1 //===- LoopRotation.cpp - Loop Rotation Pass ------------------------------===// 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 Loop Rotation Pass. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "loop-rotate" 15 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/Function.h" 17 #include "llvm/IntrinsicInst.h" 18 #include "llvm/Analysis/CodeMetrics.h" 19 #include "llvm/Analysis/LoopPass.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/ScalarEvolution.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/Transforms/Utils/Local.h" 24 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 25 #include "llvm/Transforms/Utils/SSAUpdater.h" 26 #include "llvm/Transforms/Utils/ValueMapper.h" 27 #include "llvm/Support/CFG.h" 28 #include "llvm/Support/Debug.h" 29 #include "llvm/ADT/Statistic.h" 30 using namespace llvm; 31 32 #define MAX_HEADER_SIZE 16 33 34 STATISTIC(NumRotated, "Number of loops rotated"); 35 namespace { 36 37 class LoopRotate : public LoopPass { 38 public: 39 static char ID; // Pass ID, replacement for typeid 40 LoopRotate() : LoopPass(ID) { 41 initializeLoopRotatePass(*PassRegistry::getPassRegistry()); 42 } 43 44 // LCSSA form makes instruction renaming easier. 45 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 46 AU.addPreserved<DominatorTree>(); 47 AU.addRequired<LoopInfo>(); 48 AU.addPreserved<LoopInfo>(); 49 AU.addRequiredID(LoopSimplifyID); 50 AU.addPreservedID(LoopSimplifyID); 51 AU.addRequiredID(LCSSAID); 52 AU.addPreservedID(LCSSAID); 53 AU.addPreserved<ScalarEvolution>(); 54 } 55 56 bool runOnLoop(Loop *L, LPPassManager &LPM); 57 void simplifyLoopLatch(Loop *L); 58 bool rotateLoop(Loop *L); 59 60 private: 61 LoopInfo *LI; 62 }; 63 } 64 65 char LoopRotate::ID = 0; 66 INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false) 67 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 68 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 69 INITIALIZE_PASS_DEPENDENCY(LCSSA) 70 INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false) 71 72 Pass *llvm::createLoopRotatePass() { return new LoopRotate(); } 73 74 /// Rotate Loop L as many times as possible. Return true if 75 /// the loop is rotated at least once. 76 bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) { 77 LI = &getAnalysis<LoopInfo>(); 78 79 // Simplify the loop latch before attempting to rotate the header 80 // upward. Rotation may not be needed if the loop tail can be folded into the 81 // loop exit. 82 simplifyLoopLatch(L); 83 84 // One loop can be rotated multiple times. 85 bool MadeChange = false; 86 while (rotateLoop(L)) 87 MadeChange = true; 88 89 return MadeChange; 90 } 91 92 /// RewriteUsesOfClonedInstructions - We just cloned the instructions from the 93 /// old header into the preheader. If there were uses of the values produced by 94 /// these instruction that were outside of the loop, we have to insert PHI nodes 95 /// to merge the two values. Do this now. 96 static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader, 97 BasicBlock *OrigPreheader, 98 ValueToValueMapTy &ValueMap) { 99 // Remove PHI node entries that are no longer live. 100 BasicBlock::iterator I, E = OrigHeader->end(); 101 for (I = OrigHeader->begin(); PHINode *PN = dyn_cast<PHINode>(I); ++I) 102 PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader)); 103 104 // Now fix up users of the instructions in OrigHeader, inserting PHI nodes 105 // as necessary. 106 SSAUpdater SSA; 107 for (I = OrigHeader->begin(); I != E; ++I) { 108 Value *OrigHeaderVal = I; 109 110 // If there are no uses of the value (e.g. because it returns void), there 111 // is nothing to rewrite. 112 if (OrigHeaderVal->use_empty()) 113 continue; 114 115 Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal]; 116 117 // The value now exits in two versions: the initial value in the preheader 118 // and the loop "next" value in the original header. 119 SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName()); 120 SSA.AddAvailableValue(OrigHeader, OrigHeaderVal); 121 SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal); 122 123 // Visit each use of the OrigHeader instruction. 124 for (Value::use_iterator UI = OrigHeaderVal->use_begin(), 125 UE = OrigHeaderVal->use_end(); UI != UE; ) { 126 // Grab the use before incrementing the iterator. 127 Use &U = UI.getUse(); 128 129 // Increment the iterator before removing the use from the list. 130 ++UI; 131 132 // SSAUpdater can't handle a non-PHI use in the same block as an 133 // earlier def. We can easily handle those cases manually. 134 Instruction *UserInst = cast<Instruction>(U.getUser()); 135 if (!isa<PHINode>(UserInst)) { 136 BasicBlock *UserBB = UserInst->getParent(); 137 138 // The original users in the OrigHeader are already using the 139 // original definitions. 140 if (UserBB == OrigHeader) 141 continue; 142 143 // Users in the OrigPreHeader need to use the value to which the 144 // original definitions are mapped. 145 if (UserBB == OrigPreheader) { 146 U = OrigPreHeaderVal; 147 continue; 148 } 149 } 150 151 // Anything else can be handled by SSAUpdater. 152 SSA.RewriteUse(U); 153 } 154 } 155 } 156 157 /// Determine whether the instructions in this range my be safely and cheaply 158 /// speculated. This is not an important enough situation to develop complex 159 /// heuristics. We handle a single arithmetic instruction along with any type 160 /// conversions. 161 static bool shouldSpeculateInstrs(BasicBlock::iterator Begin, 162 BasicBlock::iterator End) { 163 bool seenIncrement = false; 164 for (BasicBlock::iterator I = Begin; I != End; ++I) { 165 166 if (!isSafeToSpeculativelyExecute(I)) 167 return false; 168 169 if (isa<DbgInfoIntrinsic>(I)) 170 continue; 171 172 switch (I->getOpcode()) { 173 default: 174 return false; 175 case Instruction::GetElementPtr: 176 // GEPs are cheap if all indices are constant. 177 if (!cast<GEPOperator>(I)->hasAllConstantIndices()) 178 return false; 179 // fall-thru to increment case 180 case Instruction::Add: 181 case Instruction::Sub: 182 case Instruction::And: 183 case Instruction::Or: 184 case Instruction::Xor: 185 case Instruction::Shl: 186 case Instruction::LShr: 187 case Instruction::AShr: 188 if (seenIncrement) 189 return false; 190 seenIncrement = true; 191 break; 192 case Instruction::Trunc: 193 case Instruction::ZExt: 194 case Instruction::SExt: 195 // ignore type conversions 196 break; 197 } 198 } 199 return true; 200 } 201 202 /// Fold the loop tail into the loop exit by speculating the loop tail 203 /// instructions. Typically, this is a single post-increment. In the case of a 204 /// simple 2-block loop, hoisting the increment can be much better than 205 /// duplicating the entire loop header. In the cast of loops with early exits, 206 /// rotation will not work anyway, but simplifyLoopLatch will put the loop in 207 /// canonical form so downstream passes can handle it. 208 /// 209 /// I don't believe this invalidates SCEV. 210 void LoopRotate::simplifyLoopLatch(Loop *L) { 211 BasicBlock *Latch = L->getLoopLatch(); 212 if (!Latch || Latch->hasAddressTaken()) 213 return; 214 215 BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator()); 216 if (!Jmp || !Jmp->isUnconditional()) 217 return; 218 219 BasicBlock *LastExit = Latch->getSinglePredecessor(); 220 if (!LastExit || !L->isLoopExiting(LastExit)) 221 return; 222 223 BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator()); 224 if (!BI) 225 return; 226 227 if (!shouldSpeculateInstrs(Latch->begin(), Jmp)) 228 return; 229 230 DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into " 231 << LastExit->getName() << "\n"); 232 233 // Hoist the instructions from Latch into LastExit. 234 LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp); 235 236 unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1; 237 BasicBlock *Header = Jmp->getSuccessor(0); 238 assert(Header == L->getHeader() && "expected a backward branch"); 239 240 // Remove Latch from the CFG so that LastExit becomes the new Latch. 241 BI->setSuccessor(FallThruPath, Header); 242 Latch->replaceSuccessorsPhiUsesWith(LastExit); 243 Jmp->eraseFromParent(); 244 245 // Nuke the Latch block. 246 assert(Latch->empty() && "unable to evacuate Latch"); 247 LI->removeBlock(Latch); 248 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) 249 DT->eraseNode(Latch); 250 Latch->eraseFromParent(); 251 } 252 253 /// Rotate loop LP. Return true if the loop is rotated. 254 bool LoopRotate::rotateLoop(Loop *L) { 255 // If the loop has only one block then there is not much to rotate. 256 if (L->getBlocks().size() == 1) 257 return false; 258 259 BasicBlock *OrigHeader = L->getHeader(); 260 BasicBlock *OrigLatch = L->getLoopLatch(); 261 262 BranchInst *BI = dyn_cast<BranchInst>(OrigHeader->getTerminator()); 263 if (BI == 0 || BI->isUnconditional()) 264 return false; 265 266 // If the loop header is not one of the loop exiting blocks then 267 // either this loop is already rotated or it is not 268 // suitable for loop rotation transformations. 269 if (!L->isLoopExiting(OrigHeader)) 270 return false; 271 272 // If the loop latch already contains a branch that leaves the loop then the 273 // loop is already rotated. 274 if (OrigLatch == 0 || L->isLoopExiting(OrigLatch)) 275 return false; 276 277 // Check size of original header and reject loop if it is very big. 278 { 279 CodeMetrics Metrics; 280 Metrics.analyzeBasicBlock(OrigHeader); 281 if (Metrics.NumInsts > MAX_HEADER_SIZE) 282 return false; 283 } 284 285 // Now, this loop is suitable for rotation. 286 BasicBlock *OrigPreheader = L->getLoopPreheader(); 287 288 // If the loop could not be converted to canonical form, it must have an 289 // indirectbr in it, just give up. 290 if (OrigPreheader == 0) 291 return false; 292 293 // Anything ScalarEvolution may know about this loop or the PHI nodes 294 // in its header will soon be invalidated. 295 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>()) 296 SE->forgetLoop(L); 297 298 DEBUG(dbgs() << "LoopRotation: rotating "; L->dump()); 299 300 // Find new Loop header. NewHeader is a Header's one and only successor 301 // that is inside loop. Header's other successor is outside the 302 // loop. Otherwise loop is not suitable for rotation. 303 BasicBlock *Exit = BI->getSuccessor(0); 304 BasicBlock *NewHeader = BI->getSuccessor(1); 305 if (L->contains(Exit)) 306 std::swap(Exit, NewHeader); 307 assert(NewHeader && "Unable to determine new loop header"); 308 assert(L->contains(NewHeader) && !L->contains(Exit) && 309 "Unable to determine loop header and exit blocks"); 310 311 // This code assumes that the new header has exactly one predecessor. 312 // Remove any single-entry PHI nodes in it. 313 assert(NewHeader->getSinglePredecessor() && 314 "New header doesn't have one pred!"); 315 FoldSingleEntryPHINodes(NewHeader); 316 317 // Begin by walking OrigHeader and populating ValueMap with an entry for 318 // each Instruction. 319 BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end(); 320 ValueToValueMapTy ValueMap; 321 322 // For PHI nodes, the value available in OldPreHeader is just the 323 // incoming value from OldPreHeader. 324 for (; PHINode *PN = dyn_cast<PHINode>(I); ++I) 325 ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader); 326 327 // For the rest of the instructions, either hoist to the OrigPreheader if 328 // possible or create a clone in the OldPreHeader if not. 329 TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator(); 330 while (I != E) { 331 Instruction *Inst = I++; 332 333 // If the instruction's operands are invariant and it doesn't read or write 334 // memory, then it is safe to hoist. Doing this doesn't change the order of 335 // execution in the preheader, but does prevent the instruction from 336 // executing in each iteration of the loop. This means it is safe to hoist 337 // something that might trap, but isn't safe to hoist something that reads 338 // memory (without proving that the loop doesn't write). 339 if (L->hasLoopInvariantOperands(Inst) && 340 !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() && 341 !isa<TerminatorInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst) && 342 !isa<AllocaInst>(Inst)) { 343 Inst->moveBefore(LoopEntryBranch); 344 continue; 345 } 346 347 // Otherwise, create a duplicate of the instruction. 348 Instruction *C = Inst->clone(); 349 350 // Eagerly remap the operands of the instruction. 351 RemapInstruction(C, ValueMap, 352 RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); 353 354 // With the operands remapped, see if the instruction constant folds or is 355 // otherwise simplifyable. This commonly occurs because the entry from PHI 356 // nodes allows icmps and other instructions to fold. 357 Value *V = SimplifyInstruction(C); 358 if (V && LI->replacementPreservesLCSSAForm(C, V)) { 359 // If so, then delete the temporary instruction and stick the folded value 360 // in the map. 361 delete C; 362 ValueMap[Inst] = V; 363 } else { 364 // Otherwise, stick the new instruction into the new block! 365 C->setName(Inst->getName()); 366 C->insertBefore(LoopEntryBranch); 367 ValueMap[Inst] = C; 368 } 369 } 370 371 // Along with all the other instructions, we just cloned OrigHeader's 372 // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's 373 // successors by duplicating their incoming values for OrigHeader. 374 TerminatorInst *TI = OrigHeader->getTerminator(); 375 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 376 for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin(); 377 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 378 PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader); 379 380 // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove 381 // OrigPreHeader's old terminator (the original branch into the loop), and 382 // remove the corresponding incoming values from the PHI nodes in OrigHeader. 383 LoopEntryBranch->eraseFromParent(); 384 385 // If there were any uses of instructions in the duplicated block outside the 386 // loop, update them, inserting PHI nodes as required 387 RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap); 388 389 // NewHeader is now the header of the loop. 390 L->moveToHeader(NewHeader); 391 assert(L->getHeader() == NewHeader && "Latch block is our new header"); 392 393 394 // At this point, we've finished our major CFG changes. As part of cloning 395 // the loop into the preheader we've simplified instructions and the 396 // duplicated conditional branch may now be branching on a constant. If it is 397 // branching on a constant and if that constant means that we enter the loop, 398 // then we fold away the cond branch to an uncond branch. This simplifies the 399 // loop in cases important for nested loops, and it also means we don't have 400 // to split as many edges. 401 BranchInst *PHBI = cast<BranchInst>(OrigPreheader->getTerminator()); 402 assert(PHBI->isConditional() && "Should be clone of BI condbr!"); 403 if (!isa<ConstantInt>(PHBI->getCondition()) || 404 PHBI->getSuccessor(cast<ConstantInt>(PHBI->getCondition())->isZero()) 405 != NewHeader) { 406 // The conditional branch can't be folded, handle the general case. 407 // Update DominatorTree to reflect the CFG change we just made. Then split 408 // edges as necessary to preserve LoopSimplify form. 409 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) { 410 // Everything that was dominated by the old loop header is now dominated 411 // by the original loop preheader. Conceptually the header was merged 412 // into the preheader, even though we reuse the actual block as a new 413 // loop latch. 414 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); 415 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(), 416 OrigHeaderNode->end()); 417 DomTreeNode *OrigPreheaderNode = DT->getNode(OrigPreheader); 418 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) 419 DT->changeImmediateDominator(HeaderChildren[I], OrigPreheaderNode); 420 421 assert(DT->getNode(Exit)->getIDom() == OrigPreheaderNode); 422 assert(DT->getNode(NewHeader)->getIDom() == OrigPreheaderNode); 423 424 // Update OrigHeader to be dominated by the new header block. 425 DT->changeImmediateDominator(OrigHeader, OrigLatch); 426 } 427 428 // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and 429 // thus is not a preheader anymore. 430 // Split the edge to form a real preheader. 431 BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this); 432 NewPH->setName(NewHeader->getName() + ".lr.ph"); 433 434 // Preserve canonical loop form, which means that 'Exit' should have only 435 // one predecessor. 436 BasicBlock *ExitSplit = SplitCriticalEdge(L->getLoopLatch(), Exit, this); 437 ExitSplit->moveBefore(Exit); 438 } else { 439 // We can fold the conditional branch in the preheader, this makes things 440 // simpler. The first step is to remove the extra edge to the Exit block. 441 Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/); 442 BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI); 443 NewBI->setDebugLoc(PHBI->getDebugLoc()); 444 PHBI->eraseFromParent(); 445 446 // With our CFG finalized, update DomTree if it is available. 447 if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>()) { 448 // Update OrigHeader to be dominated by the new header block. 449 DT->changeImmediateDominator(NewHeader, OrigPreheader); 450 DT->changeImmediateDominator(OrigHeader, OrigLatch); 451 452 // Brute force incremental dominator tree update. Call 453 // findNearestCommonDominator on all CFG predecessors of each child of the 454 // original header. 455 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); 456 SmallVector<DomTreeNode *, 8> HeaderChildren(OrigHeaderNode->begin(), 457 OrigHeaderNode->end()); 458 bool Changed; 459 do { 460 Changed = false; 461 for (unsigned I = 0, E = HeaderChildren.size(); I != E; ++I) { 462 DomTreeNode *Node = HeaderChildren[I]; 463 BasicBlock *BB = Node->getBlock(); 464 465 pred_iterator PI = pred_begin(BB); 466 BasicBlock *NearestDom = *PI; 467 for (pred_iterator PE = pred_end(BB); PI != PE; ++PI) 468 NearestDom = DT->findNearestCommonDominator(NearestDom, *PI); 469 470 // Remember if this changes the DomTree. 471 if (Node->getIDom()->getBlock() != NearestDom) { 472 DT->changeImmediateDominator(BB, NearestDom); 473 Changed = true; 474 } 475 } 476 477 // If the dominator changed, this may have an effect on other 478 // predecessors, continue until we reach a fixpoint. 479 } while (Changed); 480 } 481 } 482 483 assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation"); 484 assert(L->getLoopLatch() && "Invalid loop latch after loop rotation"); 485 486 // Now that the CFG and DomTree are in a consistent state again, try to merge 487 // the OrigHeader block into OrigLatch. This will succeed if they are 488 // connected by an unconditional branch. This is just a cleanup so the 489 // emitted code isn't too gross in this common case. 490 MergeBlockIntoPredecessor(OrigHeader, this); 491 492 DEBUG(dbgs() << "LoopRotation: into "; L->dump()); 493 494 ++NumRotated; 495 return true; 496 } 497 498