LoopSimplify.cpp revision 2d0a91cd6c3df32014d547255d6a615bd1bc84fb
1//===- LoopSimplify.cpp - Loop Canonicalization 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 pass performs several transformations to transform natural loops into a 11// simpler form, which makes subsequent analyses and transformations simpler and 12// more effective. 13// 14// Loop pre-header insertion guarantees that there is a single, non-critical 15// entry edge from outside of the loop to the loop header. This simplifies a 16// number of analyses and transformations, such as LICM. 17// 18// Loop exit-block insertion guarantees that all exit blocks from the loop 19// (blocks which are outside of the loop that have predecessors inside of the 20// loop) only have predecessors from inside of the loop (and are thus dominated 21// by the loop header). This simplifies transformations such as store-sinking 22// that are built into LICM. 23// 24// This pass also guarantees that loops will have exactly one backedge. 25// 26// Indirectbr instructions introduce several complications. If the loop 27// contains or is entered by an indirectbr instruction, it may not be possible 28// to transform the loop and make these guarantees. Client code should check 29// that these conditions are true before relying on them. 30// 31// Note that the simplifycfg pass will clean up blocks which are split out but 32// end up being unnecessary, so usage of this pass should not pessimize 33// generated code. 34// 35// This pass obviously modifies the CFG, but updates loop information and 36// dominator information. 37// 38//===----------------------------------------------------------------------===// 39 40#define DEBUG_TYPE "loopsimplify" 41#include "llvm/Transforms/Scalar.h" 42#include "llvm/Constants.h" 43#include "llvm/Instructions.h" 44#include "llvm/Function.h" 45#include "llvm/LLVMContext.h" 46#include "llvm/Type.h" 47#include "llvm/Analysis/AliasAnalysis.h" 48#include "llvm/Analysis/Dominators.h" 49#include "llvm/Analysis/LoopPass.h" 50#include "llvm/Analysis/ScalarEvolution.h" 51#include "llvm/Transforms/Utils/BasicBlockUtils.h" 52#include "llvm/Transforms/Utils/Local.h" 53#include "llvm/Support/CFG.h" 54#include "llvm/Support/Debug.h" 55#include "llvm/ADT/SetOperations.h" 56#include "llvm/ADT/SetVector.h" 57#include "llvm/ADT/Statistic.h" 58#include "llvm/ADT/DepthFirstIterator.h" 59using namespace llvm; 60 61STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted"); 62STATISTIC(NumNested , "Number of nested loops split out"); 63 64namespace { 65 struct LoopSimplify : public LoopPass { 66 static char ID; // Pass identification, replacement for typeid 67 LoopSimplify() : LoopPass(&ID) {} 68 69 // AA - If we have an alias analysis object to update, this is it, otherwise 70 // this is null. 71 AliasAnalysis *AA; 72 LoopInfo *LI; 73 DominatorTree *DT; 74 Loop *L; 75 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 76 77 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 78 // We need loop information to identify the loops... 79 AU.addRequiredTransitive<LoopInfo>(); 80 AU.addRequiredTransitive<DominatorTree>(); 81 82 AU.addPreserved<LoopInfo>(); 83 AU.addPreserved<DominatorTree>(); 84 AU.addPreserved<DominanceFrontier>(); 85 AU.addPreserved<AliasAnalysis>(); 86 AU.addPreserved<ScalarEvolution>(); 87 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. 88 } 89 90 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees. 91 void verifyAnalysis() const; 92 93 private: 94 bool ProcessLoop(Loop *L, LPPassManager &LPM); 95 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit); 96 BasicBlock *InsertPreheaderForLoop(Loop *L); 97 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM); 98 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader); 99 void PlaceSplitBlockCarefully(BasicBlock *NewBB, 100 SmallVectorImpl<BasicBlock*> &SplitPreds, 101 Loop *L); 102 }; 103} 104 105char LoopSimplify::ID = 0; 106static RegisterPass<LoopSimplify> 107X("loopsimplify", "Canonicalize natural loops", true); 108 109// Publically exposed interface to pass... 110const PassInfo *const llvm::LoopSimplifyID = &X; 111Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } 112 113/// runOnLoop - Run down all loops in the CFG (recursively, but we could do 114/// it in any convenient order) inserting preheaders... 115/// 116bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) { 117 L = l; 118 bool Changed = false; 119 LI = &getAnalysis<LoopInfo>(); 120 AA = getAnalysisIfAvailable<AliasAnalysis>(); 121 DT = &getAnalysis<DominatorTree>(); 122 123 Changed |= ProcessLoop(L, LPM); 124 125 return Changed; 126} 127 128/// ProcessLoop - Walk the loop structure in depth first order, ensuring that 129/// all loops have preheaders. 130/// 131bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) { 132 bool Changed = false; 133ReprocessLoop: 134 135 // Check to see that no blocks (other than the header) in this loop have 136 // predecessors that are not in the loop. This is not valid for natural 137 // loops, but can occur if the blocks are unreachable. Since they are 138 // unreachable we can just shamelessly delete those CFG edges! 139 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 140 BB != E; ++BB) { 141 if (*BB == L->getHeader()) continue; 142 143 SmallPtrSet<BasicBlock *, 4> BadPreds; 144 for (pred_iterator PI = pred_begin(*BB), PE = pred_end(*BB); PI != PE; ++PI) 145 if (!L->contains(*PI)) 146 BadPreds.insert(*PI); 147 148 // Delete each unique out-of-loop (and thus dead) predecessor. 149 for (SmallPtrSet<BasicBlock *, 4>::iterator I = BadPreds.begin(), 150 E = BadPreds.end(); I != E; ++I) { 151 152 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "; 153 WriteAsOperand(dbgs(), *I, false); 154 dbgs() << "\n"); 155 156 // Inform each successor of each dead pred. 157 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 158 (*SI)->removePredecessor(*I); 159 // Zap the dead pred's terminator and replace it with unreachable. 160 TerminatorInst *TI = (*I)->getTerminator(); 161 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 162 (*I)->getTerminator()->eraseFromParent(); 163 new UnreachableInst((*I)->getContext(), *I); 164 Changed = true; 165 } 166 } 167 168 // If there are exiting blocks with branches on undef, resolve the undef in 169 // the direction which will exit the loop. This will help simplify loop 170 // trip count computations. 171 SmallVector<BasicBlock*, 8> ExitingBlocks; 172 L->getExitingBlocks(ExitingBlocks); 173 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(), 174 E = ExitingBlocks.end(); I != E; ++I) 175 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator())) 176 if (BI->isConditional()) { 177 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) { 178 179 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "; 180 WriteAsOperand(dbgs(), *I, false); 181 dbgs() << "\n"); 182 183 BI->setCondition(ConstantInt::get(Cond->getType(), 184 !L->contains(BI->getSuccessor(0)))); 185 Changed = true; 186 } 187 } 188 189 // Does the loop already have a preheader? If so, don't insert one. 190 BasicBlock *Preheader = L->getLoopPreheader(); 191 if (!Preheader) { 192 Preheader = InsertPreheaderForLoop(L); 193 if (Preheader) { 194 NumInserted++; 195 Changed = true; 196 } 197 } 198 199 // Next, check to make sure that all exit nodes of the loop only have 200 // predecessors that are inside of the loop. This check guarantees that the 201 // loop preheader/header will dominate the exit blocks. If the exit block has 202 // predecessors from outside of the loop, split the edge now. 203 SmallVector<BasicBlock*, 8> ExitBlocks; 204 L->getExitBlocks(ExitBlocks); 205 206 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(), 207 ExitBlocks.end()); 208 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(), 209 E = ExitBlockSet.end(); I != E; ++I) { 210 BasicBlock *ExitBlock = *I; 211 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); 212 PI != PE; ++PI) 213 // Must be exactly this loop: no subloops, parent loops, or non-loop preds 214 // allowed. 215 if (!L->contains(*PI)) { 216 if (RewriteLoopExitBlock(L, ExitBlock)) { 217 NumInserted++; 218 Changed = true; 219 } 220 break; 221 } 222 } 223 224 // If the header has more than two predecessors at this point (from the 225 // preheader and from multiple backedges), we must adjust the loop. 226 BasicBlock *LoopLatch = L->getLoopLatch(); 227 if (!LoopLatch) { 228 // If this is really a nested loop, rip it out into a child loop. Don't do 229 // this for loops with a giant number of backedges, just factor them into a 230 // common backedge instead. 231 if (L->getNumBackEdges() < 8) { 232 if (SeparateNestedLoop(L, LPM)) { 233 ++NumNested; 234 // This is a big restructuring change, reprocess the whole loop. 235 Changed = true; 236 // GCC doesn't tail recursion eliminate this. 237 goto ReprocessLoop; 238 } 239 } 240 241 // If we either couldn't, or didn't want to, identify nesting of the loops, 242 // insert a new block that all backedges target, then make it jump to the 243 // loop header. 244 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader); 245 if (LoopLatch) { 246 NumInserted++; 247 Changed = true; 248 } 249 } 250 251 // Scan over the PHI nodes in the loop header. Since they now have only two 252 // incoming values (the loop is canonicalized), we may have simplified the PHI 253 // down to 'X = phi [X, Y]', which should be replaced with 'Y'. 254 PHINode *PN; 255 for (BasicBlock::iterator I = L->getHeader()->begin(); 256 (PN = dyn_cast<PHINode>(I++)); ) 257 if (Value *V = PN->hasConstantValue(DT)) { 258 if (AA) AA->deleteValue(PN); 259 PN->replaceAllUsesWith(V); 260 PN->eraseFromParent(); 261 } 262 263 // If this loop has multiple exits and the exits all go to the same 264 // block, attempt to merge the exits. This helps several passes, such 265 // as LoopRotation, which do not support loops with multiple exits. 266 // SimplifyCFG also does this (and this code uses the same utility 267 // function), however this code is loop-aware, where SimplifyCFG is 268 // not. That gives it the advantage of being able to hoist 269 // loop-invariant instructions out of the way to open up more 270 // opportunities, and the disadvantage of having the responsibility 271 // to preserve dominator information. 272 bool UniqueExit = true; 273 if (!ExitBlocks.empty()) 274 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i) 275 if (ExitBlocks[i] != ExitBlocks[0]) { 276 UniqueExit = false; 277 break; 278 } 279 if (UniqueExit) { 280 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 281 BasicBlock *ExitingBlock = ExitingBlocks[i]; 282 if (!ExitingBlock->getSinglePredecessor()) continue; 283 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 284 if (!BI || !BI->isConditional()) continue; 285 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition()); 286 if (!CI || CI->getParent() != ExitingBlock) continue; 287 288 // Attempt to hoist out all instructions except for the 289 // comparison and the branch. 290 bool AllInvariant = true; 291 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) { 292 Instruction *Inst = I++; 293 if (Inst == CI) 294 continue; 295 if (!L->makeLoopInvariant(Inst, Changed, 296 Preheader ? Preheader->getTerminator() : 0)) { 297 AllInvariant = false; 298 break; 299 } 300 } 301 if (!AllInvariant) continue; 302 303 // The block has now been cleared of all instructions except for 304 // a comparison and a conditional branch. SimplifyCFG may be able 305 // to fold it now. 306 if (!FoldBranchToCommonDest(BI)) continue; 307 308 // Success. The block is now dead, so remove it from the loop, 309 // update the dominator tree and dominance frontier, and delete it. 310 311 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "; 312 WriteAsOperand(dbgs(), ExitingBlock, false); 313 dbgs() << "\n"); 314 315 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock)); 316 Changed = true; 317 LI->removeBlock(ExitingBlock); 318 319 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>(); 320 DomTreeNode *Node = DT->getNode(ExitingBlock); 321 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children = 322 Node->getChildren(); 323 while (!Children.empty()) { 324 DomTreeNode *Child = Children.front(); 325 DT->changeImmediateDominator(Child, Node->getIDom()); 326 if (DF) DF->changeImmediateDominator(Child->getBlock(), 327 Node->getIDom()->getBlock(), 328 DT); 329 } 330 DT->eraseNode(ExitingBlock); 331 if (DF) DF->removeBlock(ExitingBlock); 332 333 BI->getSuccessor(0)->removePredecessor(ExitingBlock); 334 BI->getSuccessor(1)->removePredecessor(ExitingBlock); 335 ExitingBlock->eraseFromParent(); 336 } 337 } 338 339 return Changed; 340} 341 342/// InsertPreheaderForLoop - Once we discover that a loop doesn't have a 343/// preheader, this method is called to insert one. This method has two phases: 344/// preheader insertion and analysis updating. 345/// 346BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) { 347 BasicBlock *Header = L->getHeader(); 348 349 // Compute the set of predecessors of the loop that are not in the loop. 350 SmallVector<BasicBlock*, 8> OutsideBlocks; 351 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header); 352 PI != PE; ++PI) 353 if (!L->contains(*PI)) { // Coming in from outside the loop? 354 // If the loop is branched to from an indirect branch, we won't 355 // be able to fully transform the loop, because it prohibits 356 // edge splitting. 357 if (isa<IndirectBrInst>((*PI)->getTerminator())) return 0; 358 359 // Keep track of it. 360 OutsideBlocks.push_back(*PI); 361 } 362 363 // Split out the loop pre-header. 364 BasicBlock *NewBB = 365 SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(), 366 ".preheader", this); 367 368 DEBUG(dbgs() << "LoopSimplify: Creating pre-header "; 369 WriteAsOperand(dbgs(), NewBB, false); 370 dbgs() << "\n"); 371 372 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 373 // code layout too horribly. 374 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L); 375 376 return NewBB; 377} 378 379/// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit 380/// blocks. This method is used to split exit blocks that have predecessors 381/// outside of the loop. 382BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) { 383 SmallVector<BasicBlock*, 8> LoopBlocks; 384 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) 385 if (L->contains(*I)) { 386 // Don't do this if the loop is exited via an indirect branch. 387 if (isa<IndirectBrInst>((*I)->getTerminator())) return 0; 388 389 LoopBlocks.push_back(*I); 390 } 391 392 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?"); 393 BasicBlock *NewBB = SplitBlockPredecessors(Exit, &LoopBlocks[0], 394 LoopBlocks.size(), ".loopexit", 395 this); 396 397 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "; 398 WriteAsOperand(dbgs(), NewBB, false); 399 dbgs() << "\n"); 400 401 return NewBB; 402} 403 404/// AddBlockAndPredsToSet - Add the specified block, and all of its 405/// predecessors, to the specified set, if it's not already in there. Stop 406/// predecessor traversal when we reach StopBlock. 407static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock, 408 std::set<BasicBlock*> &Blocks) { 409 std::vector<BasicBlock *> WorkList; 410 WorkList.push_back(InputBB); 411 do { 412 BasicBlock *BB = WorkList.back(); WorkList.pop_back(); 413 if (Blocks.insert(BB).second && BB != StopBlock) 414 // If BB is not already processed and it is not a stop block then 415 // insert its predecessor in the work list 416 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) { 417 BasicBlock *WBB = *I; 418 WorkList.push_back(WBB); 419 } 420 } while(!WorkList.empty()); 421} 422 423/// FindPHIToPartitionLoops - The first part of loop-nestification is to find a 424/// PHI node that tells us how to partition the loops. 425static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT, 426 AliasAnalysis *AA) { 427 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { 428 PHINode *PN = cast<PHINode>(I); 429 ++I; 430 if (Value *V = PN->hasConstantValue(DT)) { 431 // This is a degenerate PHI already, don't modify it! 432 PN->replaceAllUsesWith(V); 433 if (AA) AA->deleteValue(PN); 434 PN->eraseFromParent(); 435 continue; 436 } 437 438 // Scan this PHI node looking for a use of the PHI node by itself. 439 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 440 if (PN->getIncomingValue(i) == PN && 441 L->contains(PN->getIncomingBlock(i))) 442 // We found something tasty to remove. 443 return PN; 444 } 445 return 0; 446} 447 448// PlaceSplitBlockCarefully - If the block isn't already, move the new block to 449// right after some 'outside block' block. This prevents the preheader from 450// being placed inside the loop body, e.g. when the loop hasn't been rotated. 451void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB, 452 SmallVectorImpl<BasicBlock*> &SplitPreds, 453 Loop *L) { 454 // Check to see if NewBB is already well placed. 455 Function::iterator BBI = NewBB; --BBI; 456 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 457 if (&*BBI == SplitPreds[i]) 458 return; 459 } 460 461 // If it isn't already after an outside block, move it after one. This is 462 // always good as it makes the uncond branch from the outside block into a 463 // fall-through. 464 465 // Figure out *which* outside block to put this after. Prefer an outside 466 // block that neighbors a BB actually in the loop. 467 BasicBlock *FoundBB = 0; 468 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 469 Function::iterator BBI = SplitPreds[i]; 470 if (++BBI != NewBB->getParent()->end() && 471 L->contains(BBI)) { 472 FoundBB = SplitPreds[i]; 473 break; 474 } 475 } 476 477 // If our heuristic for a *good* bb to place this after doesn't find 478 // anything, just pick something. It's likely better than leaving it within 479 // the loop. 480 if (!FoundBB) 481 FoundBB = SplitPreds[0]; 482 NewBB->moveAfter(FoundBB); 483} 484 485 486/// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of 487/// them out into a nested loop. This is important for code that looks like 488/// this: 489/// 490/// Loop: 491/// ... 492/// br cond, Loop, Next 493/// ... 494/// br cond2, Loop, Out 495/// 496/// To identify this common case, we look at the PHI nodes in the header of the 497/// loop. PHI nodes with unchanging values on one backedge correspond to values 498/// that change in the "outer" loop, but not in the "inner" loop. 499/// 500/// If we are able to separate out a loop, return the new outer loop that was 501/// created. 502/// 503Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) { 504 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA); 505 if (PN == 0) return 0; // No known way to partition. 506 507 // Pull out all predecessors that have varying values in the loop. This 508 // handles the case when a PHI node has multiple instances of itself as 509 // arguments. 510 SmallVector<BasicBlock*, 8> OuterLoopPreds; 511 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 512 if (PN->getIncomingValue(i) != PN || 513 !L->contains(PN->getIncomingBlock(i))) { 514 // We can't split indirectbr edges. 515 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator())) 516 return 0; 517 518 OuterLoopPreds.push_back(PN->getIncomingBlock(i)); 519 } 520 521 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n"); 522 523 BasicBlock *Header = L->getHeader(); 524 BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0], 525 OuterLoopPreds.size(), 526 ".outer", this); 527 528 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 529 // code layout too horribly. 530 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L); 531 532 // Create the new outer loop. 533 Loop *NewOuter = new Loop(); 534 535 // Change the parent loop to use the outer loop as its child now. 536 if (Loop *Parent = L->getParentLoop()) 537 Parent->replaceChildLoopWith(L, NewOuter); 538 else 539 LI->changeTopLevelLoop(L, NewOuter); 540 541 // L is now a subloop of our outer loop. 542 NewOuter->addChildLoop(L); 543 544 // Add the new loop to the pass manager queue. 545 LPM.insertLoopIntoQueue(NewOuter); 546 547 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 548 I != E; ++I) 549 NewOuter->addBlockEntry(*I); 550 551 // Now reset the header in L, which had been moved by 552 // SplitBlockPredecessors for the outer loop. 553 L->moveToHeader(Header); 554 555 // Determine which blocks should stay in L and which should be moved out to 556 // the Outer loop now. 557 std::set<BasicBlock*> BlocksInL; 558 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) 559 if (DT->dominates(Header, *PI)) 560 AddBlockAndPredsToSet(*PI, Header, BlocksInL); 561 562 563 // Scan all of the loop children of L, moving them to OuterLoop if they are 564 // not part of the inner loop. 565 const std::vector<Loop*> &SubLoops = L->getSubLoops(); 566 for (size_t I = 0; I != SubLoops.size(); ) 567 if (BlocksInL.count(SubLoops[I]->getHeader())) 568 ++I; // Loop remains in L 569 else 570 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I)); 571 572 // Now that we know which blocks are in L and which need to be moved to 573 // OuterLoop, move any blocks that need it. 574 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 575 BasicBlock *BB = L->getBlocks()[i]; 576 if (!BlocksInL.count(BB)) { 577 // Move this block to the parent, updating the exit blocks sets 578 L->removeBlockFromLoop(BB); 579 if ((*LI)[BB] == L) 580 LI->changeLoopFor(BB, NewOuter); 581 --i; 582 } 583 } 584 585 return NewOuter; 586} 587 588 589 590/// InsertUniqueBackedgeBlock - This method is called when the specified loop 591/// has more than one backedge in it. If this occurs, revector all of these 592/// backedges to target a new basic block and have that block branch to the loop 593/// header. This ensures that loops have exactly one backedge. 594/// 595BasicBlock * 596LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) { 597 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!"); 598 599 // Get information about the loop 600 BasicBlock *Header = L->getHeader(); 601 Function *F = Header->getParent(); 602 603 // Unique backedge insertion currently depends on having a preheader. 604 if (!Preheader) 605 return 0; 606 607 // Figure out which basic blocks contain back-edges to the loop header. 608 std::vector<BasicBlock*> BackedgeBlocks; 609 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I) 610 if (*I != Preheader) BackedgeBlocks.push_back(*I); 611 612 // Create and insert the new backedge block... 613 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(), 614 Header->getName()+".backedge", F); 615 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock); 616 617 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "; 618 WriteAsOperand(dbgs(), BEBlock, false); 619 dbgs() << "\n"); 620 621 // Move the new backedge block to right after the last backedge block. 622 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos; 623 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); 624 625 // Now that the block has been inserted into the function, create PHI nodes in 626 // the backedge block which correspond to any PHI nodes in the header block. 627 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 628 PHINode *PN = cast<PHINode>(I); 629 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".be", 630 BETerminator); 631 NewPN->reserveOperandSpace(BackedgeBlocks.size()); 632 if (AA) AA->copyValue(PN, NewPN); 633 634 // Loop over the PHI node, moving all entries except the one for the 635 // preheader over to the new PHI node. 636 unsigned PreheaderIdx = ~0U; 637 bool HasUniqueIncomingValue = true; 638 Value *UniqueValue = 0; 639 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 640 BasicBlock *IBB = PN->getIncomingBlock(i); 641 Value *IV = PN->getIncomingValue(i); 642 if (IBB == Preheader) { 643 PreheaderIdx = i; 644 } else { 645 NewPN->addIncoming(IV, IBB); 646 if (HasUniqueIncomingValue) { 647 if (UniqueValue == 0) 648 UniqueValue = IV; 649 else if (UniqueValue != IV) 650 HasUniqueIncomingValue = false; 651 } 652 } 653 } 654 655 // Delete all of the incoming values from the old PN except the preheader's 656 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??"); 657 if (PreheaderIdx != 0) { 658 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); 659 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); 660 } 661 // Nuke all entries except the zero'th. 662 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) 663 PN->removeIncomingValue(e-i, false); 664 665 // Finally, add the newly constructed PHI node as the entry for the BEBlock. 666 PN->addIncoming(NewPN, BEBlock); 667 668 // As an optimization, if all incoming values in the new PhiNode (which is a 669 // subset of the incoming values of the old PHI node) have the same value, 670 // eliminate the PHI Node. 671 if (HasUniqueIncomingValue) { 672 NewPN->replaceAllUsesWith(UniqueValue); 673 if (AA) AA->deleteValue(NewPN); 674 BEBlock->getInstList().erase(NewPN); 675 } 676 } 677 678 // Now that all of the PHI nodes have been inserted and adjusted, modify the 679 // backedge blocks to just to the BEBlock instead of the header. 680 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { 681 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator(); 682 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op) 683 if (TI->getSuccessor(Op) == Header) 684 TI->setSuccessor(Op, BEBlock); 685 } 686 687 //===--- Update all analyses which we must preserve now -----------------===// 688 689 // Update Loop Information - we know that this block is now in the current 690 // loop and all parent loops. 691 L->addBasicBlockToLoop(BEBlock, LI->getBase()); 692 693 // Update dominator information 694 DT->splitBlock(BEBlock); 695 if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>()) 696 DF->splitBlock(BEBlock); 697 698 return BEBlock; 699} 700 701void LoopSimplify::verifyAnalysis() const { 702 // It used to be possible to just assert L->isLoopSimplifyForm(), however 703 // with the introduction of indirectbr, there are now cases where it's 704 // not possible to transform a loop as necessary. We can at least check 705 // that there is an indirectbr near any time there's trouble. 706 707 // Indirectbr can interfere with preheader and unique backedge insertion. 708 if (!L->getLoopPreheader() || !L->getLoopLatch()) { 709 bool HasIndBrPred = false; 710 for (pred_iterator PI = pred_begin(L->getHeader()), 711 PE = pred_end(L->getHeader()); PI != PE; ++PI) 712 if (isa<IndirectBrInst>((*PI)->getTerminator())) { 713 HasIndBrPred = true; 714 break; 715 } 716 assert(HasIndBrPred && 717 "LoopSimplify has no excuse for missing loop header info!"); 718 } 719 720 // Indirectbr can interfere with exit block canonicalization. 721 if (!L->hasDedicatedExits()) { 722 bool HasIndBrExiting = false; 723 SmallVector<BasicBlock*, 8> ExitingBlocks; 724 L->getExitingBlocks(ExitingBlocks); 725 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) 726 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) { 727 HasIndBrExiting = true; 728 break; 729 } 730 assert(HasIndBrExiting && 731 "LoopSimplify has no excuse for missing exit block info!"); 732 } 733} 734