1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===// 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 file implements some loop unrolling utilities. It does not define any 11 // actual pass or policy, but provides a single function to perform loop 12 // unrolling. 13 // 14 // The process of unrolling can produce extraneous basic blocks linked with 15 // unconditional branches. This will be corrected in the future. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #define DEBUG_TYPE "loop-unroll" 20 #include "llvm/Transforms/Utils/UnrollLoop.h" 21 #include "llvm/BasicBlock.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/InstructionSimplify.h" 24 #include "llvm/Analysis/LoopIterator.h" 25 #include "llvm/Analysis/LoopPass.h" 26 #include "llvm/Analysis/ScalarEvolution.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/raw_ostream.h" 29 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 30 #include "llvm/Transforms/Utils/Cloning.h" 31 #include "llvm/Transforms/Utils/Local.h" 32 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 33 using namespace llvm; 34 35 // TODO: Should these be here or in LoopUnroll? 36 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled"); 37 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)"); 38 39 /// RemapInstruction - Convert the instruction operands from referencing the 40 /// current values into those specified by VMap. 41 static inline void RemapInstruction(Instruction *I, 42 ValueToValueMapTy &VMap) { 43 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 44 Value *Op = I->getOperand(op); 45 ValueToValueMapTy::iterator It = VMap.find(Op); 46 if (It != VMap.end()) 47 I->setOperand(op, It->second); 48 } 49 50 if (PHINode *PN = dyn_cast<PHINode>(I)) { 51 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 52 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i)); 53 if (It != VMap.end()) 54 PN->setIncomingBlock(i, cast<BasicBlock>(It->second)); 55 } 56 } 57 } 58 59 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it 60 /// only has one predecessor, and that predecessor only has one successor. 61 /// The LoopInfo Analysis that is passed will be kept consistent. 62 /// Returns the new combined block. 63 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI, 64 LPPassManager *LPM) { 65 // Merge basic blocks into their predecessor if there is only one distinct 66 // pred, and if there is only one distinct successor of the predecessor, and 67 // if there are no PHI nodes. 68 BasicBlock *OnlyPred = BB->getSinglePredecessor(); 69 if (!OnlyPred) return 0; 70 71 if (OnlyPred->getTerminator()->getNumSuccessors() != 1) 72 return 0; 73 74 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred); 75 76 // Resolve any PHI nodes at the start of the block. They are all 77 // guaranteed to have exactly one entry if they exist, unless there are 78 // multiple duplicate (but guaranteed to be equal) entries for the 79 // incoming edges. This occurs when there are multiple edges from 80 // OnlyPred to OnlySucc. 81 FoldSingleEntryPHINodes(BB); 82 83 // Delete the unconditional branch from the predecessor... 84 OnlyPred->getInstList().pop_back(); 85 86 // Make all PHI nodes that referred to BB now refer to Pred as their 87 // source... 88 BB->replaceAllUsesWith(OnlyPred); 89 90 // Move all definitions in the successor to the predecessor... 91 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList()); 92 93 std::string OldName = BB->getName(); 94 95 // Erase basic block from the function... 96 97 // ScalarEvolution holds references to loop exit blocks. 98 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) { 99 if (Loop *L = LI->getLoopFor(BB)) 100 SE->forgetLoop(L); 101 } 102 LI->removeBlock(BB); 103 BB->eraseFromParent(); 104 105 // Inherit predecessor's name if it exists... 106 if (!OldName.empty() && !OnlyPred->hasName()) 107 OnlyPred->setName(OldName); 108 109 return OnlyPred; 110 } 111 112 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true 113 /// if unrolling was successful, or false if the loop was unmodified. Unrolling 114 /// can only fail when the loop's latch block is not terminated by a conditional 115 /// branch instruction. However, if the trip count (and multiple) are not known, 116 /// loop unrolling will mostly produce more code that is no faster. 117 /// 118 /// TripCount is generally defined as the number of times the loop header 119 /// executes. UnrollLoop relaxes the definition to permit early exits: here 120 /// TripCount is the iteration on which control exits LatchBlock if no early 121 /// exits were taken. Note that UnrollLoop assumes that the loop counter test 122 /// terminates LatchBlock in order to remove unnecesssary instances of the 123 /// test. In other words, control may exit the loop prior to TripCount 124 /// iterations via an early branch, but control may not exit the loop from the 125 /// LatchBlock's terminator prior to TripCount iterations. 126 /// 127 /// Similarly, TripMultiple divides the number of times that the LatchBlock may 128 /// execute without exiting the loop. 129 /// 130 /// The LoopInfo Analysis that is passed will be kept consistent. 131 /// 132 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be 133 /// removed from the LoopPassManager as well. LPM can also be NULL. 134 /// 135 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are 136 /// available it must also preserve those analyses. 137 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, 138 unsigned TripMultiple, LoopInfo *LI, LPPassManager *LPM) { 139 BasicBlock *Preheader = L->getLoopPreheader(); 140 if (!Preheader) { 141 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n"); 142 return false; 143 } 144 145 BasicBlock *LatchBlock = L->getLoopLatch(); 146 if (!LatchBlock) { 147 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n"); 148 return false; 149 } 150 151 BasicBlock *Header = L->getHeader(); 152 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator()); 153 154 if (!BI || BI->isUnconditional()) { 155 // The loop-rotate pass can be helpful to avoid this in many cases. 156 DEBUG(dbgs() << 157 " Can't unroll; loop not terminated by a conditional branch.\n"); 158 return false; 159 } 160 161 if (Header->hasAddressTaken()) { 162 // The loop-rotate pass can be helpful to avoid this in many cases. 163 DEBUG(dbgs() << 164 " Won't unroll loop: address of header block is taken.\n"); 165 return false; 166 } 167 168 // Notify ScalarEvolution that the loop will be substantially changed, 169 // if not outright eliminated. 170 ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>(); 171 if (SE) 172 SE->forgetLoop(L); 173 174 if (TripCount != 0) 175 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n"); 176 if (TripMultiple != 1) 177 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n"); 178 179 // Effectively "DCE" unrolled iterations that are beyond the tripcount 180 // and will never be executed. 181 if (TripCount != 0 && Count > TripCount) 182 Count = TripCount; 183 184 assert(Count > 0); 185 assert(TripMultiple > 0); 186 assert(TripCount == 0 || TripCount % TripMultiple == 0); 187 188 // Are we eliminating the loop control altogether? 189 bool CompletelyUnroll = Count == TripCount; 190 191 // If we know the trip count, we know the multiple... 192 unsigned BreakoutTrip = 0; 193 if (TripCount != 0) { 194 BreakoutTrip = TripCount % Count; 195 TripMultiple = 0; 196 } else { 197 // Figure out what multiple to use. 198 BreakoutTrip = TripMultiple = 199 (unsigned)GreatestCommonDivisor64(Count, TripMultiple); 200 } 201 202 if (CompletelyUnroll) { 203 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName() 204 << " with trip count " << TripCount << "!\n"); 205 } else { 206 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName() 207 << " by " << Count); 208 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) { 209 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip); 210 } else if (TripMultiple != 1) { 211 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch"); 212 } 213 DEBUG(dbgs() << "!\n"); 214 } 215 216 std::vector<BasicBlock*> LoopBlocks = L->getBlocks(); 217 218 bool ContinueOnTrue = L->contains(BI->getSuccessor(0)); 219 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue); 220 221 // For the first iteration of the loop, we should use the precloned values for 222 // PHI nodes. Insert associations now. 223 ValueToValueMapTy LastValueMap; 224 std::vector<PHINode*> OrigPHINode; 225 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 226 OrigPHINode.push_back(cast<PHINode>(I)); 227 } 228 229 std::vector<BasicBlock*> Headers; 230 std::vector<BasicBlock*> Latches; 231 Headers.push_back(Header); 232 Latches.push_back(LatchBlock); 233 234 // The current on-the-fly SSA update requires blocks to be processed in 235 // reverse postorder so that LastValueMap contains the correct value at each 236 // exit. 237 LoopBlocksDFS DFS(L); 238 DFS.perform(LI); 239 240 // Stash the DFS iterators before adding blocks to the loop. 241 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO(); 242 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO(); 243 244 for (unsigned It = 1; It != Count; ++It) { 245 std::vector<BasicBlock*> NewBlocks; 246 247 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) { 248 ValueToValueMapTy VMap; 249 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It)); 250 Header->getParent()->getBasicBlockList().push_back(New); 251 252 // Loop over all of the PHI nodes in the block, changing them to use the 253 // incoming values from the previous block. 254 if (*BB == Header) 255 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) { 256 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]); 257 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock); 258 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) 259 if (It > 1 && L->contains(InValI)) 260 InVal = LastValueMap[InValI]; 261 VMap[OrigPHINode[i]] = InVal; 262 New->getInstList().erase(NewPHI); 263 } 264 265 // Update our running map of newest clones 266 LastValueMap[*BB] = New; 267 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end(); 268 VI != VE; ++VI) 269 LastValueMap[VI->first] = VI->second; 270 271 L->addBasicBlockToLoop(New, LI->getBase()); 272 273 // Add phi entries for newly created values to all exit blocks. 274 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB); 275 SI != SE; ++SI) { 276 if (L->contains(*SI)) 277 continue; 278 for (BasicBlock::iterator BBI = (*SI)->begin(); 279 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) { 280 Value *Incoming = phi->getIncomingValueForBlock(*BB); 281 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming); 282 if (It != LastValueMap.end()) 283 Incoming = It->second; 284 phi->addIncoming(Incoming, New); 285 } 286 } 287 // Keep track of new headers and latches as we create them, so that 288 // we can insert the proper branches later. 289 if (*BB == Header) 290 Headers.push_back(New); 291 if (*BB == LatchBlock) 292 Latches.push_back(New); 293 294 NewBlocks.push_back(New); 295 } 296 297 // Remap all instructions in the most recent iteration 298 for (unsigned i = 0; i < NewBlocks.size(); ++i) 299 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 300 E = NewBlocks[i]->end(); I != E; ++I) 301 ::RemapInstruction(I, LastValueMap); 302 } 303 304 // Loop over the PHI nodes in the original block, setting incoming values. 305 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) { 306 PHINode *PN = OrigPHINode[i]; 307 if (CompletelyUnroll) { 308 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader)); 309 Header->getInstList().erase(PN); 310 } 311 else if (Count > 1) { 312 Value *InVal = PN->removeIncomingValue(LatchBlock, false); 313 // If this value was defined in the loop, take the value defined by the 314 // last iteration of the loop. 315 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) { 316 if (L->contains(InValI)) 317 InVal = LastValueMap[InVal]; 318 } 319 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch"); 320 PN->addIncoming(InVal, Latches.back()); 321 } 322 } 323 324 // Now that all the basic blocks for the unrolled iterations are in place, 325 // set up the branches to connect them. 326 for (unsigned i = 0, e = Latches.size(); i != e; ++i) { 327 // The original branch was replicated in each unrolled iteration. 328 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator()); 329 330 // The branch destination. 331 unsigned j = (i + 1) % e; 332 BasicBlock *Dest = Headers[j]; 333 bool NeedConditional = true; 334 335 // For a complete unroll, make the last iteration end with a branch 336 // to the exit block. 337 if (CompletelyUnroll && j == 0) { 338 Dest = LoopExit; 339 NeedConditional = false; 340 } 341 342 // If we know the trip count or a multiple of it, we can safely use an 343 // unconditional branch for some iterations. 344 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) { 345 NeedConditional = false; 346 } 347 348 if (NeedConditional) { 349 // Update the conditional branch's successor for the following 350 // iteration. 351 Term->setSuccessor(!ContinueOnTrue, Dest); 352 } else { 353 // Remove phi operands at this loop exit 354 if (Dest != LoopExit) { 355 BasicBlock *BB = Latches[i]; 356 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 357 SI != SE; ++SI) { 358 if (*SI == Headers[i]) 359 continue; 360 for (BasicBlock::iterator BBI = (*SI)->begin(); 361 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) { 362 Phi->removeIncomingValue(BB, false); 363 } 364 } 365 } 366 // Replace the conditional branch with an unconditional one. 367 BranchInst::Create(Dest, Term); 368 Term->eraseFromParent(); 369 } 370 } 371 372 // Merge adjacent basic blocks, if possible. 373 for (unsigned i = 0, e = Latches.size(); i != e; ++i) { 374 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator()); 375 if (Term->isUnconditional()) { 376 BasicBlock *Dest = Term->getSuccessor(0); 377 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM)) 378 std::replace(Latches.begin(), Latches.end(), Dest, Fold); 379 } 380 } 381 382 // FIXME: Reconstruct dom info, because it is not preserved properly. 383 // Incrementally updating domtree after loop unrolling would be easy. 384 if (DominatorTree *DT = LPM->getAnalysisIfAvailable<DominatorTree>()) 385 DT->runOnFunction(*L->getHeader()->getParent()); 386 387 // Simplify any new induction variables in the partially unrolled loop. 388 if (SE && !CompletelyUnroll) { 389 SmallVector<WeakVH, 16> DeadInsts; 390 simplifyLoopIVs(L, SE, LPM, DeadInsts); 391 392 // Aggressively clean up dead instructions that simplifyLoopIVs already 393 // identified. Any remaining should be cleaned up below. 394 while (!DeadInsts.empty()) 395 if (Instruction *Inst = 396 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 397 RecursivelyDeleteTriviallyDeadInstructions(Inst); 398 } 399 400 // At this point, the code is well formed. We now do a quick sweep over the 401 // inserted code, doing constant propagation and dead code elimination as we 402 // go. 403 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks(); 404 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(), 405 BBE = NewLoopBlocks.end(); BB != BBE; ++BB) 406 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) { 407 Instruction *Inst = I++; 408 409 if (isInstructionTriviallyDead(Inst)) 410 (*BB)->getInstList().erase(Inst); 411 else if (Value *V = SimplifyInstruction(Inst)) 412 if (LI->replacementPreservesLCSSAForm(Inst, V)) { 413 Inst->replaceAllUsesWith(V); 414 (*BB)->getInstList().erase(Inst); 415 } 416 } 417 418 NumCompletelyUnrolled += CompletelyUnroll; 419 ++NumUnrolled; 420 // Remove the loop from the LoopPassManager if it's completely removed. 421 if (CompletelyUnroll && LPM != NULL) 422 LPM->deleteLoopFromQueue(L); 423 424 return true; 425 } 426