CodeGenPrepare.cpp revision 1df9859c40492511b8aa4321eb76496005d3b75b
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/ProfileInfo.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Target/TargetLowering.h" 28#include "llvm/Transforms/Utils/AddrModeMatcher.h" 29#include "llvm/Transforms/Utils/BasicBlockUtils.h" 30#include "llvm/Transforms/Utils/Local.h" 31#include "llvm/ADT/DenseMap.h" 32#include "llvm/ADT/SmallSet.h" 33#include "llvm/Assembly/Writer.h" 34#include "llvm/Support/CallSite.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/GetElementPtrTypeIterator.h" 37#include "llvm/Support/PatternMatch.h" 38#include "llvm/Support/raw_ostream.h" 39using namespace llvm; 40using namespace llvm::PatternMatch; 41 42namespace { 43 class CodeGenPrepare : public FunctionPass { 44 /// TLI - Keep a pointer of a TargetLowering to consult for determining 45 /// transformation profitability. 46 const TargetLowering *TLI; 47 ProfileInfo *PFI; 48 49 /// BackEdges - Keep a set of all the loop back edges. 50 /// 51 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges; 52 public: 53 static char ID; // Pass identification, replacement for typeid 54 explicit CodeGenPrepare(const TargetLowering *tli = 0) 55 : FunctionPass(&ID), TLI(tli) {} 56 bool runOnFunction(Function &F); 57 58 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 59 AU.addPreserved<ProfileInfo>(); 60 } 61 62 virtual void releaseMemory() { 63 BackEdges.clear(); 64 } 65 66 private: 67 bool EliminateMostlyEmptyBlocks(Function &F); 68 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 69 void EliminateMostlyEmptyBlock(BasicBlock *BB); 70 bool OptimizeBlock(BasicBlock &BB); 71 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 72 DenseMap<Value*,Value*> &SunkAddrs); 73 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 74 DenseMap<Value*,Value*> &SunkAddrs); 75 bool MoveExtToFormExtLoad(Instruction *I); 76 bool OptimizeExtUses(Instruction *I); 77 void findLoopBackEdges(const Function &F); 78 }; 79} 80 81char CodeGenPrepare::ID = 0; 82static RegisterPass<CodeGenPrepare> X("codegenprepare", 83 "Optimize for code generation"); 84 85FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 86 return new CodeGenPrepare(TLI); 87} 88 89/// findLoopBackEdges - Do a DFS walk to find loop back edges. 90/// 91void CodeGenPrepare::findLoopBackEdges(const Function &F) { 92 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 93 FindFunctionBackedges(F, Edges); 94 95 BackEdges.insert(Edges.begin(), Edges.end()); 96} 97 98 99bool CodeGenPrepare::runOnFunction(Function &F) { 100 bool EverMadeChange = false; 101 102 PFI = getAnalysisIfAvailable<ProfileInfo>(); 103 // First pass, eliminate blocks that contain only PHI nodes and an 104 // unconditional branch. 105 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 106 107 // Now find loop back edges. 108 findLoopBackEdges(F); 109 110 bool MadeChange = true; 111 while (MadeChange) { 112 MadeChange = false; 113 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 114 MadeChange |= OptimizeBlock(*BB); 115 EverMadeChange |= MadeChange; 116 } 117 return EverMadeChange; 118} 119 120/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 121/// debug info directives, and an unconditional branch. Passes before isel 122/// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 123/// isel. Start by eliminating these blocks so we can split them the way we 124/// want them. 125bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 126 bool MadeChange = false; 127 // Note that this intentionally skips the entry block. 128 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 129 BasicBlock *BB = I++; 130 131 // If this block doesn't end with an uncond branch, ignore it. 132 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 133 if (!BI || !BI->isUnconditional()) 134 continue; 135 136 // If the instruction before the branch (skipping debug info) isn't a phi 137 // node, then other stuff is happening here. 138 BasicBlock::iterator BBI = BI; 139 if (BBI != BB->begin()) { 140 --BBI; 141 while (isa<DbgInfoIntrinsic>(BBI)) { 142 if (BBI == BB->begin()) 143 break; 144 --BBI; 145 } 146 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 147 continue; 148 } 149 150 // Do not break infinite loops. 151 BasicBlock *DestBB = BI->getSuccessor(0); 152 if (DestBB == BB) 153 continue; 154 155 if (!CanMergeBlocks(BB, DestBB)) 156 continue; 157 158 EliminateMostlyEmptyBlock(BB); 159 MadeChange = true; 160 } 161 return MadeChange; 162} 163 164/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 165/// single uncond branch between them, and BB contains no other non-phi 166/// instructions. 167bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 168 const BasicBlock *DestBB) const { 169 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 170 // the successor. If there are more complex condition (e.g. preheaders), 171 // don't mess around with them. 172 BasicBlock::const_iterator BBI = BB->begin(); 173 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 174 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 175 UI != E; ++UI) { 176 const Instruction *User = cast<Instruction>(*UI); 177 if (User->getParent() != DestBB || !isa<PHINode>(User)) 178 return false; 179 // If User is inside DestBB block and it is a PHINode then check 180 // incoming value. If incoming value is not from BB then this is 181 // a complex condition (e.g. preheaders) we want to avoid here. 182 if (User->getParent() == DestBB) { 183 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 184 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 185 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 186 if (Insn && Insn->getParent() == BB && 187 Insn->getParent() != UPN->getIncomingBlock(I)) 188 return false; 189 } 190 } 191 } 192 } 193 194 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 195 // and DestBB may have conflicting incoming values for the block. If so, we 196 // can't merge the block. 197 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 198 if (!DestBBPN) return true; // no conflict. 199 200 // Collect the preds of BB. 201 SmallPtrSet<const BasicBlock*, 16> BBPreds; 202 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 203 // It is faster to get preds from a PHI than with pred_iterator. 204 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 205 BBPreds.insert(BBPN->getIncomingBlock(i)); 206 } else { 207 BBPreds.insert(pred_begin(BB), pred_end(BB)); 208 } 209 210 // Walk the preds of DestBB. 211 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 212 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 213 if (BBPreds.count(Pred)) { // Common predecessor? 214 BBI = DestBB->begin(); 215 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 216 const Value *V1 = PN->getIncomingValueForBlock(Pred); 217 const Value *V2 = PN->getIncomingValueForBlock(BB); 218 219 // If V2 is a phi node in BB, look up what the mapped value will be. 220 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 221 if (V2PN->getParent() == BB) 222 V2 = V2PN->getIncomingValueForBlock(Pred); 223 224 // If there is a conflict, bail out. 225 if (V1 != V2) return false; 226 } 227 } 228 } 229 230 return true; 231} 232 233 234/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 235/// an unconditional branch in it. 236void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 237 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 238 BasicBlock *DestBB = BI->getSuccessor(0); 239 240 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB); 241 242 // If the destination block has a single pred, then this is a trivial edge, 243 // just collapse it. 244 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 245 if (SinglePred != DestBB) { 246 // Remember if SinglePred was the entry block of the function. If so, we 247 // will need to move BB back to the entry position. 248 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 249 MergeBasicBlockIntoOnlyPred(DestBB, this); 250 251 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 252 BB->moveBefore(&BB->getParent()->getEntryBlock()); 253 254 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 255 return; 256 } 257 } 258 259 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 260 // to handle the new incoming edges it is about to have. 261 PHINode *PN; 262 for (BasicBlock::iterator BBI = DestBB->begin(); 263 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 264 // Remove the incoming value for BB, and remember it. 265 Value *InVal = PN->removeIncomingValue(BB, false); 266 267 // Two options: either the InVal is a phi node defined in BB or it is some 268 // value that dominates BB. 269 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 270 if (InValPhi && InValPhi->getParent() == BB) { 271 // Add all of the input values of the input PHI as inputs of this phi. 272 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 273 PN->addIncoming(InValPhi->getIncomingValue(i), 274 InValPhi->getIncomingBlock(i)); 275 } else { 276 // Otherwise, add one instance of the dominating value for each edge that 277 // we will be adding. 278 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 279 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 280 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 281 } else { 282 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 283 PN->addIncoming(InVal, *PI); 284 } 285 } 286 } 287 288 // The PHIs are now updated, change everything that refers to BB to use 289 // DestBB and remove BB. 290 BB->replaceAllUsesWith(DestBB); 291 if (PFI) { 292 PFI->replaceAllUses(BB, DestBB); 293 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB)); 294 } 295 BB->eraseFromParent(); 296 297 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 298} 299 300/// FindReusablePredBB - Check all of the predecessors of the block DestPHI 301/// lives in to see if there is a block that we can reuse as a critical edge 302/// from TIBB. 303static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) { 304 BasicBlock *Dest = DestPHI->getParent(); 305 306 /// TIPHIValues - This array is lazily computed to determine the values of 307 /// PHIs in Dest that TI would provide. 308 SmallVector<Value*, 32> TIPHIValues; 309 310 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB. 311 unsigned TIBBEntryNo = 0; 312 313 // Check to see if Dest has any blocks that can be used as a split edge for 314 // this terminator. 315 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) { 316 BasicBlock *Pred = DestPHI->getIncomingBlock(pi); 317 // To be usable, the pred has to end with an uncond branch to the dest. 318 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 319 if (!PredBr || !PredBr->isUnconditional()) 320 continue; 321 // Must be empty other than the branch and debug info. 322 BasicBlock::iterator I = Pred->begin(); 323 while (isa<DbgInfoIntrinsic>(I)) 324 I++; 325 if (&*I != PredBr) 326 continue; 327 // Cannot be the entry block; its label does not get emitted. 328 if (Pred == &Dest->getParent()->getEntryBlock()) 329 continue; 330 331 // Finally, since we know that Dest has phi nodes in it, we have to make 332 // sure that jumping to Pred will have the same effect as going to Dest in 333 // terms of PHI values. 334 PHINode *PN; 335 unsigned PHINo = 0; 336 unsigned PredEntryNo = pi; 337 338 bool FoundMatch = true; 339 for (BasicBlock::iterator I = Dest->begin(); 340 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 341 if (PHINo == TIPHIValues.size()) { 342 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB) 343 TIBBEntryNo = PN->getBasicBlockIndex(TIBB); 344 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo)); 345 } 346 347 // If the PHI entry doesn't work, we can't use this pred. 348 if (PN->getIncomingBlock(PredEntryNo) != Pred) 349 PredEntryNo = PN->getBasicBlockIndex(Pred); 350 351 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) { 352 FoundMatch = false; 353 break; 354 } 355 } 356 357 // If we found a workable predecessor, change TI to branch to Succ. 358 if (FoundMatch) 359 return Pred; 360 } 361 return 0; 362} 363 364 365/// SplitEdgeNicely - Split the critical edge from TI to its specified 366/// successor if it will improve codegen. We only do this if the successor has 367/// phi nodes (otherwise critical edges are ok). If there is already another 368/// predecessor of the succ that is empty (and thus has no phi nodes), use it 369/// instead of introducing a new block. 370static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, 371 SmallSet<std::pair<const BasicBlock*, 372 const BasicBlock*>, 8> &BackEdges, 373 Pass *P) { 374 BasicBlock *TIBB = TI->getParent(); 375 BasicBlock *Dest = TI->getSuccessor(SuccNum); 376 assert(isa<PHINode>(Dest->begin()) && 377 "This should only be called if Dest has a PHI!"); 378 PHINode *DestPHI = cast<PHINode>(Dest->begin()); 379 380 // Do not split edges to EH landing pads. 381 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) 382 if (Invoke->getSuccessor(1) == Dest) 383 return; 384 385 // As a hack, never split backedges of loops. Even though the copy for any 386 // PHIs inserted on the backedge would be dead for exits from the loop, we 387 // assume that the cost of *splitting* the backedge would be too high. 388 if (BackEdges.count(std::make_pair(TIBB, Dest))) 389 return; 390 391 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) { 392 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>(); 393 if (PFI) 394 PFI->splitEdge(TIBB, Dest, ReuseBB); 395 Dest->removePredecessor(TIBB); 396 TI->setSuccessor(SuccNum, ReuseBB); 397 return; 398 } 399 400 SplitCriticalEdge(TI, SuccNum, P, true); 401} 402 403 404/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 405/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), 406/// sink it into user blocks to reduce the number of virtual 407/// registers that must be created and coalesced. 408/// 409/// Return true if any changes are made. 410/// 411static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 412 // If this is a noop copy, 413 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 414 EVT DstVT = TLI.getValueType(CI->getType()); 415 416 // This is an fp<->int conversion? 417 if (SrcVT.isInteger() != DstVT.isInteger()) 418 return false; 419 420 // If this is an extension, it will be a zero or sign extension, which 421 // isn't a noop. 422 if (SrcVT.bitsLT(DstVT)) return false; 423 424 // If these values will be promoted, find out what they will be promoted 425 // to. This helps us consider truncates on PPC as noop copies when they 426 // are. 427 if (TLI.getTypeAction(CI->getContext(), SrcVT) == TargetLowering::Promote) 428 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); 429 if (TLI.getTypeAction(CI->getContext(), DstVT) == TargetLowering::Promote) 430 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); 431 432 // If, after promotion, these are the same types, this is a noop copy. 433 if (SrcVT != DstVT) 434 return false; 435 436 BasicBlock *DefBB = CI->getParent(); 437 438 /// InsertedCasts - Only insert a cast in each block once. 439 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 440 441 bool MadeChange = false; 442 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 443 UI != E; ) { 444 Use &TheUse = UI.getUse(); 445 Instruction *User = cast<Instruction>(*UI); 446 447 // Figure out which BB this cast is used in. For PHI's this is the 448 // appropriate predecessor block. 449 BasicBlock *UserBB = User->getParent(); 450 if (PHINode *PN = dyn_cast<PHINode>(User)) { 451 UserBB = PN->getIncomingBlock(UI); 452 } 453 454 // Preincrement use iterator so we don't invalidate it. 455 ++UI; 456 457 // If this user is in the same block as the cast, don't change the cast. 458 if (UserBB == DefBB) continue; 459 460 // If we have already inserted a cast into this block, use it. 461 CastInst *&InsertedCast = InsertedCasts[UserBB]; 462 463 if (!InsertedCast) { 464 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 465 466 InsertedCast = 467 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 468 InsertPt); 469 MadeChange = true; 470 } 471 472 // Replace a use of the cast with a use of the new cast. 473 TheUse = InsertedCast; 474 } 475 476 // If we removed all uses, nuke the cast. 477 if (CI->use_empty()) { 478 CI->eraseFromParent(); 479 MadeChange = true; 480 } 481 482 return MadeChange; 483} 484 485/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 486/// the number of virtual registers that must be created and coalesced. This is 487/// a clear win except on targets with multiple condition code registers 488/// (PowerPC), where it might lose; some adjustment may be wanted there. 489/// 490/// Return true if any changes are made. 491static bool OptimizeCmpExpression(CmpInst *CI) { 492 BasicBlock *DefBB = CI->getParent(); 493 494 /// InsertedCmp - Only insert a cmp in each block once. 495 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 496 497 bool MadeChange = false; 498 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 499 UI != E; ) { 500 Use &TheUse = UI.getUse(); 501 Instruction *User = cast<Instruction>(*UI); 502 503 // Preincrement use iterator so we don't invalidate it. 504 ++UI; 505 506 // Don't bother for PHI nodes. 507 if (isa<PHINode>(User)) 508 continue; 509 510 // Figure out which BB this cmp is used in. 511 BasicBlock *UserBB = User->getParent(); 512 513 // If this user is in the same block as the cmp, don't change the cmp. 514 if (UserBB == DefBB) continue; 515 516 // If we have already inserted a cmp into this block, use it. 517 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 518 519 if (!InsertedCmp) { 520 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 521 522 InsertedCmp = 523 CmpInst::Create(CI->getOpcode(), 524 CI->getPredicate(), CI->getOperand(0), 525 CI->getOperand(1), "", InsertPt); 526 MadeChange = true; 527 } 528 529 // Replace a use of the cmp with a use of the new cmp. 530 TheUse = InsertedCmp; 531 } 532 533 // If we removed all uses, nuke the cmp. 534 if (CI->use_empty()) 535 CI->eraseFromParent(); 536 537 return MadeChange; 538} 539 540//===----------------------------------------------------------------------===// 541// Memory Optimization 542//===----------------------------------------------------------------------===// 543 544/// IsNonLocalValue - Return true if the specified values are defined in a 545/// different basic block than BB. 546static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 547 if (Instruction *I = dyn_cast<Instruction>(V)) 548 return I->getParent() != BB; 549 return false; 550} 551 552/// OptimizeMemoryInst - Load and Store Instructions often have 553/// addressing modes that can do significant amounts of computation. As such, 554/// instruction selection will try to get the load or store to do as much 555/// computation as possible for the program. The problem is that isel can only 556/// see within a single block. As such, we sink as much legal addressing mode 557/// stuff into the block as possible. 558/// 559/// This method is used to optimize both load/store and inline asms with memory 560/// operands. 561bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 562 const Type *AccessTy, 563 DenseMap<Value*,Value*> &SunkAddrs) { 564 // Figure out what addressing mode will be built up for this operation. 565 SmallVector<Instruction*, 16> AddrModeInsts; 566 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst, 567 AddrModeInsts, *TLI); 568 569 // Check to see if any of the instructions supersumed by this addr mode are 570 // non-local to I's BB. 571 bool AnyNonLocal = false; 572 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 573 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 574 AnyNonLocal = true; 575 break; 576 } 577 } 578 579 // If all the instructions matched are already in this BB, don't do anything. 580 if (!AnyNonLocal) { 581 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"); 582 return false; 583 } 584 585 // Insert this computation right after this user. Since our caller is 586 // scanning from the top of the BB to the bottom, reuse of the expr are 587 // guaranteed to happen later. 588 BasicBlock::iterator InsertPt = MemoryInst; 589 590 // Now that we determined the addressing expression we want to use and know 591 // that we have to sink it into this block. Check to see if we have already 592 // done this for some other load/store instr in this block. If so, reuse the 593 // computation. 594 Value *&SunkAddr = SunkAddrs[Addr]; 595 if (SunkAddr) { 596 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 597 << *MemoryInst); 598 if (SunkAddr->getType() != Addr->getType()) 599 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 600 } else { 601 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 602 << *MemoryInst); 603 const Type *IntPtrTy = 604 TLI->getTargetData()->getIntPtrType(AccessTy->getContext()); 605 606 Value *Result = 0; 607 608 // Start with the base register. Do this first so that subsequent address 609 // matching finds it last, which will prevent it from trying to match it 610 // as the scaled value in case it happens to be a mul. That would be 611 // problematic if we've sunk a different mul for the scale, because then 612 // we'd end up sinking both muls. 613 if (AddrMode.BaseReg) { 614 Value *V = AddrMode.BaseReg; 615 if (V->getType()->isPointerTy()) 616 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 617 if (V->getType() != IntPtrTy) 618 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true, 619 "sunkaddr", InsertPt); 620 Result = V; 621 } 622 623 // Add the scale value. 624 if (AddrMode.Scale) { 625 Value *V = AddrMode.ScaledReg; 626 if (V->getType() == IntPtrTy) { 627 // done. 628 } else if (V->getType()->isPointerTy()) { 629 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 630 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 631 cast<IntegerType>(V->getType())->getBitWidth()) { 632 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 633 } else { 634 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 635 } 636 if (AddrMode.Scale != 1) 637 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 638 AddrMode.Scale), 639 "sunkaddr", InsertPt); 640 if (Result) 641 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 642 else 643 Result = V; 644 } 645 646 // Add in the BaseGV if present. 647 if (AddrMode.BaseGV) { 648 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 649 InsertPt); 650 if (Result) 651 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 652 else 653 Result = V; 654 } 655 656 // Add in the Base Offset if present. 657 if (AddrMode.BaseOffs) { 658 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 659 if (Result) 660 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 661 else 662 Result = V; 663 } 664 665 if (Result == 0) 666 SunkAddr = Constant::getNullValue(Addr->getType()); 667 else 668 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 669 } 670 671 MemoryInst->replaceUsesOfWith(Addr, SunkAddr); 672 673 if (Addr->use_empty()) 674 RecursivelyDeleteTriviallyDeadInstructions(Addr); 675 return true; 676} 677 678/// OptimizeInlineAsmInst - If there are any memory operands, use 679/// OptimizeMemoryInst to sink their address computing into the block when 680/// possible / profitable. 681bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 682 DenseMap<Value*,Value*> &SunkAddrs) { 683 bool MadeChange = false; 684 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 685 686 // Do a prepass over the constraints, canonicalizing them, and building up the 687 // ConstraintOperands list. 688 std::vector<InlineAsm::ConstraintInfo> 689 ConstraintInfos = IA->ParseConstraints(); 690 691 /// ConstraintOperands - Information about all of the constraints. 692 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 693 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 694 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 695 ConstraintOperands. 696 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 697 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 698 699 // Compute the value type for each operand. 700 switch (OpInfo.Type) { 701 case InlineAsm::isOutput: 702 if (OpInfo.isIndirect) 703 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 704 break; 705 case InlineAsm::isInput: 706 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 707 break; 708 case InlineAsm::isClobber: 709 // Nothing to do. 710 break; 711 } 712 713 // Compute the constraint code and ConstraintType to use. 714 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 715 OpInfo.ConstraintType == TargetLowering::C_Memory); 716 717 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 718 OpInfo.isIndirect) { 719 Value *OpVal = OpInfo.CallOperandVal; 720 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 721 } 722 } 723 724 return MadeChange; 725} 726 727/// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same 728/// basic block as the load, unless conditions are unfavorable. This allows 729/// SelectionDAG to fold the extend into the load. 730/// 731bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) { 732 // Look for a load being extended. 733 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0)); 734 if (!LI) return false; 735 736 // If they're already in the same block, there's nothing to do. 737 if (LI->getParent() == I->getParent()) 738 return false; 739 740 // If the load has other users and the truncate is not free, this probably 741 // isn't worthwhile. 742 if (!LI->hasOneUse() && 743 TLI && !TLI->isTruncateFree(I->getType(), LI->getType())) 744 return false; 745 746 // Check whether the target supports casts folded into loads. 747 unsigned LType; 748 if (isa<ZExtInst>(I)) 749 LType = ISD::ZEXTLOAD; 750 else { 751 assert(isa<SExtInst>(I) && "Unexpected ext type!"); 752 LType = ISD::SEXTLOAD; 753 } 754 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType()))) 755 return false; 756 757 // Move the extend into the same block as the load, so that SelectionDAG 758 // can fold it. 759 I->removeFromParent(); 760 I->insertAfter(LI); 761 return true; 762} 763 764bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 765 BasicBlock *DefBB = I->getParent(); 766 767 // If both result of the {s|z}xt and its source are live out, rewrite all 768 // other uses of the source with result of extension. 769 Value *Src = I->getOperand(0); 770 if (Src->hasOneUse()) 771 return false; 772 773 // Only do this xform if truncating is free. 774 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 775 return false; 776 777 // Only safe to perform the optimization if the source is also defined in 778 // this block. 779 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 780 return false; 781 782 bool DefIsLiveOut = false; 783 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 784 UI != E; ++UI) { 785 Instruction *User = cast<Instruction>(*UI); 786 787 // Figure out which BB this ext is used in. 788 BasicBlock *UserBB = User->getParent(); 789 if (UserBB == DefBB) continue; 790 DefIsLiveOut = true; 791 break; 792 } 793 if (!DefIsLiveOut) 794 return false; 795 796 // Make sure non of the uses are PHI nodes. 797 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 798 UI != E; ++UI) { 799 Instruction *User = cast<Instruction>(*UI); 800 BasicBlock *UserBB = User->getParent(); 801 if (UserBB == DefBB) continue; 802 // Be conservative. We don't want this xform to end up introducing 803 // reloads just before load / store instructions. 804 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 805 return false; 806 } 807 808 // InsertedTruncs - Only insert one trunc in each block once. 809 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 810 811 bool MadeChange = false; 812 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 813 UI != E; ++UI) { 814 Use &TheUse = UI.getUse(); 815 Instruction *User = cast<Instruction>(*UI); 816 817 // Figure out which BB this ext is used in. 818 BasicBlock *UserBB = User->getParent(); 819 if (UserBB == DefBB) continue; 820 821 // Both src and def are live in this block. Rewrite the use. 822 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 823 824 if (!InsertedTrunc) { 825 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 826 827 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 828 } 829 830 // Replace a use of the {s|z}ext source with a use of the result. 831 TheUse = InsertedTrunc; 832 833 MadeChange = true; 834 } 835 836 return MadeChange; 837} 838 839// In this pass we look for GEP and cast instructions that are used 840// across basic blocks and rewrite them to improve basic-block-at-a-time 841// selection. 842bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 843 bool MadeChange = false; 844 845 // Split all critical edges where the dest block has a PHI. 846 TerminatorInst *BBTI = BB.getTerminator(); 847 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) { 848 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) { 849 BasicBlock *SuccBB = BBTI->getSuccessor(i); 850 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true)) 851 SplitEdgeNicely(BBTI, i, BackEdges, this); 852 } 853 } 854 855 // Keep track of non-local addresses that have been sunk into this block. 856 // This allows us to avoid inserting duplicate code for blocks with multiple 857 // load/stores of the same address. 858 DenseMap<Value*, Value*> SunkAddrs; 859 860 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 861 Instruction *I = BBI++; 862 863 if (CastInst *CI = dyn_cast<CastInst>(I)) { 864 // If the source of the cast is a constant, then this should have 865 // already been constant folded. The only reason NOT to constant fold 866 // it is if something (e.g. LSR) was careful to place the constant 867 // evaluation in a block other than then one that uses it (e.g. to hoist 868 // the address of globals out of a loop). If this is the case, we don't 869 // want to forward-subst the cast. 870 if (isa<Constant>(CI->getOperand(0))) 871 continue; 872 873 bool Change = false; 874 if (TLI) { 875 Change = OptimizeNoopCopyExpression(CI, *TLI); 876 MadeChange |= Change; 877 } 878 879 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) { 880 MadeChange |= MoveExtToFormExtLoad(I); 881 MadeChange |= OptimizeExtUses(I); 882 } 883 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 884 MadeChange |= OptimizeCmpExpression(CI); 885 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 886 if (TLI) 887 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 888 SunkAddrs); 889 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 890 if (TLI) 891 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 892 SI->getOperand(0)->getType(), 893 SunkAddrs); 894 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 895 if (GEPI->hasAllZeroIndices()) { 896 /// The GEP operand must be a pointer, so must its result -> BitCast 897 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 898 GEPI->getName(), GEPI); 899 GEPI->replaceAllUsesWith(NC); 900 GEPI->eraseFromParent(); 901 MadeChange = true; 902 BBI = NC; 903 } 904 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 905 // If we found an inline asm expession, and if the target knows how to 906 // lower it to normal LLVM code, do so now. 907 if (TLI && isa<InlineAsm>(CI->getCalledValue())) { 908 if (TLI->ExpandInlineAsm(CI)) { 909 BBI = BB.begin(); 910 // Avoid processing instructions out of order, which could cause 911 // reuse before a value is defined. 912 SunkAddrs.clear(); 913 } else 914 // Sink address computing for memory operands into the block. 915 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 916 } 917 } 918 } 919 920 return MadeChange; 921} 922