CodeGenPrepare.cpp revision 95bb00414eba82cc1c058b558cd28bc28134c561
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/InlineAsm.h" 22#include "llvm/Instructions.h" 23#include "llvm/IntrinsicInst.h" 24#include "llvm/Pass.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Analysis/ProfileInfo.h" 27#include "llvm/Target/TargetData.h" 28#include "llvm/Target/TargetLowering.h" 29#include "llvm/Transforms/Utils/AddrModeMatcher.h" 30#include "llvm/Transforms/Utils/BasicBlockUtils.h" 31#include "llvm/Transforms/Utils/Local.h" 32#include "llvm/Transforms/Utils/BuildLibCalls.h" 33#include "llvm/ADT/DenseMap.h" 34#include "llvm/ADT/SmallSet.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/Assembly/Writer.h" 37#include "llvm/Support/CallSite.h" 38#include "llvm/Support/CommandLine.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/GetElementPtrTypeIterator.h" 41#include "llvm/Support/PatternMatch.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/Support/IRBuilder.h" 44using namespace llvm; 45using namespace llvm::PatternMatch; 46 47STATISTIC(NumElim, "Number of blocks eliminated"); 48 49static cl::opt<bool> 50CriticalEdgeSplit("cgp-critical-edge-splitting", 51 cl::desc("Split critical edges during codegen prepare"), 52 cl::init(false), cl::Hidden); 53 54namespace { 55 class CodeGenPrepare : public FunctionPass { 56 /// TLI - Keep a pointer of a TargetLowering to consult for determining 57 /// transformation profitability. 58 const TargetLowering *TLI; 59 ProfileInfo *PFI; 60 61 /// BackEdges - Keep a set of all the loop back edges. 62 /// 63 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges; 64 public: 65 static char ID; // Pass identification, replacement for typeid 66 explicit CodeGenPrepare(const TargetLowering *tli = 0) 67 : FunctionPass(ID), TLI(tli) { 68 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); 69 } 70 bool runOnFunction(Function &F); 71 72 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 73 AU.addPreserved<ProfileInfo>(); 74 } 75 76 virtual void releaseMemory() { 77 BackEdges.clear(); 78 } 79 80 private: 81 bool EliminateMostlyEmptyBlocks(Function &F); 82 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 83 void EliminateMostlyEmptyBlock(BasicBlock *BB); 84 bool OptimizeBlock(BasicBlock &BB); 85 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 86 DenseMap<Value*,Value*> &SunkAddrs); 87 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 88 DenseMap<Value*,Value*> &SunkAddrs); 89 bool OptimizeCallInst(CallInst *CI); 90 bool MoveExtToFormExtLoad(Instruction *I); 91 bool OptimizeExtUses(Instruction *I); 92 void findLoopBackEdges(const Function &F); 93 }; 94} 95 96char CodeGenPrepare::ID = 0; 97INITIALIZE_PASS(CodeGenPrepare, "codegenprepare", 98 "Optimize for code generation", false, false) 99 100FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 101 return new CodeGenPrepare(TLI); 102} 103 104/// findLoopBackEdges - Do a DFS walk to find loop back edges. 105/// 106void CodeGenPrepare::findLoopBackEdges(const Function &F) { 107 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 108 FindFunctionBackedges(F, Edges); 109 110 BackEdges.insert(Edges.begin(), Edges.end()); 111} 112 113 114bool CodeGenPrepare::runOnFunction(Function &F) { 115 bool EverMadeChange = false; 116 117 PFI = getAnalysisIfAvailable<ProfileInfo>(); 118 // First pass, eliminate blocks that contain only PHI nodes and an 119 // unconditional branch. 120 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 121 122 // Now find loop back edges, but only if they are being used to decide which 123 // critical edges to split. 124 if (CriticalEdgeSplit) 125 findLoopBackEdges(F); 126 127 bool MadeChange = true; 128 while (MadeChange) { 129 MadeChange = false; 130 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 131 MadeChange |= OptimizeBlock(*BB); 132 EverMadeChange |= MadeChange; 133 } 134 return EverMadeChange; 135} 136 137/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 138/// debug info directives, and an unconditional branch. Passes before isel 139/// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 140/// isel. Start by eliminating these blocks so we can split them the way we 141/// want them. 142bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 143 bool MadeChange = false; 144 // Note that this intentionally skips the entry block. 145 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 146 BasicBlock *BB = I++; 147 148 // If this block doesn't end with an uncond branch, ignore it. 149 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 150 if (!BI || !BI->isUnconditional()) 151 continue; 152 153 // If the instruction before the branch (skipping debug info) isn't a phi 154 // node, then other stuff is happening here. 155 BasicBlock::iterator BBI = BI; 156 if (BBI != BB->begin()) { 157 --BBI; 158 while (isa<DbgInfoIntrinsic>(BBI)) { 159 if (BBI == BB->begin()) 160 break; 161 --BBI; 162 } 163 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 164 continue; 165 } 166 167 // Do not break infinite loops. 168 BasicBlock *DestBB = BI->getSuccessor(0); 169 if (DestBB == BB) 170 continue; 171 172 if (!CanMergeBlocks(BB, DestBB)) 173 continue; 174 175 EliminateMostlyEmptyBlock(BB); 176 MadeChange = true; 177 } 178 return MadeChange; 179} 180 181/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 182/// single uncond branch between them, and BB contains no other non-phi 183/// instructions. 184bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 185 const BasicBlock *DestBB) const { 186 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 187 // the successor. If there are more complex condition (e.g. preheaders), 188 // don't mess around with them. 189 BasicBlock::const_iterator BBI = BB->begin(); 190 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 191 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end(); 192 UI != E; ++UI) { 193 const Instruction *User = cast<Instruction>(*UI); 194 if (User->getParent() != DestBB || !isa<PHINode>(User)) 195 return false; 196 // If User is inside DestBB block and it is a PHINode then check 197 // incoming value. If incoming value is not from BB then this is 198 // a complex condition (e.g. preheaders) we want to avoid here. 199 if (User->getParent() == DestBB) { 200 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 201 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 202 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 203 if (Insn && Insn->getParent() == BB && 204 Insn->getParent() != UPN->getIncomingBlock(I)) 205 return false; 206 } 207 } 208 } 209 } 210 211 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 212 // and DestBB may have conflicting incoming values for the block. If so, we 213 // can't merge the block. 214 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 215 if (!DestBBPN) return true; // no conflict. 216 217 // Collect the preds of BB. 218 SmallPtrSet<const BasicBlock*, 16> BBPreds; 219 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 220 // It is faster to get preds from a PHI than with pred_iterator. 221 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 222 BBPreds.insert(BBPN->getIncomingBlock(i)); 223 } else { 224 BBPreds.insert(pred_begin(BB), pred_end(BB)); 225 } 226 227 // Walk the preds of DestBB. 228 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 229 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 230 if (BBPreds.count(Pred)) { // Common predecessor? 231 BBI = DestBB->begin(); 232 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 233 const Value *V1 = PN->getIncomingValueForBlock(Pred); 234 const Value *V2 = PN->getIncomingValueForBlock(BB); 235 236 // If V2 is a phi node in BB, look up what the mapped value will be. 237 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 238 if (V2PN->getParent() == BB) 239 V2 = V2PN->getIncomingValueForBlock(Pred); 240 241 // If there is a conflict, bail out. 242 if (V1 != V2) return false; 243 } 244 } 245 } 246 247 return true; 248} 249 250 251/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 252/// an unconditional branch in it. 253void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 254 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 255 BasicBlock *DestBB = BI->getSuccessor(0); 256 257 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB); 258 259 // If the destination block has a single pred, then this is a trivial edge, 260 // just collapse it. 261 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 262 if (SinglePred != DestBB) { 263 // Remember if SinglePred was the entry block of the function. If so, we 264 // will need to move BB back to the entry position. 265 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 266 MergeBasicBlockIntoOnlyPred(DestBB, this); 267 268 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 269 BB->moveBefore(&BB->getParent()->getEntryBlock()); 270 271 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 272 return; 273 } 274 } 275 276 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 277 // to handle the new incoming edges it is about to have. 278 PHINode *PN; 279 for (BasicBlock::iterator BBI = DestBB->begin(); 280 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 281 // Remove the incoming value for BB, and remember it. 282 Value *InVal = PN->removeIncomingValue(BB, false); 283 284 // Two options: either the InVal is a phi node defined in BB or it is some 285 // value that dominates BB. 286 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 287 if (InValPhi && InValPhi->getParent() == BB) { 288 // Add all of the input values of the input PHI as inputs of this phi. 289 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 290 PN->addIncoming(InValPhi->getIncomingValue(i), 291 InValPhi->getIncomingBlock(i)); 292 } else { 293 // Otherwise, add one instance of the dominating value for each edge that 294 // we will be adding. 295 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 296 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 297 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 298 } else { 299 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 300 PN->addIncoming(InVal, *PI); 301 } 302 } 303 } 304 305 // The PHIs are now updated, change everything that refers to BB to use 306 // DestBB and remove BB. 307 BB->replaceAllUsesWith(DestBB); 308 if (PFI) { 309 PFI->replaceAllUses(BB, DestBB); 310 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB)); 311 } 312 BB->eraseFromParent(); 313 ++NumElim; 314 315 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 316} 317 318/// FindReusablePredBB - Check all of the predecessors of the block DestPHI 319/// lives in to see if there is a block that we can reuse as a critical edge 320/// from TIBB. 321static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) { 322 BasicBlock *Dest = DestPHI->getParent(); 323 324 /// TIPHIValues - This array is lazily computed to determine the values of 325 /// PHIs in Dest that TI would provide. 326 SmallVector<Value*, 32> TIPHIValues; 327 328 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB. 329 unsigned TIBBEntryNo = 0; 330 331 // Check to see if Dest has any blocks that can be used as a split edge for 332 // this terminator. 333 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) { 334 BasicBlock *Pred = DestPHI->getIncomingBlock(pi); 335 // To be usable, the pred has to end with an uncond branch to the dest. 336 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 337 if (!PredBr || !PredBr->isUnconditional()) 338 continue; 339 // Must be empty other than the branch and debug info. 340 BasicBlock::iterator I = Pred->begin(); 341 while (isa<DbgInfoIntrinsic>(I)) 342 I++; 343 if (&*I != PredBr) 344 continue; 345 // Cannot be the entry block; its label does not get emitted. 346 if (Pred == &Dest->getParent()->getEntryBlock()) 347 continue; 348 349 // Finally, since we know that Dest has phi nodes in it, we have to make 350 // sure that jumping to Pred will have the same effect as going to Dest in 351 // terms of PHI values. 352 PHINode *PN; 353 unsigned PHINo = 0; 354 unsigned PredEntryNo = pi; 355 356 bool FoundMatch = true; 357 for (BasicBlock::iterator I = Dest->begin(); 358 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 359 if (PHINo == TIPHIValues.size()) { 360 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB) 361 TIBBEntryNo = PN->getBasicBlockIndex(TIBB); 362 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo)); 363 } 364 365 // If the PHI entry doesn't work, we can't use this pred. 366 if (PN->getIncomingBlock(PredEntryNo) != Pred) 367 PredEntryNo = PN->getBasicBlockIndex(Pred); 368 369 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) { 370 FoundMatch = false; 371 break; 372 } 373 } 374 375 // If we found a workable predecessor, change TI to branch to Succ. 376 if (FoundMatch) 377 return Pred; 378 } 379 return 0; 380} 381 382 383/// SplitEdgeNicely - Split the critical edge from TI to its specified 384/// successor if it will improve codegen. We only do this if the successor has 385/// phi nodes (otherwise critical edges are ok). If there is already another 386/// predecessor of the succ that is empty (and thus has no phi nodes), use it 387/// instead of introducing a new block. 388static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, 389 SmallSet<std::pair<const BasicBlock*, 390 const BasicBlock*>, 8> &BackEdges, 391 Pass *P) { 392 BasicBlock *TIBB = TI->getParent(); 393 BasicBlock *Dest = TI->getSuccessor(SuccNum); 394 assert(isa<PHINode>(Dest->begin()) && 395 "This should only be called if Dest has a PHI!"); 396 PHINode *DestPHI = cast<PHINode>(Dest->begin()); 397 398 // Do not split edges to EH landing pads. 399 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) 400 if (Invoke->getSuccessor(1) == Dest) 401 return; 402 403 // As a hack, never split backedges of loops. Even though the copy for any 404 // PHIs inserted on the backedge would be dead for exits from the loop, we 405 // assume that the cost of *splitting* the backedge would be too high. 406 if (BackEdges.count(std::make_pair(TIBB, Dest))) 407 return; 408 409 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) { 410 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>(); 411 if (PFI) 412 PFI->splitEdge(TIBB, Dest, ReuseBB); 413 Dest->removePredecessor(TIBB); 414 TI->setSuccessor(SuccNum, ReuseBB); 415 return; 416 } 417 418 SplitCriticalEdge(TI, SuccNum, P, true); 419} 420 421 422/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 423/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), 424/// sink it into user blocks to reduce the number of virtual 425/// registers that must be created and coalesced. 426/// 427/// Return true if any changes are made. 428/// 429static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 430 // If this is a noop copy, 431 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 432 EVT DstVT = TLI.getValueType(CI->getType()); 433 434 // This is an fp<->int conversion? 435 if (SrcVT.isInteger() != DstVT.isInteger()) 436 return false; 437 438 // If this is an extension, it will be a zero or sign extension, which 439 // isn't a noop. 440 if (SrcVT.bitsLT(DstVT)) return false; 441 442 // If these values will be promoted, find out what they will be promoted 443 // to. This helps us consider truncates on PPC as noop copies when they 444 // are. 445 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 446 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); 447 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 448 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); 449 450 // If, after promotion, these are the same types, this is a noop copy. 451 if (SrcVT != DstVT) 452 return false; 453 454 BasicBlock *DefBB = CI->getParent(); 455 456 /// InsertedCasts - Only insert a cast in each block once. 457 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 458 459 bool MadeChange = false; 460 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 461 UI != E; ) { 462 Use &TheUse = UI.getUse(); 463 Instruction *User = cast<Instruction>(*UI); 464 465 // Figure out which BB this cast is used in. For PHI's this is the 466 // appropriate predecessor block. 467 BasicBlock *UserBB = User->getParent(); 468 if (PHINode *PN = dyn_cast<PHINode>(User)) { 469 UserBB = PN->getIncomingBlock(UI); 470 } 471 472 // Preincrement use iterator so we don't invalidate it. 473 ++UI; 474 475 // If this user is in the same block as the cast, don't change the cast. 476 if (UserBB == DefBB) continue; 477 478 // If we have already inserted a cast into this block, use it. 479 CastInst *&InsertedCast = InsertedCasts[UserBB]; 480 481 if (!InsertedCast) { 482 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 483 484 InsertedCast = 485 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 486 InsertPt); 487 MadeChange = true; 488 } 489 490 // Replace a use of the cast with a use of the new cast. 491 TheUse = InsertedCast; 492 } 493 494 // If we removed all uses, nuke the cast. 495 if (CI->use_empty()) { 496 CI->eraseFromParent(); 497 MadeChange = true; 498 } 499 500 return MadeChange; 501} 502 503/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 504/// the number of virtual registers that must be created and coalesced. This is 505/// a clear win except on targets with multiple condition code registers 506/// (PowerPC), where it might lose; some adjustment may be wanted there. 507/// 508/// Return true if any changes are made. 509static bool OptimizeCmpExpression(CmpInst *CI) { 510 BasicBlock *DefBB = CI->getParent(); 511 512 /// InsertedCmp - Only insert a cmp in each block once. 513 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 514 515 bool MadeChange = false; 516 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 517 UI != E; ) { 518 Use &TheUse = UI.getUse(); 519 Instruction *User = cast<Instruction>(*UI); 520 521 // Preincrement use iterator so we don't invalidate it. 522 ++UI; 523 524 // Don't bother for PHI nodes. 525 if (isa<PHINode>(User)) 526 continue; 527 528 // Figure out which BB this cmp is used in. 529 BasicBlock *UserBB = User->getParent(); 530 531 // If this user is in the same block as the cmp, don't change the cmp. 532 if (UserBB == DefBB) continue; 533 534 // If we have already inserted a cmp into this block, use it. 535 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 536 537 if (!InsertedCmp) { 538 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 539 540 InsertedCmp = 541 CmpInst::Create(CI->getOpcode(), 542 CI->getPredicate(), CI->getOperand(0), 543 CI->getOperand(1), "", InsertPt); 544 MadeChange = true; 545 } 546 547 // Replace a use of the cmp with a use of the new cmp. 548 TheUse = InsertedCmp; 549 } 550 551 // If we removed all uses, nuke the cmp. 552 if (CI->use_empty()) 553 CI->eraseFromParent(); 554 555 return MadeChange; 556} 557 558namespace { 559class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls { 560protected: 561 void replaceCall(Value *With) { 562 CI->replaceAllUsesWith(With); 563 CI->eraseFromParent(); 564 } 565 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const { 566 if (ConstantInt *SizeCI = 567 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) 568 return SizeCI->isAllOnesValue(); 569 return false; 570 } 571}; 572} // end anonymous namespace 573 574bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) { 575 // Lower all uses of llvm.objectsize.* 576 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); 577 if (II && II->getIntrinsicID() == Intrinsic::objectsize) { 578 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1); 579 const Type *ReturnTy = CI->getType(); 580 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 581 CI->replaceAllUsesWith(RetVal); 582 CI->eraseFromParent(); 583 return true; 584 } 585 586 // From here on out we're working with named functions. 587 if (CI->getCalledFunction() == 0) return false; 588 589 // We'll need TargetData from here on out. 590 const TargetData *TD = TLI ? TLI->getTargetData() : 0; 591 if (!TD) return false; 592 593 // Lower all default uses of _chk calls. This is very similar 594 // to what InstCombineCalls does, but here we are only lowering calls 595 // that have the default "don't know" as the objectsize. Anything else 596 // should be left alone. 597 CodeGenPrepareFortifiedLibCalls Simplifier; 598 return Simplifier.fold(CI, TD); 599} 600//===----------------------------------------------------------------------===// 601// Memory Optimization 602//===----------------------------------------------------------------------===// 603 604/// IsNonLocalValue - Return true if the specified values are defined in a 605/// different basic block than BB. 606static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 607 if (Instruction *I = dyn_cast<Instruction>(V)) 608 return I->getParent() != BB; 609 return false; 610} 611 612/// OptimizeMemoryInst - Load and Store Instructions often have 613/// addressing modes that can do significant amounts of computation. As such, 614/// instruction selection will try to get the load or store to do as much 615/// computation as possible for the program. The problem is that isel can only 616/// see within a single block. As such, we sink as much legal addressing mode 617/// stuff into the block as possible. 618/// 619/// This method is used to optimize both load/store and inline asms with memory 620/// operands. 621bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 622 const Type *AccessTy, 623 DenseMap<Value*,Value*> &SunkAddrs) { 624 Value *Repl = Addr; 625 626 // Try to collapse single-value PHI nodes. This is necessary to undo 627 // unprofitable PRE transformations. 628 SmallVector<Value*, 8> worklist; 629 SmallPtrSet<Value*, 16> Visited; 630 worklist.push_back(Addr); 631 632 // Use a worklist to iteratively look through PHI nodes, and ensure that 633 // the addressing mode obtained from the non-PHI roots of the graph 634 // are equivalent. 635 Value *Consensus = 0; 636 unsigned NumUses = 0; 637 SmallVector<Instruction*, 16> AddrModeInsts; 638 ExtAddrMode AddrMode; 639 while (!worklist.empty()) { 640 Value *V = worklist.back(); 641 worklist.pop_back(); 642 643 // Break use-def graph loops. 644 if (Visited.count(V)) { 645 Consensus = 0; 646 break; 647 } 648 649 Visited.insert(V); 650 651 // For a PHI node, push all of its incoming values. 652 if (PHINode *P = dyn_cast<PHINode>(V)) { 653 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) 654 worklist.push_back(P->getIncomingValue(i)); 655 continue; 656 } 657 658 // For non-PHIs, determine the addressing mode being computed. 659 SmallVector<Instruction*, 16> NewAddrModeInsts; 660 ExtAddrMode NewAddrMode = 661 AddressingModeMatcher::Match(V, AccessTy,MemoryInst, 662 NewAddrModeInsts, *TLI); 663 664 // Ensure that the obtained addressing mode is equivalent to that obtained 665 // for all other roots of the PHI traversal. Also, when choosing one 666 // such root as representative, select the one with the most uses in order 667 // to keep the cost modeling heuristics in AddressingModeMatcher applicable. 668 if (!Consensus || NewAddrMode == AddrMode) { 669 if (V->getNumUses() > NumUses) { 670 Consensus = V; 671 NumUses = V->getNumUses(); 672 AddrMode = NewAddrMode; 673 AddrModeInsts = NewAddrModeInsts; 674 } 675 continue; 676 } 677 678 Consensus = 0; 679 break; 680 } 681 682 // If the addressing mode couldn't be determined, or if multiple different 683 // ones were determined, bail out now. 684 if (!Consensus) return false; 685 686 // Check to see if any of the instructions supersumed by this addr mode are 687 // non-local to I's BB. 688 bool AnyNonLocal = false; 689 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 690 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 691 AnyNonLocal = true; 692 break; 693 } 694 } 695 696 // If all the instructions matched are already in this BB, don't do anything. 697 if (!AnyNonLocal) { 698 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"); 699 return false; 700 } 701 702 // Insert this computation right after this user. Since our caller is 703 // scanning from the top of the BB to the bottom, reuse of the expr are 704 // guaranteed to happen later. 705 BasicBlock::iterator InsertPt = MemoryInst; 706 707 // Now that we determined the addressing expression we want to use and know 708 // that we have to sink it into this block. Check to see if we have already 709 // done this for some other load/store instr in this block. If so, reuse the 710 // computation. 711 Value *&SunkAddr = SunkAddrs[Addr]; 712 if (SunkAddr) { 713 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 714 << *MemoryInst); 715 if (SunkAddr->getType() != Addr->getType()) 716 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 717 } else { 718 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 719 << *MemoryInst); 720 const Type *IntPtrTy = 721 TLI->getTargetData()->getIntPtrType(AccessTy->getContext()); 722 723 Value *Result = 0; 724 725 // Start with the base register. Do this first so that subsequent address 726 // matching finds it last, which will prevent it from trying to match it 727 // as the scaled value in case it happens to be a mul. That would be 728 // problematic if we've sunk a different mul for the scale, because then 729 // we'd end up sinking both muls. 730 if (AddrMode.BaseReg) { 731 Value *V = AddrMode.BaseReg; 732 if (V->getType()->isPointerTy()) 733 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 734 if (V->getType() != IntPtrTy) 735 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true, 736 "sunkaddr", InsertPt); 737 Result = V; 738 } 739 740 // Add the scale value. 741 if (AddrMode.Scale) { 742 Value *V = AddrMode.ScaledReg; 743 if (V->getType() == IntPtrTy) { 744 // done. 745 } else if (V->getType()->isPointerTy()) { 746 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 747 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 748 cast<IntegerType>(V->getType())->getBitWidth()) { 749 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 750 } else { 751 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 752 } 753 if (AddrMode.Scale != 1) 754 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 755 AddrMode.Scale), 756 "sunkaddr", InsertPt); 757 if (Result) 758 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 759 else 760 Result = V; 761 } 762 763 // Add in the BaseGV if present. 764 if (AddrMode.BaseGV) { 765 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 766 InsertPt); 767 if (Result) 768 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 769 else 770 Result = V; 771 } 772 773 // Add in the Base Offset if present. 774 if (AddrMode.BaseOffs) { 775 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 776 if (Result) 777 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 778 else 779 Result = V; 780 } 781 782 if (Result == 0) 783 SunkAddr = Constant::getNullValue(Addr->getType()); 784 else 785 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 786 } 787 788 MemoryInst->replaceUsesOfWith(Repl, SunkAddr); 789 790 if (Repl->use_empty()) { 791 RecursivelyDeleteTriviallyDeadInstructions(Repl); 792 // This address is now available for reassignment, so erase the table entry; 793 // we don't want to match some completely different instruction. 794 SunkAddrs[Addr] = 0; 795 } 796 return true; 797} 798 799/// OptimizeInlineAsmInst - If there are any memory operands, use 800/// OptimizeMemoryInst to sink their address computing into the block when 801/// possible / profitable. 802bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 803 DenseMap<Value*,Value*> &SunkAddrs) { 804 bool MadeChange = false; 805 806 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI->ParseConstraints(CS); 807 unsigned ArgNo = 0; 808 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 809 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; 810 811 // Compute the constraint code and ConstraintType to use. 812 TLI->ComputeConstraintToUse(OpInfo, SDValue()); 813 814 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 815 OpInfo.isIndirect) { 816 Value *OpVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 817 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 818 } else if (OpInfo.Type == InlineAsm::isInput) 819 ArgNo++; 820 } 821 822 return MadeChange; 823} 824 825/// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same 826/// basic block as the load, unless conditions are unfavorable. This allows 827/// SelectionDAG to fold the extend into the load. 828/// 829bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) { 830 // Look for a load being extended. 831 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0)); 832 if (!LI) return false; 833 834 // If they're already in the same block, there's nothing to do. 835 if (LI->getParent() == I->getParent()) 836 return false; 837 838 // If the load has other users and the truncate is not free, this probably 839 // isn't worthwhile. 840 if (!LI->hasOneUse() && 841 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) || 842 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) && 843 !TLI->isTruncateFree(I->getType(), LI->getType())) 844 return false; 845 846 // Check whether the target supports casts folded into loads. 847 unsigned LType; 848 if (isa<ZExtInst>(I)) 849 LType = ISD::ZEXTLOAD; 850 else { 851 assert(isa<SExtInst>(I) && "Unexpected ext type!"); 852 LType = ISD::SEXTLOAD; 853 } 854 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType()))) 855 return false; 856 857 // Move the extend into the same block as the load, so that SelectionDAG 858 // can fold it. 859 I->removeFromParent(); 860 I->insertAfter(LI); 861 return true; 862} 863 864bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 865 BasicBlock *DefBB = I->getParent(); 866 867 // If the result of a {s|z}ext and its source are both live out, rewrite all 868 // other uses of the source with result of extension. 869 Value *Src = I->getOperand(0); 870 if (Src->hasOneUse()) 871 return false; 872 873 // Only do this xform if truncating is free. 874 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 875 return false; 876 877 // Only safe to perform the optimization if the source is also defined in 878 // this block. 879 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 880 return false; 881 882 bool DefIsLiveOut = false; 883 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 884 UI != E; ++UI) { 885 Instruction *User = cast<Instruction>(*UI); 886 887 // Figure out which BB this ext is used in. 888 BasicBlock *UserBB = User->getParent(); 889 if (UserBB == DefBB) continue; 890 DefIsLiveOut = true; 891 break; 892 } 893 if (!DefIsLiveOut) 894 return false; 895 896 // Make sure non of the uses are PHI nodes. 897 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 898 UI != E; ++UI) { 899 Instruction *User = cast<Instruction>(*UI); 900 BasicBlock *UserBB = User->getParent(); 901 if (UserBB == DefBB) continue; 902 // Be conservative. We don't want this xform to end up introducing 903 // reloads just before load / store instructions. 904 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 905 return false; 906 } 907 908 // InsertedTruncs - Only insert one trunc in each block once. 909 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 910 911 bool MadeChange = false; 912 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 913 UI != E; ++UI) { 914 Use &TheUse = UI.getUse(); 915 Instruction *User = cast<Instruction>(*UI); 916 917 // Figure out which BB this ext is used in. 918 BasicBlock *UserBB = User->getParent(); 919 if (UserBB == DefBB) continue; 920 921 // Both src and def are live in this block. Rewrite the use. 922 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 923 924 if (!InsertedTrunc) { 925 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 926 927 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 928 } 929 930 // Replace a use of the {s|z}ext source with a use of the result. 931 TheUse = InsertedTrunc; 932 933 MadeChange = true; 934 } 935 936 return MadeChange; 937} 938 939// In this pass we look for GEP and cast instructions that are used 940// across basic blocks and rewrite them to improve basic-block-at-a-time 941// selection. 942bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 943 bool MadeChange = false; 944 945 // Split all critical edges where the dest block has a PHI. 946 if (CriticalEdgeSplit) { 947 TerminatorInst *BBTI = BB.getTerminator(); 948 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) { 949 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) { 950 BasicBlock *SuccBB = BBTI->getSuccessor(i); 951 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true)) 952 SplitEdgeNicely(BBTI, i, BackEdges, this); 953 } 954 } 955 } 956 957 // Keep track of non-local addresses that have been sunk into this block. 958 // This allows us to avoid inserting duplicate code for blocks with multiple 959 // load/stores of the same address. 960 DenseMap<Value*, Value*> SunkAddrs; 961 962 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 963 Instruction *I = BBI++; 964 965 if (PHINode *P = dyn_cast<PHINode>(I)) { 966 // It is possible for very late stage optimizations (such as SimplifyCFG) 967 // to introduce PHI nodes too late to be cleaned up. If we detect such a 968 // trivial PHI, go ahead and zap it here. 969 if (Value *V = SimplifyInstruction(P)) { 970 P->replaceAllUsesWith(V); 971 P->eraseFromParent(); 972 } 973 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 974 // If the source of the cast is a constant, then this should have 975 // already been constant folded. The only reason NOT to constant fold 976 // it is if something (e.g. LSR) was careful to place the constant 977 // evaluation in a block other than then one that uses it (e.g. to hoist 978 // the address of globals out of a loop). If this is the case, we don't 979 // want to forward-subst the cast. 980 if (isa<Constant>(CI->getOperand(0))) 981 continue; 982 983 bool Change = false; 984 if (TLI) { 985 Change = OptimizeNoopCopyExpression(CI, *TLI); 986 MadeChange |= Change; 987 } 988 989 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) { 990 MadeChange |= MoveExtToFormExtLoad(I); 991 MadeChange |= OptimizeExtUses(I); 992 } 993 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 994 MadeChange |= OptimizeCmpExpression(CI); 995 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 996 if (TLI) 997 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 998 SunkAddrs); 999 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1000 if (TLI) 1001 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 1002 SI->getOperand(0)->getType(), 1003 SunkAddrs); 1004 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1005 if (GEPI->hasAllZeroIndices()) { 1006 /// The GEP operand must be a pointer, so must its result -> BitCast 1007 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1008 GEPI->getName(), GEPI); 1009 GEPI->replaceAllUsesWith(NC); 1010 GEPI->eraseFromParent(); 1011 MadeChange = true; 1012 BBI = NC; 1013 } 1014 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 1015 // If we found an inline asm expession, and if the target knows how to 1016 // lower it to normal LLVM code, do so now. 1017 if (TLI && isa<InlineAsm>(CI->getCalledValue())) { 1018 if (TLI->ExpandInlineAsm(CI)) { 1019 BBI = BB.begin(); 1020 // Avoid processing instructions out of order, which could cause 1021 // reuse before a value is defined. 1022 SunkAddrs.clear(); 1023 } else 1024 // Sink address computing for memory operands into the block. 1025 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 1026 } else { 1027 // Other CallInst optimizations that don't need to muck with the 1028 // enclosing iterator here. 1029 MadeChange |= OptimizeCallInst(CI); 1030 } 1031 } 1032 } 1033 1034 return MadeChange; 1035} 1036