LoopIdiomRecognize.cpp revision e2c43920919c6fe376613d1d8331897dc1ba3d57
1//===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===// 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 implements an idiom recognizer that transforms simple loops into a 11// non-loop form. In cases that this kicks in, it can be a significant 12// performance win. 13// 14//===----------------------------------------------------------------------===// 15 16#define DEBUG_TYPE "loop-idiom" 17#include "llvm/Transforms/Scalar.h" 18#include "llvm/Analysis/AliasAnalysis.h" 19#include "llvm/Analysis/LoopPass.h" 20#include "llvm/Analysis/ScalarEvolutionExpressions.h" 21#include "llvm/Analysis/ScalarEvolutionExpander.h" 22#include "llvm/Analysis/ValueTracking.h" 23#include "llvm/Target/TargetData.h" 24#include "llvm/Transforms/Utils/Local.h" 25#include "llvm/Support/Debug.h" 26#include "llvm/Support/IRBuilder.h" 27#include "llvm/Support/raw_ostream.h" 28using namespace llvm; 29 30// TODO: Recognize "N" size array multiplies: replace with call to blas or 31// something. 32 33namespace { 34 class LoopIdiomRecognize : public LoopPass { 35 Loop *CurLoop; 36 const TargetData *TD; 37 ScalarEvolution *SE; 38 public: 39 static char ID; 40 explicit LoopIdiomRecognize() : LoopPass(ID) { 41 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry()); 42 } 43 44 bool runOnLoop(Loop *L, LPPassManager &LPM); 45 46 bool processLoopStore(StoreInst *SI, const SCEV *BECount); 47 48 bool processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize, 49 Value *SplatValue, 50 const SCEVAddRecExpr *Ev, 51 const SCEV *BECount); 52 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, 53 const SCEVAddRecExpr *StoreEv, 54 const SCEVAddRecExpr *LoadEv, 55 const SCEV *BECount); 56 57 /// This transformation requires natural loop information & requires that 58 /// loop preheaders be inserted into the CFG. 59 /// 60 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 61 AU.addRequired<LoopInfo>(); 62 AU.addPreserved<LoopInfo>(); 63 AU.addRequiredID(LoopSimplifyID); 64 AU.addPreservedID(LoopSimplifyID); 65 AU.addRequiredID(LCSSAID); 66 AU.addPreservedID(LCSSAID); 67 AU.addRequired<AliasAnalysis>(); 68 AU.addPreserved<AliasAnalysis>(); 69 AU.addRequired<ScalarEvolution>(); 70 AU.addPreserved<ScalarEvolution>(); 71 AU.addPreserved<DominatorTree>(); 72 } 73 }; 74} 75 76char LoopIdiomRecognize::ID = 0; 77INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", 78 false, false) 79INITIALIZE_PASS_DEPENDENCY(LoopInfo) 80INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 81INITIALIZE_PASS_DEPENDENCY(LCSSA) 82INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 83INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 84INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", 85 false, false) 86 87Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); } 88 89/// DeleteDeadInstruction - Delete this instruction. Before we do, go through 90/// and zero out all the operands of this instruction. If any of them become 91/// dead, delete them and the computation tree that feeds them. 92/// 93static void DeleteDeadInstruction(Instruction *I, ScalarEvolution &SE) { 94 SmallVector<Instruction*, 32> NowDeadInsts; 95 96 NowDeadInsts.push_back(I); 97 98 // Before we touch this instruction, remove it from SE! 99 do { 100 Instruction *DeadInst = NowDeadInsts.pop_back_val(); 101 102 // This instruction is dead, zap it, in stages. Start by removing it from 103 // SCEV. 104 SE.forgetValue(DeadInst); 105 106 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) { 107 Value *Op = DeadInst->getOperand(op); 108 DeadInst->setOperand(op, 0); 109 110 // If this operand just became dead, add it to the NowDeadInsts list. 111 if (!Op->use_empty()) continue; 112 113 if (Instruction *OpI = dyn_cast<Instruction>(Op)) 114 if (isInstructionTriviallyDead(OpI)) 115 NowDeadInsts.push_back(OpI); 116 } 117 118 DeadInst->eraseFromParent(); 119 120 } while (!NowDeadInsts.empty()); 121} 122 123bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { 124 CurLoop = L; 125 126 // We only look at trivial single basic block loops. 127 // TODO: eventually support more complex loops, scanning the header. 128 if (L->getBlocks().size() != 1) 129 return false; 130 131 // The trip count of the loop must be analyzable. 132 SE = &getAnalysis<ScalarEvolution>(); 133 if (!SE->hasLoopInvariantBackedgeTakenCount(L)) 134 return false; 135 const SCEV *BECount = SE->getBackedgeTakenCount(L); 136 if (isa<SCEVCouldNotCompute>(BECount)) return false; 137 138 // We require target data for now. 139 TD = getAnalysisIfAvailable<TargetData>(); 140 if (TD == 0) return false; 141 142 BasicBlock *BB = L->getHeader(); 143 DEBUG(dbgs() << "loop-idiom Scanning: F[" << BB->getParent()->getName() 144 << "] Loop %" << BB->getName() << "\n"); 145 146 bool MadeChange = false; 147 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { 148 // Look for store instructions, which may be memsets. 149 StoreInst *SI = dyn_cast<StoreInst>(I++); 150 if (SI == 0 || SI->isVolatile()) continue; 151 152 WeakVH InstPtr(SI); 153 if (!processLoopStore(SI, BECount)) continue; 154 155 MadeChange = true; 156 157 // If processing the store invalidated our iterator, start over from the 158 // head of the loop. 159 if (InstPtr == 0) 160 I = BB->begin(); 161 } 162 163 return MadeChange; 164} 165 166/// scanBlock - Look over a block to see if we can promote anything out of it. 167bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) { 168 Value *StoredVal = SI->getValueOperand(); 169 Value *StorePtr = SI->getPointerOperand(); 170 171 // Reject stores that are so large that they overflow an unsigned. 172 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType()); 173 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) 174 return false; 175 176 // See if the pointer expression is an AddRec like {base,+,1} on the current 177 // loop, which indicates a strided store. If we have something else, it's a 178 // random store we can't handle. 179 const SCEVAddRecExpr *StoreEv = 180 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr)); 181 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) 182 return false; 183 184 // Check to see if the stride matches the size of the store. If so, then we 185 // know that every byte is touched in the loop. 186 unsigned StoreSize = (unsigned)SizeInBits >> 3; 187 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1)); 188 189 // TODO: Could also handle negative stride here someday, that will require the 190 // validity check in mayLoopModRefLocation to be updated though. 191 if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) 192 return false; 193 194 // If the stored value is a byte-wise value (like i32 -1), then it may be 195 // turned into a memset of i8 -1, assuming that all the consequtive bytes 196 // are stored. A store of i32 0x01020304 can never be turned into a memset. 197 if (Value *SplatValue = isBytewiseValue(StoredVal)) 198 if (processLoopStoreOfSplatValue(SI, StoreSize, SplatValue, StoreEv, 199 BECount)) 200 return true; 201 202 // If the stored value is a strided load in the same loop with the same stride 203 // this this may be transformable into a memcpy. This kicks in for stuff like 204 // for (i) A[i] = B[i]; 205 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) { 206 const SCEVAddRecExpr *LoadEv = 207 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0))); 208 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() && 209 StoreEv->getOperand(1) == LoadEv->getOperand(1) && !LI->isVolatile()) 210 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount)) 211 return true; 212 } 213 // errs() << "UNHANDLED strided store: " << *Ev << " - " << *SI << "\n"; 214 215 return false; 216} 217 218/// mayLoopModRefLocation - Return true if the specified loop might do a load or 219/// store to the same location that the specified store could store to, which is 220/// a loop-strided access. 221static bool mayLoopModRefLocation(Value *Ptr, Loop *L, const SCEV *BECount, 222 unsigned StoreSize, AliasAnalysis &AA, 223 StoreInst *IgnoredStore) { 224 // Get the location that may be stored across the loop. Since the access is 225 // strided positively through memory, we say that the modified location starts 226 // at the pointer and has infinite size. 227 uint64_t AccessSize = AliasAnalysis::UnknownSize; 228 229 // If the loop iterates a fixed number of times, we can refine the access size 230 // to be exactly the size of the memset, which is (BECount+1)*StoreSize 231 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount)) 232 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize; 233 234 // TODO: For this to be really effective, we have to dive into the pointer 235 // operand in the store. Store to &A[i] of 100 will always return may alias 236 // with store of &A[100], we need to StoreLoc to be "A" with size of 100, 237 // which will then no-alias a store to &A[100]. 238 AliasAnalysis::Location StoreLoc(Ptr, AccessSize); 239 240 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; 241 ++BI) 242 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I) 243 if (&*I != IgnoredStore && 244 AA.getModRefInfo(I, StoreLoc) != AliasAnalysis::NoModRef) 245 return true; 246 247 return false; 248} 249 250/// processLoopStoreOfSplatValue - We see a strided store of a memsetable value. 251/// If we can transform this into a memset in the loop preheader, do so. 252bool LoopIdiomRecognize:: 253processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize, 254 Value *SplatValue, 255 const SCEVAddRecExpr *Ev, const SCEV *BECount) { 256 // Verify that the stored value is loop invariant. If not, we can't promote 257 // the memset. 258 if (!CurLoop->isLoopInvariant(SplatValue)) 259 return false; 260 261 // Okay, we have a strided store "p[i]" of a splattable value. We can turn 262 // this into a memset in the loop preheader now if we want. However, this 263 // would be unsafe to do if there is anything else in the loop that may read 264 // or write to the aliased location. Check for an alias. 265 if (mayLoopModRefLocation(SI->getPointerOperand(), CurLoop, BECount, 266 StoreSize, getAnalysis<AliasAnalysis>(), SI)) 267 return false; 268 269 // Okay, everything looks good, insert the memset. 270 BasicBlock *Preheader = CurLoop->getLoopPreheader(); 271 272 IRBuilder<> Builder(Preheader->getTerminator()); 273 274 // The trip count of the loop and the base pointer of the addrec SCEV is 275 // guaranteed to be loop invariant, which means that it should dominate the 276 // header. Just insert code for it in the preheader. 277 SCEVExpander Expander(*SE); 278 279 unsigned AddrSpace = SI->getPointerAddressSpace(); 280 Value *BasePtr = 281 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace), 282 Preheader->getTerminator()); 283 284 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to 285 // pointer size if it isn't already. 286 const Type *IntPtr = TD->getIntPtrType(SI->getContext()); 287 unsigned BESize = SE->getTypeSizeInBits(BECount->getType()); 288 if (BESize < TD->getPointerSizeInBits()) 289 BECount = SE->getZeroExtendExpr(BECount, IntPtr); 290 else if (BESize > TD->getPointerSizeInBits()) 291 BECount = SE->getTruncateExpr(BECount, IntPtr); 292 293 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), 294 true, true /*nooverflow*/); 295 if (StoreSize != 1) 296 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), 297 true, true /*nooverflow*/); 298 299 Value *NumBytes = 300 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); 301 302 Value *NewCall = 303 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, SI->getAlignment()); 304 305 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" 306 << " from store to: " << *Ev << " at: " << *SI << "\n"); 307 (void)NewCall; 308 309 // Okay, the memset has been formed. Zap the original store and anything that 310 // feeds into it. 311 DeleteDeadInstruction(SI, *SE); 312 return true; 313} 314 315/// processLoopStoreOfLoopLoad - We see a strided store whose value is a 316/// same-strided load. 317bool LoopIdiomRecognize:: 318processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, 319 const SCEVAddRecExpr *StoreEv, 320 const SCEVAddRecExpr *LoadEv, 321 const SCEV *BECount) { 322 LoadInst *LI = cast<LoadInst>(SI->getValueOperand()); 323 324 // Okay, we have a strided store "p[i]" of a loaded value. We can turn 325 // this into a memcmp in the loop preheader now if we want. However, this 326 // would be unsafe to do if there is anything else in the loop that may read 327 // or write to the aliased location (including the load feeding the stores). 328 // Check for an alias. 329 if (mayLoopModRefLocation(SI->getPointerOperand(), CurLoop, BECount, 330 StoreSize, getAnalysis<AliasAnalysis>(), SI)) 331 return false; 332 333 // Okay, everything looks good, insert the memcpy. 334 BasicBlock *Preheader = CurLoop->getLoopPreheader(); 335 336 IRBuilder<> Builder(Preheader->getTerminator()); 337 338 // The trip count of the loop and the base pointer of the addrec SCEV is 339 // guaranteed to be loop invariant, which means that it should dominate the 340 // header. Just insert code for it in the preheader. 341 SCEVExpander Expander(*SE); 342 343 Value *LoadBasePtr = 344 Expander.expandCodeFor(LoadEv->getStart(), 345 Builder.getInt8PtrTy(LI->getPointerAddressSpace()), 346 Preheader->getTerminator()); 347 Value *StoreBasePtr = 348 Expander.expandCodeFor(StoreEv->getStart(), 349 Builder.getInt8PtrTy(SI->getPointerAddressSpace()), 350 Preheader->getTerminator()); 351 352 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to 353 // pointer size if it isn't already. 354 const Type *IntPtr = TD->getIntPtrType(SI->getContext()); 355 unsigned BESize = SE->getTypeSizeInBits(BECount->getType()); 356 if (BESize < TD->getPointerSizeInBits()) 357 BECount = SE->getZeroExtendExpr(BECount, IntPtr); 358 else if (BESize > TD->getPointerSizeInBits()) 359 BECount = SE->getTruncateExpr(BECount, IntPtr); 360 361 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), 362 true, true /*nooverflow*/); 363 if (StoreSize != 1) 364 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), 365 true, true /*nooverflow*/); 366 367 Value *NumBytes = 368 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); 369 370 Value *NewCall = 371 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, 372 std::min(SI->getAlignment(), LI->getAlignment())); 373 374 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n" 375 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" 376 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"); 377 (void)NewCall; 378 379 // Okay, the memset has been formed. Zap the original store and anything that 380 // feeds into it. 381 DeleteDeadInstruction(SI, *SE); 382 return true; 383} 384