrModeInsts; 866 continue; 867 } else if (NewAddrMode == AddrMode) { 868 if (!IsNumUsesConsensusValid) { 869 NumUsesConsensus = Consensus->getNumUses(); 870 IsNumUsesConsensusValid = true; 871 } 872 873 // Ensure that the obtained addressing mode is equivalent to that obtained 874 // for all other roots of the PHI traversal. Also, when choosing one 875 // such root as representative, select the one with the most uses in order 876 // to keep the cost modeling heuristics in AddressingModeMatcher 877 // applicable. 878 unsigned NumUses = V->getNumUses(); 879 if (NumUses > NumUsesConsensus) { 880 Consensus = V; 881 NumUsesConsensus = NumUses; 882 AddrModeInsts = NewAddrModeInsts; 883 } 884 continue; 885 } 886 887 Consensus = 0; 888 break; 889 } 890 891 // If the addressing mode couldn't be determined, or if multiple different 892 // ones were determined, bail out now. 893 if (!Consensus) return false; 894 895 // Check to see if any of the instructions supersumed by this addr mode are 896 // non-local to I's BB. 897 bool AnyNonLocal = false; 898 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 899 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 900 AnyNonLocal = true; 901 break; 902 } 903 } 904 905 // If all the instructions matched are already in this BB, don't do anything. 906 if (!AnyNonLocal) { 907 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"); 908 return false; 909 } 910 911 // Insert this computation right after this user. Since our caller is 912 // scanning from the top of the BB to the bottom, reuse of the expr are 913 // guaranteed to happen later. 914 IRBuilder<> Builder(MemoryInst); 915 916 // Now that we determined the addressing expression we want to use and know 917 // that we have to sink it into this block. Check to see if we have already 918 // done this for some other load/store instr in this block. If so, reuse the 919 // computation. 920 Value *&SunkAddr = SunkAddrs[Addr]; 921 if (SunkAddr) { 922 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 923 << *MemoryInst); 924 if (SunkAddr->getType() != Addr->getType()) 925 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType()); 926 } else { 927 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 928 << *MemoryInst); 929 Type *IntPtrTy = 930 TLI->getTargetData()->getIntPtrType(AccessTy->getContext()); 931 932 Value *Result = 0; 933 934 // Start with the base register. Do this first so that subsequent address 935 // matching finds it last, which will prevent it from trying to match it 936 // as the scaled value in case it happens to be a mul. That would be 937 // problematic if we've sunk a different mul for the scale, because then 938 // we'd end up sinking both muls. 939 if (AddrMode.BaseReg) { 940 Value *V = AddrMode.BaseReg; 941 if (V->getType()->isPointerTy()) 942 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); 943 if (V->getType() != IntPtrTy) 944 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); 945 Result = V; 946 } 947 948 // Add the scale value. 949 if (AddrMode.Scale) { 950 Value *V = AddrMode.ScaledReg; 951 if (V->getType() == IntPtrTy) { 952 // done. 953 } else if (V->getType()->isPointerTy()) { 954 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); 955 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 956 cast<IntegerType>(V->getType())->getBitWidth()) { 957 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); 958 } else { 959 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr"); 960 } 961 if (AddrMode.Scale != 1) 962 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), 963 "sunkaddr"); 964 if (Result) 965 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 966 else 967 Result = V; 968 } 969 970 // Add in the BaseGV if present. 971 if (AddrMode.BaseGV) { 972 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr"); 973 if (Result) 974 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 975 else 976 Result = V; 977 } 978 979 // Add in the Base Offset if present. 980 if (AddrMode.BaseOffs) { 981 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 982 if (Result) 983 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 984 else 985 Result = V; 986 } 987 988 if (Result == 0) 989 SunkAddr = Constant::getNullValue(Addr->getType()); 990 else 991 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr"); 992 } 993 994 MemoryInst->replaceUsesOfWith(Repl, SunkAddr); 995 996 // If we have no uses, recursively delete the value and all dead instructions 997 // using it. 998 if (Repl->use_empty()) { 999 // This can cause recursive deletion, which can invalidate our iterator. 1000 // Use a WeakVH to hold onto it in case this happens. 1001 WeakVH IterHandle(CurInstIterator); 1002 BasicBlock *BB = CurInstIterator->getParent(); 1003 1004 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo); 1005 1006 if (IterHandle != CurInstIterator) { 1007 // If the iterator instruction was recursively deleted, start over at the 1008 // start of the block. 1009 CurInstIterator = BB->begin(); 1010 SunkAddrs.clear(); 1011 } else { 1012 // This address is now available for reassignment, so erase the table 1013 // entry; we don't want to match some completely different instruction. 1014 SunkAddrs[Addr] = 0; 1015 } 1016 } 1017 ++NumMemoryInsts; 1018 return true; 1019 } 1020 1021 /// OptimizeInlineAsmInst - If there are any memory operands, use 1022 /// OptimizeMemoryInst to sink their address computing into the block when 1023 /// possible / profitable. 1024 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) { 1025 bool MadeChange = false; 1026 1027 TargetLowering::AsmOperandInfoVector 1028 TargetConstraints = TLI->ParseConstraints(CS); 1029 unsigned ArgNo = 0; 1030 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 1031 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; 1032 1033 // Compute the constraint code and ConstraintType to use. 1034 TLI->ComputeConstraintToUse(OpInfo, SDValue()); 1035 1036 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 1037 OpInfo.isIndirect) { 1038 Value *OpVal = CS->getArgOperand(ArgNo++); 1039 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType()); 1040 } else if (OpInfo.Type == InlineAsm::isInput) 1041 ArgNo++; 1042 } 1043 1044 return MadeChange; 1045 } 1046 1047 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same 1048 /// basic block as the load, unless conditions are unfavorable. This allows 1049 /// SelectionDAG to fold the extend into the load. 1050 /// 1051 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) { 1052 // Look for a load being extended. 1053 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0)); 1054 if (!LI) return false; 1055 1056 // If they're already in the same block, there's nothing to do. 1057 if (LI->getParent() == I->getParent()) 1058 return false; 1059 1060 // If the load has other users and the truncate is not free, this probably 1061 // isn't worthwhile. 1062 if (!LI->hasOneUse() && 1063 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) || 1064 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) && 1065 !TLI->isTruncateFree(I->getType(), LI->getType())) 1066 return false; 1067 1068 // Check whether the target supports casts folded into loads. 1069 unsigned LType; 1070 if (isa<ZExtInst>(I)) 1071 LType = ISD::ZEXTLOAD; 1072 else { 1073 assert(isa<SExtInst>(I) && "Unexpected ext type!"); 1074 LType = ISD::SEXTLOAD; 1075 } 1076 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType()))) 1077 return false; 1078 1079 // Move the extend into the same block as the load, so that SelectionDAG 1080 // can fold it. 1081 I->removeFromParent(); 1082 I->insertAfter(LI); 1083 ++NumExtsMoved; 1084 return true; 1085 } 1086 1087 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 1088 BasicBlock *DefBB = I->getParent(); 1089 1090 // If the result of a {s|z}ext and its source are both live out, rewrite all 1091 // other uses of the source with result of extension. 1092 Value *Src = I->getOperand(0); 1093 if (Src->hasOneUse()) 1094 return false; 1095 1096 // Only do this xform if truncating is free. 1097 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 1098 return false; 1099 1100 // Only safe to perform the optimization if the source is also defined in 1101 // this block. 1102 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 1103 return false; 1104 1105 bool DefIsLiveOut = false; 1106 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1107 UI != E; ++UI) { 1108 Instruction *User = cast<Instruction>(*UI); 1109 1110 // Figure out which BB this ext is used in. 1111 BasicBlock *UserBB = User->getParent(); 1112 if (UserBB == DefBB) continue; 1113 DefIsLiveOut = true; 1114 break; 1115 } 1116 if (!DefIsLiveOut) 1117 return false; 1118 1119 // Make sure non of the uses are PHI nodes. 1120 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1121 UI != E; ++UI) { 1122 Instruction *User = cast<Instruction>(*UI); 1123 BasicBlock *UserBB = User->getParent(); 1124 if (UserBB == DefBB) continue; 1125 // Be conservative. We don't want this xform to end up introducing 1126 // reloads just before load / store instructions. 1127 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 1128 return false; 1129 } 1130 1131 // InsertedTruncs - Only insert one trunc in each block once. 1132 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 1133 1134 bool MadeChange = false; 1135 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1136 UI != E; ++UI) { 1137 Use &TheUse = UI.getUse(); 1138 Instruction *User = cast<Instruction>(*UI); 1139 1140 // Figure out which BB this ext is used in. 1141 BasicBlock *UserBB = User->getParent(); 1142 if (UserBB == DefBB) continue; 1143 1144 // Both src and def are live in this block. Rewrite the use. 1145 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 1146 1147 if (!InsertedTrunc) { 1148 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 1149 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 1150 } 1151 1152 // Replace a use of the {s|z}ext source with a use of the result. 1153 TheUse = InsertedTrunc; 1154 ++NumExtUses; 1155 MadeChange = true; 1156 } 1157 1158 return MadeChange; 1159 } 1160 1161 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be 1162 /// turned into an explicit branch. 1163 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) { 1164 // FIXME: This should use the same heuristics as IfConversion to determine 1165 // whether a select is better represented as a branch. This requires that 1166 // branch probability metadata is preserved for the select, which is not the 1167 // case currently. 1168 1169 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 1170 1171 // If the branch is predicted right, an out of order CPU can avoid blocking on 1172 // the compare. Emit cmovs on compares with a memory operand as branches to 1173 // avoid stalls on the load from memory. If the compare has more than one use 1174 // there's probably another cmov or setcc around so it's not worth emitting a 1175 // branch. 1176 if (!Cmp) 1177 return false; 1178 1179 Value *CmpOp0 = Cmp->getOperand(0); 1180 Value *CmpOp1 = Cmp->getOperand(1); 1181 1182 // We check that the memory operand has one use to avoid uses of the loaded 1183 // value directly after the compare, making branches unprofitable. 1184 return Cmp->hasOneUse() && 1185 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) || 1186 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse())); 1187 } 1188 1189 1190 /// If we have a SelectInst that will likely profit from branch prediction, 1191 /// turn it into a branch. 1192 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) { 1193 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); 1194 1195 // Can we convert the 'select' to CF ? 1196 if (DisableSelectToBranch || OptSize || !TLI || VectorCond) 1197 return false; 1198 1199 TargetLowering::SelectSupportKind SelectKind; 1200 if (VectorCond) 1201 SelectKind = TargetLowering::VectorMaskSelect; 1202 else if (SI->getType()->isVectorTy()) 1203 SelectKind = TargetLowering::ScalarCondVectorVal; 1204 else 1205 SelectKind = TargetLowering::ScalarValSelect; 1206 1207 // Do we have efficient codegen support for this kind of 'selects' ? 1208 if (TLI->isSelectSupported(SelectKind)) { 1209 // We have efficient codegen support for the select instruction. 1210 // Check if it is profitable to keep this 'select'. 1211 if (!TLI->isPredictableSelectExpensive() || 1212 !isFormingBranchFromSelectProfitable(SI)) 1213 return false; 1214 } 1215 1216 ModifiedDT = true; 1217 1218 // First, we split the block containing the select into 2 blocks. 1219 BasicBlock *StartBlock = SI->getParent(); 1220 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI)); 1221 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); 1222 1223 // Create a new block serving as the landing pad for the branch. 1224 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid", 1225 NextBlock->getParent(), NextBlock); 1226 1227 // Move the unconditional branch from the block with the select in it into our 1228 // landing pad block. 1229 StartBlock->getTerminator()->eraseFromParent(); 1230 BranchInst::Create(NextBlock, SmallBlock); 1231 1232 // Insert the real conditional branch based on the original condition. 1233 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI); 1234 1235 // The select itself is replaced with a PHI Node. 1236 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin()); 1237 PN->takeName(SI); 1238 PN->addIncoming(SI->getTrueValue(), StartBlock); 1239 PN->addIncoming(SI->getFalseValue(), SmallBlock); 1240 SI->replaceAllUsesWith(PN); 1241 SI->eraseFromParent(); 1242 1243 // Instruct OptimizeBlock to skip to the next block. 1244 CurInstIterator = StartBlock->end(); 1245 ++NumSelectsExpanded; 1246 return true; 1247 } 1248 1249 bool CodeGenPrepare::OptimizeInst(Instruction *I) { 1250 if (PHINode *P = dyn_cast<PHINode>(I)) { 1251 // It is possible for very late stage optimizations (such as SimplifyCFG) 1252 // to introduce PHI nodes too late to be cleaned up. If we detect such a 1253 // trivial PHI, go ahead and zap it here. 1254 if (Value *V = SimplifyInstruction(P)) { 1255 P->replaceAllUsesWith(V); 1256 P->eraseFromParent(); 1257 ++NumPHIsElim; 1258 return true; 1259 } 1260 return false; 1261 } 1262 1263 if (CastInst *CI = dyn_cast<CastInst>(I)) { 1264 // If the source of the cast is a constant, then this should have 1265 // already been constant folded. The only reason NOT to constant fold 1266 // it is if something (e.g. LSR) was careful to place the constant 1267 // evaluation in a block other than then one that uses it (e.g. to hoist 1268 // the address of globals out of a loop). If this is the case, we don't 1269 // want to forward-subst the cast. 1270 if (isa<Constant>(CI->getOperand(0))) 1271 return false; 1272 1273 if (TLI && OptimizeNoopCopyExpression(CI, *TLI)) 1274 return true; 1275 1276 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) { 1277 bool MadeChange = MoveExtToFormExtLoad(I); 1278 return MadeChange | OptimizeExtUses(I); 1279 } 1280 return false; 1281 } 1282 1283 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 1284 return OptimizeCmpExpression(CI); 1285 1286 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1287 if (TLI) 1288 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType()); 1289 return false; 1290 } 1291 1292 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1293 if (TLI) 1294 return OptimizeMemoryInst(I, SI->getOperand(1), 1295 SI->getOperand(0)->getType()); 1296 return false; 1297 } 1298 1299 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1300 if (GEPI->hasAllZeroIndices()) { 1301 /// The GEP operand must be a pointer, so must its result -> BitCast 1302 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1303 GEPI->getName(), GEPI); 1304 GEPI->replaceAllUsesWith(NC); 1305 GEPI->eraseFromParent(); 1306 ++NumGEPsElim; 1307 OptimizeInst(NC); 1308 return true; 1309 } 1310 return false; 1311 } 1312 1313 if (CallInst *CI = dyn_cast<CallInst>(I)) 1314 return OptimizeCallInst(CI); 1315 1316 if (ReturnInst *RI = dyn_cast<ReturnInst>(I)) 1317 return DupRetToEnableTailCallOpts(RI); 1318 1319 if (SelectInst *SI = dyn_cast<SelectInst>(I)) 1320 return OptimizeSelectInst(SI); 1321 1322 return false; 1323 } 1324 1325 // In this pass we look for GEP and cast instructions that are used 1326 // across basic blocks and rewrite them to improve basic-block-at-a-time 1327 // selection. 1328 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 1329 SunkAddrs.clear(); 1330 bool MadeChange = false; 1331 1332 CurInstIterator = BB.begin(); 1333 for (BasicBlock::iterator E = BB.end(); CurInstIterator != E; ) 1334 MadeChange |= OptimizeInst(CurInstIterator++); 1335 1336 return MadeChange; 1337 } 1338 1339 // llvm.dbg.value is far away from the value then iSel may not be able 1340 // handle it properly. iSel will drop llvm.dbg.value if it can not 1341 // find a node corresponding to the value. 1342 bool CodeGenPrepare::PlaceDbgValues(Function &F) { 1343 bool MadeChange = false; 1344 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) { 1345 Instruction *PrevNonDbgInst = NULL; 1346 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) { 1347 Instruction *Insn = BI; ++BI; 1348 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn); 1349 if (!DVI) { 1350 PrevNonDbgInst = Insn; 1351 continue; 1352 } 1353 1354 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue()); 1355 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) { 1356 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI); 1357 DVI->removeFromParent(); 1358 if (isa<PHINode>(VI)) 1359 DVI->insertBefore(VI->getParent()->getFirstInsertionPt()); 1360 else 1361 DVI->insertAfter(VI); 1362 MadeChange = true; 1363 ++NumDbgValueMoved; 1364 } 1365 } 1366 } 1367 return MadeChange; 1368 } 1369