LoopSimplify.cpp revision 02dd53e1c5b941ca5f60fca1b95ebcaf9ccd1dfc
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/IntrinsicInst.h" 45#include "llvm/Function.h" 46#include "llvm/LLVMContext.h" 47#include "llvm/Type.h" 48#include "llvm/Analysis/AliasAnalysis.h" 49#include "llvm/Analysis/Dominators.h" 50#include "llvm/Analysis/LoopPass.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.addRequired<DominatorTree>(); 80 AU.addPreserved<DominatorTree>(); 81 82 AU.addRequired<LoopInfo>(); 83 AU.addPreserved<LoopInfo>(); 84 85 AU.addPreserved<AliasAnalysis>(); 86 AU.addPreserved("scalar-evolution"); 87 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. 88 AU.addPreserved<DominanceFrontier>(); 89 AU.addPreservedID(LCSSAID); 90 } 91 92 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees. 93 void verifyAnalysis() const; 94 95 private: 96 bool ProcessLoop(Loop *L, LPPassManager &LPM); 97 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit); 98 BasicBlock *InsertPreheaderForLoop(Loop *L); 99 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM); 100 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader); 101 void PlaceSplitBlockCarefully(BasicBlock *NewBB, 102 SmallVectorImpl<BasicBlock*> &SplitPreds, 103 Loop *L); 104 }; 105} 106 107char LoopSimplify::ID = 0; 108INITIALIZE_PASS(LoopSimplify, "loopsimplify", 109 "Canonicalize natural loops", true, false); 110 111// Publically exposed interface to pass... 112char &llvm::LoopSimplifyID = LoopSimplify::ID; 113Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } 114 115/// runOnLoop - Run down all loops in the CFG (recursively, but we could do 116/// it in any convenient order) inserting preheaders... 117/// 118bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) { 119 L = l; 120 bool Changed = false; 121 LI = &getAnalysis<LoopInfo>(); 122 AA = getAnalysisIfAvailable<AliasAnalysis>(); 123 DT = &getAnalysis<DominatorTree>(); 124 125 Changed |= ProcessLoop(L, LPM); 126 127 return Changed; 128} 129 130/// ProcessLoop - Walk the loop structure in depth first order, ensuring that 131/// all loops have preheaders. 132/// 133bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) { 134 bool Changed = false; 135ReprocessLoop: 136 137 // Check to see that no blocks (other than the header) in this loop have 138 // predecessors that are not in the loop. This is not valid for natural 139 // loops, but can occur if the blocks are unreachable. Since they are 140 // unreachable we can just shamelessly delete those CFG edges! 141 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 142 BB != E; ++BB) { 143 if (*BB == L->getHeader()) continue; 144 145 SmallPtrSet<BasicBlock*, 4> BadPreds; 146 for (pred_iterator PI = pred_begin(*BB), 147 PE = pred_end(*BB); PI != PE; ++PI) { 148 BasicBlock *P = *PI; 149 if (!L->contains(P)) 150 BadPreds.insert(P); 151 } 152 153 // Delete each unique out-of-loop (and thus dead) predecessor. 154 for (SmallPtrSet<BasicBlock*, 4>::iterator I = BadPreds.begin(), 155 E = BadPreds.end(); I != E; ++I) { 156 157 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "; 158 WriteAsOperand(dbgs(), *I, false); 159 dbgs() << "\n"); 160 161 // Inform each successor of each dead pred. 162 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 163 (*SI)->removePredecessor(*I); 164 // Zap the dead pred's terminator and replace it with unreachable. 165 TerminatorInst *TI = (*I)->getTerminator(); 166 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 167 (*I)->getTerminator()->eraseFromParent(); 168 new UnreachableInst((*I)->getContext(), *I); 169 Changed = true; 170 } 171 } 172 173 // If there are exiting blocks with branches on undef, resolve the undef in 174 // the direction which will exit the loop. This will help simplify loop 175 // trip count computations. 176 SmallVector<BasicBlock*, 8> ExitingBlocks; 177 L->getExitingBlocks(ExitingBlocks); 178 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(), 179 E = ExitingBlocks.end(); I != E; ++I) 180 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator())) 181 if (BI->isConditional()) { 182 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) { 183 184 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "; 185 WriteAsOperand(dbgs(), *I, false); 186 dbgs() << "\n"); 187 188 BI->setCondition(ConstantInt::get(Cond->getType(), 189 !L->contains(BI->getSuccessor(0)))); 190 Changed = true; 191 } 192 } 193 194 // Does the loop already have a preheader? If so, don't insert one. 195 BasicBlock *Preheader = L->getLoopPreheader(); 196 if (!Preheader) { 197 Preheader = InsertPreheaderForLoop(L); 198 if (Preheader) { 199 ++NumInserted; 200 Changed = true; 201 } 202 } 203 204 // Next, check to make sure that all exit nodes of the loop only have 205 // predecessors that are inside of the loop. This check guarantees that the 206 // loop preheader/header will dominate the exit blocks. If the exit block has 207 // predecessors from outside of the loop, split the edge now. 208 SmallVector<BasicBlock*, 8> ExitBlocks; 209 L->getExitBlocks(ExitBlocks); 210 211 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(), 212 ExitBlocks.end()); 213 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(), 214 E = ExitBlockSet.end(); I != E; ++I) { 215 BasicBlock *ExitBlock = *I; 216 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); 217 PI != PE; ++PI) 218 // Must be exactly this loop: no subloops, parent loops, or non-loop preds 219 // allowed. 220 if (!L->contains(*PI)) { 221 if (RewriteLoopExitBlock(L, ExitBlock)) { 222 ++NumInserted; 223 Changed = true; 224 } 225 break; 226 } 227 } 228 229 // If the header has more than two predecessors at this point (from the 230 // preheader and from multiple backedges), we must adjust the loop. 231 BasicBlock *LoopLatch = L->getLoopLatch(); 232 if (!LoopLatch) { 233 // If this is really a nested loop, rip it out into a child loop. Don't do 234 // this for loops with a giant number of backedges, just factor them into a 235 // common backedge instead. 236 if (L->getNumBackEdges() < 8) { 237 if (SeparateNestedLoop(L, LPM)) { 238 ++NumNested; 239 // This is a big restructuring change, reprocess the whole loop. 240 Changed = true; 241 // GCC doesn't tail recursion eliminate this. 242 goto ReprocessLoop; 243 } 244 } 245 246 // If we either couldn't, or didn't want to, identify nesting of the loops, 247 // insert a new block that all backedges target, then make it jump to the 248 // loop header. 249 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader); 250 if (LoopLatch) { 251 ++NumInserted; 252 Changed = true; 253 } 254 } 255 256 // Scan over the PHI nodes in the loop header. Since they now have only two 257 // incoming values (the loop is canonicalized), we may have simplified the PHI 258 // down to 'X = phi [X, Y]', which should be replaced with 'Y'. 259 PHINode *PN; 260 for (BasicBlock::iterator I = L->getHeader()->begin(); 261 (PN = dyn_cast<PHINode>(I++)); ) 262 if (Value *V = PN->hasConstantValue(DT)) { 263 if (AA) AA->deleteValue(PN); 264 PN->replaceAllUsesWith(V); 265 PN->eraseFromParent(); 266 } 267 268 // If this loop has multiple exits and the exits all go to the same 269 // block, attempt to merge the exits. This helps several passes, such 270 // as LoopRotation, which do not support loops with multiple exits. 271 // SimplifyCFG also does this (and this code uses the same utility 272 // function), however this code is loop-aware, where SimplifyCFG is 273 // not. That gives it the advantage of being able to hoist 274 // loop-invariant instructions out of the way to open up more 275 // opportunities, and the disadvantage of having the responsibility 276 // to preserve dominator information. 277 bool UniqueExit = true; 278 if (!ExitBlocks.empty()) 279 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i) 280 if (ExitBlocks[i] != ExitBlocks[0]) { 281 UniqueExit = false; 282 break; 283 } 284 if (UniqueExit) { 285 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 286 BasicBlock *ExitingBlock = ExitingBlocks[i]; 287 if (!ExitingBlock->getSinglePredecessor()) continue; 288 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 289 if (!BI || !BI->isConditional()) continue; 290 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition()); 291 if (!CI || CI->getParent() != ExitingBlock) continue; 292 293 // Attempt to hoist out all instructions except for the 294 // comparison and the branch. 295 bool AllInvariant = true; 296 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) { 297 Instruction *Inst = I++; 298 // Skip debug info intrinsics. 299 if (isa<DbgInfoIntrinsic>(Inst)) 300 continue; 301 if (Inst == CI) 302 continue; 303 if (!L->makeLoopInvariant(Inst, Changed, 304 Preheader ? Preheader->getTerminator() : 0)) { 305 AllInvariant = false; 306 break; 307 } 308 } 309 if (!AllInvariant) continue; 310 311 // The block has now been cleared of all instructions except for 312 // a comparison and a conditional branch. SimplifyCFG may be able 313 // to fold it now. 314 if (!FoldBranchToCommonDest(BI)) continue; 315 316 // Success. The block is now dead, so remove it from the loop, 317 // update the dominator tree and dominance frontier, and delete it. 318 319 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "; 320 WriteAsOperand(dbgs(), ExitingBlock, false); 321 dbgs() << "\n"); 322 323 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock)); 324 Changed = true; 325 LI->removeBlock(ExitingBlock); 326 327 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>(); 328 DomTreeNode *Node = DT->getNode(ExitingBlock); 329 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children = 330 Node->getChildren(); 331 while (!Children.empty()) { 332 DomTreeNode *Child = Children.front(); 333 DT->changeImmediateDominator(Child, Node->getIDom()); 334 if (DF) DF->changeImmediateDominator(Child->getBlock(), 335 Node->getIDom()->getBlock(), 336 DT); 337 } 338 DT->eraseNode(ExitingBlock); 339 if (DF) DF->removeBlock(ExitingBlock); 340 341 BI->getSuccessor(0)->removePredecessor(ExitingBlock); 342 BI->getSuccessor(1)->removePredecessor(ExitingBlock); 343 ExitingBlock->eraseFromParent(); 344 } 345 } 346 347 return Changed; 348} 349 350/// InsertPreheaderForLoop - Once we discover that a loop doesn't have a 351/// preheader, this method is called to insert one. This method has two phases: 352/// preheader insertion and analysis updating. 353/// 354BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) { 355 BasicBlock *Header = L->getHeader(); 356 357 // Compute the set of predecessors of the loop that are not in the loop. 358 SmallVector<BasicBlock*, 8> OutsideBlocks; 359 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header); 360 PI != PE; ++PI) { 361 BasicBlock *P = *PI; 362 if (!L->contains(P)) { // Coming in from outside the loop? 363 // If the loop is branched to from an indirect branch, we won't 364 // be able to fully transform the loop, because it prohibits 365 // edge splitting. 366 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 367 368 // Keep track of it. 369 OutsideBlocks.push_back(P); 370 } 371 } 372 373 // Split out the loop pre-header. 374 BasicBlock *NewBB = 375 SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(), 376 ".preheader", this); 377 378 DEBUG(dbgs() << "LoopSimplify: Creating pre-header "; 379 WriteAsOperand(dbgs(), NewBB, false); 380 dbgs() << "\n"); 381 382 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 383 // code layout too horribly. 384 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L); 385 386 return NewBB; 387} 388 389/// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit 390/// blocks. This method is used to split exit blocks that have predecessors 391/// outside of the loop. 392BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) { 393 SmallVector<BasicBlock*, 8> LoopBlocks; 394 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) { 395 BasicBlock *P = *I; 396 if (L->contains(P)) { 397 // Don't do this if the loop is exited via an indirect branch. 398 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 399 400 LoopBlocks.push_back(P); 401 } 402 } 403 404 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?"); 405 BasicBlock *NewBB = SplitBlockPredecessors(Exit, &LoopBlocks[0], 406 LoopBlocks.size(), ".loopexit", 407 this); 408 409 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "; 410 WriteAsOperand(dbgs(), NewBB, false); 411 dbgs() << "\n"); 412 413 return NewBB; 414} 415 416/// AddBlockAndPredsToSet - Add the specified block, and all of its 417/// predecessors, to the specified set, if it's not already in there. Stop 418/// predecessor traversal when we reach StopBlock. 419static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock, 420 std::set<BasicBlock*> &Blocks) { 421 std::vector<BasicBlock *> WorkList; 422 WorkList.push_back(InputBB); 423 do { 424 BasicBlock *BB = WorkList.back(); WorkList.pop_back(); 425 if (Blocks.insert(BB).second && BB != StopBlock) 426 // If BB is not already processed and it is not a stop block then 427 // insert its predecessor in the work list 428 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) { 429 BasicBlock *WBB = *I; 430 WorkList.push_back(WBB); 431 } 432 } while(!WorkList.empty()); 433} 434 435/// FindPHIToPartitionLoops - The first part of loop-nestification is to find a 436/// PHI node that tells us how to partition the loops. 437static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT, 438 AliasAnalysis *AA) { 439 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { 440 PHINode *PN = cast<PHINode>(I); 441 ++I; 442 if (Value *V = PN->hasConstantValue(DT)) { 443 // This is a degenerate PHI already, don't modify it! 444 PN->replaceAllUsesWith(V); 445 if (AA) AA->deleteValue(PN); 446 PN->eraseFromParent(); 447 continue; 448 } 449 450 // Scan this PHI node looking for a use of the PHI node by itself. 451 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 452 if (PN->getIncomingValue(i) == PN && 453 L->contains(PN->getIncomingBlock(i))) 454 // We found something tasty to remove. 455 return PN; 456 } 457 return 0; 458} 459 460// PlaceSplitBlockCarefully - If the block isn't already, move the new block to 461// right after some 'outside block' block. This prevents the preheader from 462// being placed inside the loop body, e.g. when the loop hasn't been rotated. 463void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB, 464 SmallVectorImpl<BasicBlock*> &SplitPreds, 465 Loop *L) { 466 // Check to see if NewBB is already well placed. 467 Function::iterator BBI = NewBB; --BBI; 468 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 469 if (&*BBI == SplitPreds[i]) 470 return; 471 } 472 473 // If it isn't already after an outside block, move it after one. This is 474 // always good as it makes the uncond branch from the outside block into a 475 // fall-through. 476 477 // Figure out *which* outside block to put this after. Prefer an outside 478 // block that neighbors a BB actually in the loop. 479 BasicBlock *FoundBB = 0; 480 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 481 Function::iterator BBI = SplitPreds[i]; 482 if (++BBI != NewBB->getParent()->end() && 483 L->contains(BBI)) { 484 FoundBB = SplitPreds[i]; 485 break; 486 } 487 } 488 489 // If our heuristic for a *good* bb to place this after doesn't find 490 // anything, just pick something. It's likely better than leaving it within 491 // the loop. 492 if (!FoundBB) 493 FoundBB = SplitPreds[0]; 494 NewBB->moveAfter(FoundBB); 495} 496 497 498/// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of 499/// them out into a nested loop. This is important for code that looks like 500/// this: 501/// 502/// Loop: 503/// ... 504/// br cond, Loop, Next 505/// ... 506/// br cond2, Loop, Out 507/// 508/// To identify this common case, we look at the PHI nodes in the header of the 509/// loop. PHI nodes with unchanging values on one backedge correspond to values 510/// that change in the "outer" loop, but not in the "inner" loop. 511/// 512/// If we are able to separate out a loop, return the new outer loop that was 513/// created. 514/// 515Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) { 516 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA); 517 if (PN == 0) return 0; // No known way to partition. 518 519 // Pull out all predecessors that have varying values in the loop. This 520 // handles the case when a PHI node has multiple instances of itself as 521 // arguments. 522 SmallVector<BasicBlock*, 8> OuterLoopPreds; 523 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 524 if (PN->getIncomingValue(i) != PN || 525 !L->contains(PN->getIncomingBlock(i))) { 526 // We can't split indirectbr edges. 527 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator())) 528 return 0; 529 530 OuterLoopPreds.push_back(PN->getIncomingBlock(i)); 531 } 532 533 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n"); 534 535 BasicBlock *Header = L->getHeader(); 536 BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0], 537 OuterLoopPreds.size(), 538 ".outer", this); 539 540 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 541 // code layout too horribly. 542 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L); 543 544 // Create the new outer loop. 545 Loop *NewOuter = new Loop(); 546 547 // Change the parent loop to use the outer loop as its child now. 548 if (Loop *Parent = L->getParentLoop()) 549 Parent->replaceChildLoopWith(L, NewOuter); 550 else 551 LI->changeTopLevelLoop(L, NewOuter); 552 553 // L is now a subloop of our outer loop. 554 NewOuter->addChildLoop(L); 555 556 // Add the new loop to the pass manager queue. 557 LPM.insertLoopIntoQueue(NewOuter); 558 559 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 560 I != E; ++I) 561 NewOuter->addBlockEntry(*I); 562 563 // Now reset the header in L, which had been moved by 564 // SplitBlockPredecessors for the outer loop. 565 L->moveToHeader(Header); 566 567 // Determine which blocks should stay in L and which should be moved out to 568 // the Outer loop now. 569 std::set<BasicBlock*> BlocksInL; 570 for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) { 571 BasicBlock *P = *PI; 572 if (DT->dominates(Header, P)) 573 AddBlockAndPredsToSet(P, Header, BlocksInL); 574 } 575 576 // Scan all of the loop children of L, moving them to OuterLoop if they are 577 // not part of the inner loop. 578 const std::vector<Loop*> &SubLoops = L->getSubLoops(); 579 for (size_t I = 0; I != SubLoops.size(); ) 580 if (BlocksInL.count(SubLoops[I]->getHeader())) 581 ++I; // Loop remains in L 582 else 583 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I)); 584 585 // Now that we know which blocks are in L and which need to be moved to 586 // OuterLoop, move any blocks that need it. 587 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 588 BasicBlock *BB = L->getBlocks()[i]; 589 if (!BlocksInL.count(BB)) { 590 // Move this block to the parent, updating the exit blocks sets 591 L->removeBlockFromLoop(BB); 592 if ((*LI)[BB] == L) 593 LI->changeLoopFor(BB, NewOuter); 594 --i; 595 } 596 } 597 598 return NewOuter; 599} 600 601 602 603/// InsertUniqueBackedgeBlock - This method is called when the specified loop 604/// has more than one backedge in it. If this occurs, revector all of these 605/// backedges to target a new basic block and have that block branch to the loop 606/// header. This ensures that loops have exactly one backedge. 607/// 608BasicBlock * 609LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) { 610 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!"); 611 612 // Get information about the loop 613 BasicBlock *Header = L->getHeader(); 614 Function *F = Header->getParent(); 615 616 // Unique backedge insertion currently depends on having a preheader. 617 if (!Preheader) 618 return 0; 619 620 // Figure out which basic blocks contain back-edges to the loop header. 621 std::vector<BasicBlock*> BackedgeBlocks; 622 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){ 623 BasicBlock *P = *I; 624 625 // Indirectbr edges cannot be split, so we must fail if we find one. 626 if (isa<IndirectBrInst>(P->getTerminator())) 627 return 0; 628 629 if (P != Preheader) BackedgeBlocks.push_back(P); 630 } 631 632 // Create and insert the new backedge block... 633 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(), 634 Header->getName()+".backedge", F); 635 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock); 636 637 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "; 638 WriteAsOperand(dbgs(), BEBlock, false); 639 dbgs() << "\n"); 640 641 // Move the new backedge block to right after the last backedge block. 642 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos; 643 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); 644 645 // Now that the block has been inserted into the function, create PHI nodes in 646 // the backedge block which correspond to any PHI nodes in the header block. 647 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 648 PHINode *PN = cast<PHINode>(I); 649 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".be", 650 BETerminator); 651 NewPN->reserveOperandSpace(BackedgeBlocks.size()); 652 if (AA) AA->copyValue(PN, NewPN); 653 654 // Loop over the PHI node, moving all entries except the one for the 655 // preheader over to the new PHI node. 656 unsigned PreheaderIdx = ~0U; 657 bool HasUniqueIncomingValue = true; 658 Value *UniqueValue = 0; 659 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 660 BasicBlock *IBB = PN->getIncomingBlock(i); 661 Value *IV = PN->getIncomingValue(i); 662 if (IBB == Preheader) { 663 PreheaderIdx = i; 664 } else { 665 NewPN->addIncoming(IV, IBB); 666 if (HasUniqueIncomingValue) { 667 if (UniqueValue == 0) 668 UniqueValue = IV; 669 else if (UniqueValue != IV) 670 HasUniqueIncomingValue = false; 671 } 672 } 673 } 674 675 // Delete all of the incoming values from the old PN except the preheader's 676 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??"); 677 if (PreheaderIdx != 0) { 678 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); 679 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); 680 } 681 // Nuke all entries except the zero'th. 682 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) 683 PN->removeIncomingValue(e-i, false); 684 685 // Finally, add the newly constructed PHI node as the entry for the BEBlock. 686 PN->addIncoming(NewPN, BEBlock); 687 688 // As an optimization, if all incoming values in the new PhiNode (which is a 689 // subset of the incoming values of the old PHI node) have the same value, 690 // eliminate the PHI Node. 691 if (HasUniqueIncomingValue) { 692 NewPN->replaceAllUsesWith(UniqueValue); 693 if (AA) AA->deleteValue(NewPN); 694 BEBlock->getInstList().erase(NewPN); 695 } 696 } 697 698 // Now that all of the PHI nodes have been inserted and adjusted, modify the 699 // backedge blocks to just to the BEBlock instead of the header. 700 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { 701 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator(); 702 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op) 703 if (TI->getSuccessor(Op) == Header) 704 TI->setSuccessor(Op, BEBlock); 705 } 706 707 //===--- Update all analyses which we must preserve now -----------------===// 708 709 // Update Loop Information - we know that this block is now in the current 710 // loop and all parent loops. 711 L->addBasicBlockToLoop(BEBlock, LI->getBase()); 712 713 // Update dominator information 714 DT->splitBlock(BEBlock); 715 if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>()) 716 DF->splitBlock(BEBlock); 717 718 return BEBlock; 719} 720 721void LoopSimplify::verifyAnalysis() const { 722 // It used to be possible to just assert L->isLoopSimplifyForm(), however 723 // with the introduction of indirectbr, there are now cases where it's 724 // not possible to transform a loop as necessary. We can at least check 725 // that there is an indirectbr near any time there's trouble. 726 727 // Indirectbr can interfere with preheader and unique backedge insertion. 728 if (!L->getLoopPreheader() || !L->getLoopLatch()) { 729 bool HasIndBrPred = false; 730 for (pred_iterator PI = pred_begin(L->getHeader()), 731 PE = pred_end(L->getHeader()); PI != PE; ++PI) 732 if (isa<IndirectBrInst>((*PI)->getTerminator())) { 733 HasIndBrPred = true; 734 break; 735 } 736 assert(HasIndBrPred && 737 "LoopSimplify has no excuse for missing loop header info!"); 738 } 739 740 // Indirectbr can interfere with exit block canonicalization. 741 if (!L->hasDedicatedExits()) { 742 bool HasIndBrExiting = false; 743 SmallVector<BasicBlock*, 8> ExitingBlocks; 744 L->getExitingBlocks(ExitingBlocks); 745 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) 746 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) { 747 HasIndBrExiting = true; 748 break; 749 } 750 assert(HasIndBrExiting && 751 "LoopSimplify has no excuse for missing exit block info!"); 752 } 753} 754