LoopSimplify.cpp revision ad190145912facc6fbf2fbe58023bb238fbf2365
1//===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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// Note that the simplifycfg pass will clean up blocks which are split out but 27// end up being unnecessary, so usage of this pass should not pessimize 28// generated code. 29// 30// This pass obviously modifies the CFG, but updates loop information and 31// dominator information. 32// 33//===----------------------------------------------------------------------===// 34 35#define DEBUG_TYPE "loopsimplify" 36#include "llvm/Transforms/Scalar.h" 37#include "llvm/Constant.h" 38#include "llvm/Instructions.h" 39#include "llvm/Function.h" 40#include "llvm/Type.h" 41#include "llvm/Analysis/AliasAnalysis.h" 42#include "llvm/Analysis/Dominators.h" 43#include "llvm/Analysis/LoopInfo.h" 44#include "llvm/Support/CFG.h" 45#include "llvm/Support/Compiler.h" 46#include "llvm/ADT/SetOperations.h" 47#include "llvm/ADT/SetVector.h" 48#include "llvm/ADT/Statistic.h" 49#include "llvm/ADT/DepthFirstIterator.h" 50using namespace llvm; 51 52STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted"); 53STATISTIC(NumNested , "Number of nested loops split out"); 54 55namespace { 56 struct VISIBILITY_HIDDEN LoopSimplify : public FunctionPass { 57 // AA - If we have an alias analysis object to update, this is it, otherwise 58 // this is null. 59 AliasAnalysis *AA; 60 LoopInfo *LI; 61 62 virtual bool runOnFunction(Function &F); 63 64 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 65 // We need loop information to identify the loops... 66 AU.addRequired<LoopInfo>(); 67 AU.addRequired<DominatorTree>(); 68 AU.addRequired<ETForest>(); 69 70 AU.addPreserved<LoopInfo>(); 71 AU.addPreserved<ImmediateDominators>(); 72 AU.addPreserved<ETForest>(); 73 AU.addPreserved<DominatorTree>(); 74 AU.addPreserved<DominanceFrontier>(); 75 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. 76 } 77 private: 78 bool ProcessLoop(Loop *L); 79 BasicBlock *SplitBlockPredecessors(BasicBlock *BB, const char *Suffix, 80 const std::vector<BasicBlock*> &Preds); 81 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit); 82 void InsertPreheaderForLoop(Loop *L); 83 Loop *SeparateNestedLoop(Loop *L); 84 void InsertUniqueBackedgeBlock(Loop *L); 85 void PlaceSplitBlockCarefully(BasicBlock *NewBB, 86 std::vector<BasicBlock*> &SplitPreds, 87 Loop *L); 88 89 void UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB, 90 std::vector<BasicBlock*> &PredBlocks); 91 }; 92 93 RegisterPass<LoopSimplify> 94 X("loopsimplify", "Canonicalize natural loops", true); 95} 96 97// Publically exposed interface to pass... 98const PassInfo *llvm::LoopSimplifyID = X.getPassInfo(); 99FunctionPass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } 100 101/// runOnFunction - Run down all loops in the CFG (recursively, but we could do 102/// it in any convenient order) inserting preheaders... 103/// 104bool LoopSimplify::runOnFunction(Function &F) { 105 bool Changed = false; 106 LI = &getAnalysis<LoopInfo>(); 107 AA = getAnalysisToUpdate<AliasAnalysis>(); 108 109 // Check to see that no blocks (other than the header) in loops have 110 // predecessors that are not in loops. This is not valid for natural loops, 111 // but can occur if the blocks are unreachable. Since they are unreachable we 112 // can just shamelessly destroy their terminators to make them not branch into 113 // the loop! 114 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 115 // This case can only occur for unreachable blocks. Blocks that are 116 // unreachable can't be in loops, so filter those blocks out. 117 if (LI->getLoopFor(BB)) continue; 118 119 bool BlockUnreachable = false; 120 TerminatorInst *TI = BB->getTerminator(); 121 122 // Check to see if any successors of this block are non-loop-header loops 123 // that are not the header. 124 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 125 // If this successor is not in a loop, BB is clearly ok. 126 Loop *L = LI->getLoopFor(TI->getSuccessor(i)); 127 if (!L) continue; 128 129 // If the succ is the loop header, and if L is a top-level loop, then this 130 // is an entrance into a loop through the header, which is also ok. 131 if (L->getHeader() == TI->getSuccessor(i) && L->getParentLoop() == 0) 132 continue; 133 134 // Otherwise, this is an entrance into a loop from some place invalid. 135 // Either the loop structure is invalid and this is not a natural loop (in 136 // which case the compiler is buggy somewhere else) or BB is unreachable. 137 BlockUnreachable = true; 138 break; 139 } 140 141 // If this block is ok, check the next one. 142 if (!BlockUnreachable) continue; 143 144 // Otherwise, this block is dead. To clean up the CFG and to allow later 145 // loop transformations to ignore this case, we delete the edges into the 146 // loop by replacing the terminator. 147 148 // Remove PHI entries from the successors. 149 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 150 TI->getSuccessor(i)->removePredecessor(BB); 151 152 // Add a new unreachable instruction. 153 new UnreachableInst(TI); 154 155 // Delete the dead terminator. 156 if (AA) AA->deleteValue(&BB->back()); 157 BB->getInstList().pop_back(); 158 Changed |= true; 159 } 160 161 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 162 Changed |= ProcessLoop(*I); 163 164 return Changed; 165} 166 167/// ProcessLoop - Walk the loop structure in depth first order, ensuring that 168/// all loops have preheaders. 169/// 170bool LoopSimplify::ProcessLoop(Loop *L) { 171 bool Changed = false; 172ReprocessLoop: 173 174 // Canonicalize inner loops before outer loops. Inner loop canonicalization 175 // can provide work for the outer loop to canonicalize. 176 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 177 Changed |= ProcessLoop(*I); 178 179 assert(L->getBlocks()[0] == L->getHeader() && 180 "Header isn't first block in loop?"); 181 182 // Does the loop already have a preheader? If so, don't insert one. 183 if (L->getLoopPreheader() == 0) { 184 InsertPreheaderForLoop(L); 185 NumInserted++; 186 Changed = true; 187 } 188 189 // Next, check to make sure that all exit nodes of the loop only have 190 // predecessors that are inside of the loop. This check guarantees that the 191 // loop preheader/header will dominate the exit blocks. If the exit block has 192 // predecessors from outside of the loop, split the edge now. 193 std::vector<BasicBlock*> ExitBlocks; 194 L->getExitBlocks(ExitBlocks); 195 196 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end()); 197 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(), 198 E = ExitBlockSet.end(); I != E; ++I) { 199 BasicBlock *ExitBlock = *I; 200 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); 201 PI != PE; ++PI) 202 // Must be exactly this loop: no subloops, parent loops, or non-loop preds 203 // allowed. 204 if (!L->contains(*PI)) { 205 RewriteLoopExitBlock(L, ExitBlock); 206 NumInserted++; 207 Changed = true; 208 break; 209 } 210 } 211 212 // If the header has more than two predecessors at this point (from the 213 // preheader and from multiple backedges), we must adjust the loop. 214 unsigned NumBackedges = L->getNumBackEdges(); 215 if (NumBackedges != 1) { 216 // If this is really a nested loop, rip it out into a child loop. Don't do 217 // this for loops with a giant number of backedges, just factor them into a 218 // common backedge instead. 219 if (NumBackedges < 8) { 220 if (Loop *NL = SeparateNestedLoop(L)) { 221 ++NumNested; 222 // This is a big restructuring change, reprocess the whole loop. 223 ProcessLoop(NL); 224 Changed = true; 225 // GCC doesn't tail recursion eliminate this. 226 goto ReprocessLoop; 227 } 228 } 229 230 // If we either couldn't, or didn't want to, identify nesting of the loops, 231 // insert a new block that all backedges target, then make it jump to the 232 // loop header. 233 InsertUniqueBackedgeBlock(L); 234 NumInserted++; 235 Changed = true; 236 } 237 238 // Scan over the PHI nodes in the loop header. Since they now have only two 239 // incoming values (the loop is canonicalized), we may have simplified the PHI 240 // down to 'X = phi [X, Y]', which should be replaced with 'Y'. 241 PHINode *PN; 242 for (BasicBlock::iterator I = L->getHeader()->begin(); 243 (PN = dyn_cast<PHINode>(I++)); ) 244 if (Value *V = PN->hasConstantValue()) { 245 PN->replaceAllUsesWith(V); 246 PN->eraseFromParent(); 247 } 248 249 return Changed; 250} 251 252/// SplitBlockPredecessors - Split the specified block into two blocks. We want 253/// to move the predecessors specified in the Preds list to point to the new 254/// block, leaving the remaining predecessors pointing to BB. This method 255/// updates the SSA PHINode's, but no other analyses. 256/// 257BasicBlock *LoopSimplify::SplitBlockPredecessors(BasicBlock *BB, 258 const char *Suffix, 259 const std::vector<BasicBlock*> &Preds) { 260 261 // Create new basic block, insert right before the original block... 262 BasicBlock *NewBB = new BasicBlock(BB->getName()+Suffix, BB->getParent(), BB); 263 264 // The preheader first gets an unconditional branch to the loop header... 265 BranchInst *BI = new BranchInst(BB, NewBB); 266 267 // For every PHI node in the block, insert a PHI node into NewBB where the 268 // incoming values from the out of loop edges are moved to NewBB. We have two 269 // possible cases here. If the loop is dead, we just insert dummy entries 270 // into the PHI nodes for the new edge. If the loop is not dead, we move the 271 // incoming edges in BB into new PHI nodes in NewBB. 272 // 273 if (!Preds.empty()) { // Is the loop not obviously dead? 274 // Check to see if the values being merged into the new block need PHI 275 // nodes. If so, insert them. 276 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ) { 277 PHINode *PN = cast<PHINode>(I); 278 ++I; 279 280 // Check to see if all of the values coming in are the same. If so, we 281 // don't need to create a new PHI node. 282 Value *InVal = PN->getIncomingValueForBlock(Preds[0]); 283 for (unsigned i = 1, e = Preds.size(); i != e; ++i) 284 if (InVal != PN->getIncomingValueForBlock(Preds[i])) { 285 InVal = 0; 286 break; 287 } 288 289 // If the values coming into the block are not the same, we need a PHI. 290 if (InVal == 0) { 291 // Create the new PHI node, insert it into NewBB at the end of the block 292 PHINode *NewPHI = new PHINode(PN->getType(), PN->getName()+".ph", BI); 293 if (AA) AA->copyValue(PN, NewPHI); 294 295 // Move all of the edges from blocks outside the loop to the new PHI 296 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 297 Value *V = PN->removeIncomingValue(Preds[i], false); 298 NewPHI->addIncoming(V, Preds[i]); 299 } 300 InVal = NewPHI; 301 } else { 302 // Remove all of the edges coming into the PHI nodes from outside of the 303 // block. 304 for (unsigned i = 0, e = Preds.size(); i != e; ++i) 305 PN->removeIncomingValue(Preds[i], false); 306 } 307 308 // Add an incoming value to the PHI node in the loop for the preheader 309 // edge. 310 PN->addIncoming(InVal, NewBB); 311 312 // Can we eliminate this phi node now? 313 if (Value *V = PN->hasConstantValue(true)) { 314 Instruction *I = dyn_cast<Instruction>(V); 315 // If I is in NewBB, the ETForest call will fail, because NewBB isn't 316 // registered in ETForest yet. Handle this case explicitly. 317 if (!I || (I->getParent() != NewBB && 318 getAnalysis<ETForest>().dominates(I, PN))) { 319 PN->replaceAllUsesWith(V); 320 if (AA) AA->deleteValue(PN); 321 BB->getInstList().erase(PN); 322 } 323 } 324 } 325 326 // Now that the PHI nodes are updated, actually move the edges from 327 // Preds to point to NewBB instead of BB. 328 // 329 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 330 TerminatorInst *TI = Preds[i]->getTerminator(); 331 for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) 332 if (TI->getSuccessor(s) == BB) 333 TI->setSuccessor(s, NewBB); 334 } 335 336 } else { // Otherwise the loop is dead... 337 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) { 338 PHINode *PN = cast<PHINode>(I); 339 // Insert dummy values as the incoming value... 340 PN->addIncoming(Constant::getNullValue(PN->getType()), NewBB); 341 } 342 } 343 return NewBB; 344} 345 346/// InsertPreheaderForLoop - Once we discover that a loop doesn't have a 347/// preheader, this method is called to insert one. This method has two phases: 348/// preheader insertion and analysis updating. 349/// 350void LoopSimplify::InsertPreheaderForLoop(Loop *L) { 351 BasicBlock *Header = L->getHeader(); 352 353 // Compute the set of predecessors of the loop that are not in the loop. 354 std::vector<BasicBlock*> OutsideBlocks; 355 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header); 356 PI != PE; ++PI) 357 if (!L->contains(*PI)) // Coming in from outside the loop? 358 OutsideBlocks.push_back(*PI); // Keep track of it... 359 360 // Split out the loop pre-header. 361 BasicBlock *NewBB = 362 SplitBlockPredecessors(Header, ".preheader", OutsideBlocks); 363 364 365 //===--------------------------------------------------------------------===// 366 // Update analysis results now that we have performed the transformation 367 // 368 369 // We know that we have loop information to update... update it now. 370 if (Loop *Parent = L->getParentLoop()) 371 Parent->addBasicBlockToLoop(NewBB, *LI); 372 373 UpdateDomInfoForRevectoredPreds(NewBB, OutsideBlocks); 374 375 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 376 // code layout too horribly. 377 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L); 378} 379 380/// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit 381/// blocks. This method is used to split exit blocks that have predecessors 382/// outside of the loop. 383BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) { 384 std::vector<BasicBlock*> LoopBlocks; 385 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) 386 if (L->contains(*I)) 387 LoopBlocks.push_back(*I); 388 389 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?"); 390 BasicBlock *NewBB = SplitBlockPredecessors(Exit, ".loopexit", LoopBlocks); 391 392 // Update Loop Information - we know that the new block will be in whichever 393 // loop the Exit block is in. Note that it may not be in that immediate loop, 394 // if the successor is some other loop header. In that case, we continue 395 // walking up the loop tree to find a loop that contains both the successor 396 // block and the predecessor block. 397 Loop *SuccLoop = LI->getLoopFor(Exit); 398 while (SuccLoop && !SuccLoop->contains(L->getHeader())) 399 SuccLoop = SuccLoop->getParentLoop(); 400 if (SuccLoop) 401 SuccLoop->addBasicBlockToLoop(NewBB, *LI); 402 403 // Update dominator information (set, immdom, domtree, and domfrontier) 404 UpdateDomInfoForRevectoredPreds(NewBB, LoopBlocks); 405 return NewBB; 406} 407 408/// AddBlockAndPredsToSet - Add the specified block, and all of its 409/// predecessors, to the specified set, if it's not already in there. Stop 410/// predecessor traversal when we reach StopBlock. 411static void AddBlockAndPredsToSet(BasicBlock *BB, BasicBlock *StopBlock, 412 std::set<BasicBlock*> &Blocks) { 413 if (!Blocks.insert(BB).second) return; // already processed. 414 if (BB == StopBlock) return; // Stop here! 415 416 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) 417 AddBlockAndPredsToSet(*I, StopBlock, Blocks); 418} 419 420/// FindPHIToPartitionLoops - The first part of loop-nestification is to find a 421/// PHI node that tells us how to partition the loops. 422static PHINode *FindPHIToPartitionLoops(Loop *L, ETForest *EF, 423 AliasAnalysis *AA) { 424 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { 425 PHINode *PN = cast<PHINode>(I); 426 ++I; 427 if (Value *V = PN->hasConstantValue()) 428 if (!isa<Instruction>(V) || EF->dominates(cast<Instruction>(V), PN)) { 429 // This is a degenerate PHI already, don't modify it! 430 PN->replaceAllUsesWith(V); 431 if (AA) AA->deleteValue(PN); 432 PN->eraseFromParent(); 433 continue; 434 } 435 436 // Scan this PHI node looking for a use of the PHI node by itself. 437 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 438 if (PN->getIncomingValue(i) == PN && 439 L->contains(PN->getIncomingBlock(i))) 440 // We found something tasty to remove. 441 return PN; 442 } 443 return 0; 444} 445 446// PlaceSplitBlockCarefully - If the block isn't already, move the new block to 447// right after some 'outside block' block. This prevents the preheader from 448// being placed inside the loop body, e.g. when the loop hasn't been rotated. 449void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB, 450 std::vector<BasicBlock*>&SplitPreds, 451 Loop *L) { 452 // Check to see if NewBB is already well placed. 453 Function::iterator BBI = NewBB; --BBI; 454 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 455 if (&*BBI == SplitPreds[i]) 456 return; 457 } 458 459 // If it isn't already after an outside block, move it after one. This is 460 // always good as it makes the uncond branch from the outside block into a 461 // fall-through. 462 463 // Figure out *which* outside block to put this after. Prefer an outside 464 // block that neighbors a BB actually in the loop. 465 BasicBlock *FoundBB = 0; 466 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 467 Function::iterator BBI = SplitPreds[i]; 468 if (++BBI != NewBB->getParent()->end() && 469 L->contains(BBI)) { 470 FoundBB = SplitPreds[i]; 471 break; 472 } 473 } 474 475 // If our heuristic for a *good* bb to place this after doesn't find 476 // anything, just pick something. It's likely better than leaving it within 477 // the loop. 478 if (!FoundBB) 479 FoundBB = SplitPreds[0]; 480 NewBB->moveAfter(FoundBB); 481} 482 483 484/// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of 485/// them out into a nested loop. This is important for code that looks like 486/// this: 487/// 488/// Loop: 489/// ... 490/// br cond, Loop, Next 491/// ... 492/// br cond2, Loop, Out 493/// 494/// To identify this common case, we look at the PHI nodes in the header of the 495/// loop. PHI nodes with unchanging values on one backedge correspond to values 496/// that change in the "outer" loop, but not in the "inner" loop. 497/// 498/// If we are able to separate out a loop, return the new outer loop that was 499/// created. 500/// 501Loop *LoopSimplify::SeparateNestedLoop(Loop *L) { 502 ETForest *EF = getAnalysisToUpdate<ETForest>(); 503 PHINode *PN = FindPHIToPartitionLoops(L, EF, AA); 504 if (PN == 0) return 0; // No known way to partition. 505 506 // Pull out all predecessors that have varying values in the loop. This 507 // handles the case when a PHI node has multiple instances of itself as 508 // arguments. 509 std::vector<BasicBlock*> OuterLoopPreds; 510 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 511 if (PN->getIncomingValue(i) != PN || 512 !L->contains(PN->getIncomingBlock(i))) 513 OuterLoopPreds.push_back(PN->getIncomingBlock(i)); 514 515 BasicBlock *Header = L->getHeader(); 516 BasicBlock *NewBB = SplitBlockPredecessors(Header, ".outer", OuterLoopPreds); 517 518 // Update dominator information (set, immdom, domtree, and domfrontier) 519 UpdateDomInfoForRevectoredPreds(NewBB, OuterLoopPreds); 520 521 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 522 // code layout too horribly. 523 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L); 524 525 // Create the new outer loop. 526 Loop *NewOuter = new Loop(); 527 528 // Change the parent loop to use the outer loop as its child now. 529 if (Loop *Parent = L->getParentLoop()) 530 Parent->replaceChildLoopWith(L, NewOuter); 531 else 532 LI->changeTopLevelLoop(L, NewOuter); 533 534 // This block is going to be our new header block: add it to this loop and all 535 // parent loops. 536 NewOuter->addBasicBlockToLoop(NewBB, *LI); 537 538 // L is now a subloop of our outer loop. 539 NewOuter->addChildLoop(L); 540 541 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) 542 NewOuter->addBlockEntry(L->getBlocks()[i]); 543 544 // Determine which blocks should stay in L and which should be moved out to 545 // the Outer loop now. 546 std::set<BasicBlock*> BlocksInL; 547 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) 548 if (EF->dominates(Header, *PI)) 549 AddBlockAndPredsToSet(*PI, Header, BlocksInL); 550 551 552 // Scan all of the loop children of L, moving them to OuterLoop if they are 553 // not part of the inner loop. 554 for (Loop::iterator I = L->begin(); I != L->end(); ) 555 if (BlocksInL.count((*I)->getHeader())) 556 ++I; // Loop remains in L 557 else 558 NewOuter->addChildLoop(L->removeChildLoop(I)); 559 560 // Now that we know which blocks are in L and which need to be moved to 561 // OuterLoop, move any blocks that need it. 562 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 563 BasicBlock *BB = L->getBlocks()[i]; 564 if (!BlocksInL.count(BB)) { 565 // Move this block to the parent, updating the exit blocks sets 566 L->removeBlockFromLoop(BB); 567 if ((*LI)[BB] == L) 568 LI->changeLoopFor(BB, NewOuter); 569 --i; 570 } 571 } 572 573 return NewOuter; 574} 575 576 577 578/// InsertUniqueBackedgeBlock - This method is called when the specified loop 579/// has more than one backedge in it. If this occurs, revector all of these 580/// backedges to target a new basic block and have that block branch to the loop 581/// header. This ensures that loops have exactly one backedge. 582/// 583void LoopSimplify::InsertUniqueBackedgeBlock(Loop *L) { 584 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!"); 585 586 // Get information about the loop 587 BasicBlock *Preheader = L->getLoopPreheader(); 588 BasicBlock *Header = L->getHeader(); 589 Function *F = Header->getParent(); 590 591 // Figure out which basic blocks contain back-edges to the loop header. 592 std::vector<BasicBlock*> BackedgeBlocks; 593 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I) 594 if (*I != Preheader) BackedgeBlocks.push_back(*I); 595 596 // Create and insert the new backedge block... 597 BasicBlock *BEBlock = new BasicBlock(Header->getName()+".backedge", F); 598 BranchInst *BETerminator = new BranchInst(Header, BEBlock); 599 600 // Move the new backedge block to right after the last backedge block. 601 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos; 602 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); 603 604 // Now that the block has been inserted into the function, create PHI nodes in 605 // the backedge block which correspond to any PHI nodes in the header block. 606 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 607 PHINode *PN = cast<PHINode>(I); 608 PHINode *NewPN = new PHINode(PN->getType(), PN->getName()+".be", 609 BETerminator); 610 NewPN->reserveOperandSpace(BackedgeBlocks.size()); 611 if (AA) AA->copyValue(PN, NewPN); 612 613 // Loop over the PHI node, moving all entries except the one for the 614 // preheader over to the new PHI node. 615 unsigned PreheaderIdx = ~0U; 616 bool HasUniqueIncomingValue = true; 617 Value *UniqueValue = 0; 618 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 619 BasicBlock *IBB = PN->getIncomingBlock(i); 620 Value *IV = PN->getIncomingValue(i); 621 if (IBB == Preheader) { 622 PreheaderIdx = i; 623 } else { 624 NewPN->addIncoming(IV, IBB); 625 if (HasUniqueIncomingValue) { 626 if (UniqueValue == 0) 627 UniqueValue = IV; 628 else if (UniqueValue != IV) 629 HasUniqueIncomingValue = false; 630 } 631 } 632 } 633 634 // Delete all of the incoming values from the old PN except the preheader's 635 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??"); 636 if (PreheaderIdx != 0) { 637 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); 638 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); 639 } 640 // Nuke all entries except the zero'th. 641 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) 642 PN->removeIncomingValue(e-i, false); 643 644 // Finally, add the newly constructed PHI node as the entry for the BEBlock. 645 PN->addIncoming(NewPN, BEBlock); 646 647 // As an optimization, if all incoming values in the new PhiNode (which is a 648 // subset of the incoming values of the old PHI node) have the same value, 649 // eliminate the PHI Node. 650 if (HasUniqueIncomingValue) { 651 NewPN->replaceAllUsesWith(UniqueValue); 652 if (AA) AA->deleteValue(NewPN); 653 BEBlock->getInstList().erase(NewPN); 654 } 655 } 656 657 // Now that all of the PHI nodes have been inserted and adjusted, modify the 658 // backedge blocks to just to the BEBlock instead of the header. 659 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { 660 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator(); 661 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op) 662 if (TI->getSuccessor(Op) == Header) 663 TI->setSuccessor(Op, BEBlock); 664 } 665 666 //===--- Update all analyses which we must preserve now -----------------===// 667 668 // Update Loop Information - we know that this block is now in the current 669 // loop and all parent loops. 670 L->addBasicBlockToLoop(BEBlock, *LI); 671 672 // Update dominator information (set, immdom, domtree, and domfrontier) 673 UpdateDomInfoForRevectoredPreds(BEBlock, BackedgeBlocks); 674} 675 676// Returns true if BasicBlock A dominates at least one block in vector B 677// Helper function for UpdateDomInfoForRevectoredPreds 678static bool BlockDominatesAny(BasicBlock* A, const std::vector<BasicBlock*>& B, const ETForest& ETF) { 679 for (std::vector<BasicBlock*>::iterator BI = B.begin(), BE = B.end(); BI != BE; ++BI) { 680 if (ETF.dominates(A, *BI)) 681 return true; 682 } 683 return false; 684} 685 686/// UpdateDomInfoForRevectoredPreds - This method is used to update the four 687/// different kinds of dominator information (immediate dominators, 688/// dominator trees, et-forest and dominance frontiers) after a new block has 689/// been added to the CFG. 690/// 691/// This only supports the case when an existing block (known as "NewBBSucc"), 692/// had some of its predecessors factored into a new basic block. This 693/// transformation inserts a new basic block ("NewBB"), with a single 694/// unconditional branch to NewBBSucc, and moves some predecessors of 695/// "NewBBSucc" to now branch to NewBB. These predecessors are listed in 696/// PredBlocks, even though they are the same as 697/// pred_begin(NewBB)/pred_end(NewBB). 698/// 699void LoopSimplify::UpdateDomInfoForRevectoredPreds(BasicBlock *NewBB, 700 std::vector<BasicBlock*> &PredBlocks) { 701 assert(!PredBlocks.empty() && "No predblocks??"); 702 assert(succ_begin(NewBB) != succ_end(NewBB) && 703 ++succ_begin(NewBB) == succ_end(NewBB) && 704 "NewBB should have a single successor!"); 705 BasicBlock *NewBBSucc = *succ_begin(NewBB); 706 ETForest& ETF = getAnalysis<ETForest>(); 707 708 // The newly inserted basic block will dominate existing basic blocks iff the 709 // PredBlocks dominate all of the non-pred blocks. If all predblocks dominate 710 // the non-pred blocks, then they all must be the same block! 711 // 712 bool NewBBDominatesNewBBSucc = true; 713 { 714 BasicBlock *OnePred = PredBlocks[0]; 715 unsigned i = 1, e = PredBlocks.size(); 716 for (i = 1; !ETF.isReachableFromEntry(OnePred); ++i) { 717 assert(i != e && "Didn't find reachable pred?"); 718 OnePred = PredBlocks[i]; 719 } 720 721 for (; i != e; ++i) 722 if (PredBlocks[i] != OnePred && ETF.isReachableFromEntry(OnePred)){ 723 NewBBDominatesNewBBSucc = false; 724 break; 725 } 726 727 if (NewBBDominatesNewBBSucc) 728 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc); 729 PI != E; ++PI) 730 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) { 731 NewBBDominatesNewBBSucc = false; 732 break; 733 } 734 } 735 736 // The other scenario where the new block can dominate its successors are when 737 // all predecessors of NewBBSucc that are not NewBB are dominated by NewBBSucc 738 // already. 739 if (!NewBBDominatesNewBBSucc) { 740 NewBBDominatesNewBBSucc = true; 741 for (pred_iterator PI = pred_begin(NewBBSucc), E = pred_end(NewBBSucc); 742 PI != E; ++PI) 743 if (*PI != NewBB && !ETF.dominates(NewBBSucc, *PI)) { 744 NewBBDominatesNewBBSucc = false; 745 break; 746 } 747 } 748 749 BasicBlock *NewBBIDom = 0; 750 751 // Update immediate dominator information if we have it. 752 if (ImmediateDominators *ID = getAnalysisToUpdate<ImmediateDominators>()) { 753 unsigned i = 0; 754 for (i = 0; i < PredBlocks.size(); ++i) 755 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) { 756 NewBBIDom = PredBlocks[i]; 757 break; 758 } 759 assert(i != PredBlocks.size() && "No reachable preds?"); 760 for (i = i + 1; i < PredBlocks.size(); ++i) { 761 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) 762 NewBBIDom = ETF.nearestCommonDominator(NewBBIDom, PredBlocks[i]); 763 } 764 assert(NewBBIDom && "No immediate dominator found??"); 765 766 // Set the immediate dominator now... 767 ID->addNewBlock(NewBB, NewBBIDom); 768 769 // If NewBB strictly dominates other blocks, we need to update their idom's 770 // now. The only block that need adjustment is the NewBBSucc block, whose 771 // idom should currently be set to PredBlocks[0]. 772 if (NewBBDominatesNewBBSucc) 773 ID->setImmediateDominator(NewBBSucc, NewBB); 774 } 775 776 // Update DominatorTree information if it is active. 777 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) { 778 // If we don't have ImmediateDominator info around, calculate the idom as 779 // above. 780 if (!NewBBIDom) { 781 unsigned i = 0; 782 for (i = 0; i < PredBlocks.size(); ++i) 783 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) { 784 NewBBIDom = PredBlocks[i]; 785 break; 786 } 787 assert(i != PredBlocks.size() && "No reachable preds?"); 788 for (i = i + 1; i < PredBlocks.size(); ++i) { 789 if (ETF.dominates(&PredBlocks[i]->getParent()->getEntryBlock(), PredBlocks[i])) 790 NewBBIDom = ETF.nearestCommonDominator(NewBBIDom, PredBlocks[i]); 791 } 792 assert(NewBBIDom && "No immediate dominator found??"); 793 } 794 DominatorTree::Node *NewBBIDomNode = DT->getNode(NewBBIDom); 795 796 // Create the new dominator tree node... and set the idom of NewBB. 797 DominatorTree::Node *NewBBNode = DT->createNewNode(NewBB, NewBBIDomNode); 798 799 // If NewBB strictly dominates other blocks, then it is now the immediate 800 // dominator of NewBBSucc. Update the dominator tree as appropriate. 801 if (NewBBDominatesNewBBSucc) { 802 DominatorTree::Node *NewBBSuccNode = DT->getNode(NewBBSucc); 803 DT->changeImmediateDominator(NewBBSuccNode, NewBBNode); 804 } 805 } 806 807 // Update ET-Forest information if it is active. 808 if (ETForest *EF = getAnalysisToUpdate<ETForest>()) { 809 EF->addNewBlock(NewBB, NewBBIDom); 810 if (NewBBDominatesNewBBSucc) 811 EF->setImmediateDominator(NewBBSucc, NewBB); 812 } 813 814 // Update dominance frontier information... 815 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>()) { 816 // If NewBB dominates NewBBSucc, then DF(NewBB) is now going to be the 817 // DF(PredBlocks[0]) without the stuff that the new block does not dominate 818 // a predecessor of. 819 if (NewBBDominatesNewBBSucc) { 820 DominanceFrontier::iterator DFI = DF->find(PredBlocks[0]); 821 if (DFI != DF->end()) { 822 DominanceFrontier::DomSetType Set = DFI->second; 823 // Filter out stuff in Set that we do not dominate a predecessor of. 824 for (DominanceFrontier::DomSetType::iterator SetI = Set.begin(), 825 E = Set.end(); SetI != E;) { 826 bool DominatesPred = false; 827 for (pred_iterator PI = pred_begin(*SetI), E = pred_end(*SetI); 828 PI != E; ++PI) 829 if (ETF.dominates(NewBB, *PI)) 830 DominatesPred = true; 831 if (!DominatesPred) 832 Set.erase(SetI++); 833 else 834 ++SetI; 835 } 836 837 DF->addBasicBlock(NewBB, Set); 838 } 839 840 } else { 841 // DF(NewBB) is {NewBBSucc} because NewBB does not strictly dominate 842 // NewBBSucc, but it does dominate itself (and there is an edge (NewBB -> 843 // NewBBSucc)). NewBBSucc is the single successor of NewBB. 844 DominanceFrontier::DomSetType NewDFSet; 845 NewDFSet.insert(NewBBSucc); 846 DF->addBasicBlock(NewBB, NewDFSet); 847 } 848 849 // Now we must loop over all of the dominance frontiers in the function, 850 // replacing occurrences of NewBBSucc with NewBB in some cases. All 851 // blocks that dominate a block in PredBlocks and contained NewBBSucc in 852 // their dominance frontier must be updated to contain NewBB instead. 853 // 854 for (Function::iterator FI = NewBB->getParent()->begin(), 855 FE = NewBB->getParent()->end(); FI != FE; ++FI) { 856 DominanceFrontier::iterator DFI = DF->find(FI); 857 if (DFI == DF->end()) continue; // unreachable block. 858 859 // Only consider dominators of NewBBSucc 860 if (!DFI->second.count(NewBBSucc)) continue; 861 862 if (BlockDominatesAny(FI, PredBlocks, ETF)) { 863 // If NewBBSucc should not stay in our dominator frontier, remove it. 864 // We remove it unless there is a predecessor of NewBBSucc that we 865 // dominate, but we don't strictly dominate NewBBSucc. 866 bool ShouldRemove = true; 867 if ((BasicBlock*)FI == NewBBSucc || !ETF.dominates(FI, NewBBSucc)) { 868 // Okay, we know that PredDom does not strictly dominate NewBBSucc. 869 // Check to see if it dominates any predecessors of NewBBSucc. 870 for (pred_iterator PI = pred_begin(NewBBSucc), 871 E = pred_end(NewBBSucc); PI != E; ++PI) 872 if (ETF.dominates(FI, *PI)) { 873 ShouldRemove = false; 874 break; 875 } 876 877 if (ShouldRemove) 878 DF->removeFromFrontier(DFI, NewBBSucc); 879 DF->addToFrontier(DFI, NewBB); 880 881 break; 882 } 883 } 884 } 885 } 886} 887 888 889