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 "loop-simplify" 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/InstructionSimplify.h" 51 #include "llvm/Analysis/LoopPass.h" 52 #include "llvm/Analysis/ScalarEvolution.h" 53 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 54 #include "llvm/Transforms/Utils/Local.h" 55 #include "llvm/Support/CFG.h" 56 #include "llvm/Support/Debug.h" 57 #include "llvm/ADT/SetOperations.h" 58 #include "llvm/ADT/SetVector.h" 59 #include "llvm/ADT/Statistic.h" 60 #include "llvm/ADT/DepthFirstIterator.h" 61 using namespace llvm; 62 63 STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted"); 64 STATISTIC(NumNested , "Number of nested loops split out"); 65 66 namespace { 67 struct LoopSimplify : public LoopPass { 68 static char ID; // Pass identification, replacement for typeid 69 LoopSimplify() : LoopPass(ID) { 70 initializeLoopSimplifyPass(*PassRegistry::getPassRegistry()); 71 } 72 73 // AA - If we have an alias analysis object to update, this is it, otherwise 74 // this is null. 75 AliasAnalysis *AA; 76 LoopInfo *LI; 77 DominatorTree *DT; 78 ScalarEvolution *SE; 79 Loop *L; 80 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 81 82 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 83 // We need loop information to identify the loops... 84 AU.addRequired<DominatorTree>(); 85 AU.addPreserved<DominatorTree>(); 86 87 AU.addRequired<LoopInfo>(); 88 AU.addPreserved<LoopInfo>(); 89 90 AU.addPreserved<AliasAnalysis>(); 91 AU.addPreserved<ScalarEvolution>(); 92 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. 93 } 94 95 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees. 96 void verifyAnalysis() const; 97 98 private: 99 bool ProcessLoop(Loop *L, LPPassManager &LPM); 100 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit); 101 BasicBlock *InsertPreheaderForLoop(Loop *L); 102 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM); 103 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader); 104 void PlaceSplitBlockCarefully(BasicBlock *NewBB, 105 SmallVectorImpl<BasicBlock*> &SplitPreds, 106 Loop *L); 107 }; 108 } 109 110 char LoopSimplify::ID = 0; 111 INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify", 112 "Canonicalize natural loops", true, false) 113 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 114 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 115 INITIALIZE_PASS_END(LoopSimplify, "loop-simplify", 116 "Canonicalize natural loops", true, false) 117 118 // Publicly exposed interface to pass... 119 char &llvm::LoopSimplifyID = LoopSimplify::ID; 120 Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } 121 122 /// runOnLoop - Run down all loops in the CFG (recursively, but we could do 123 /// it in any convenient order) inserting preheaders... 124 /// 125 bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) { 126 L = l; 127 bool Changed = false; 128 LI = &getAnalysis<LoopInfo>(); 129 AA = getAnalysisIfAvailable<AliasAnalysis>(); 130 DT = &getAnalysis<DominatorTree>(); 131 SE = getAnalysisIfAvailable<ScalarEvolution>(); 132 133 Changed |= ProcessLoop(L, LPM); 134 135 return Changed; 136 } 137 138 /// ProcessLoop - Walk the loop structure in depth first order, ensuring that 139 /// all loops have preheaders. 140 /// 141 bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) { 142 bool Changed = false; 143 ReprocessLoop: 144 145 // Check to see that no blocks (other than the header) in this loop have 146 // predecessors that are not in the loop. This is not valid for natural 147 // loops, but can occur if the blocks are unreachable. Since they are 148 // unreachable we can just shamelessly delete those CFG edges! 149 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 150 BB != E; ++BB) { 151 if (*BB == L->getHeader()) continue; 152 153 SmallPtrSet<BasicBlock*, 4> BadPreds; 154 for (pred_iterator PI = pred_begin(*BB), 155 PE = pred_end(*BB); PI != PE; ++PI) { 156 BasicBlock *P = *PI; 157 if (!L->contains(P)) 158 BadPreds.insert(P); 159 } 160 161 // Delete each unique out-of-loop (and thus dead) predecessor. 162 for (SmallPtrSet<BasicBlock*, 4>::iterator I = BadPreds.begin(), 163 E = BadPreds.end(); I != E; ++I) { 164 165 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor " 166 << (*I)->getName() << "\n"); 167 168 // Inform each successor of each dead pred. 169 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 170 (*SI)->removePredecessor(*I); 171 // Zap the dead pred's terminator and replace it with unreachable. 172 TerminatorInst *TI = (*I)->getTerminator(); 173 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 174 (*I)->getTerminator()->eraseFromParent(); 175 new UnreachableInst((*I)->getContext(), *I); 176 Changed = true; 177 } 178 } 179 180 // If there are exiting blocks with branches on undef, resolve the undef in 181 // the direction which will exit the loop. This will help simplify loop 182 // trip count computations. 183 SmallVector<BasicBlock*, 8> ExitingBlocks; 184 L->getExitingBlocks(ExitingBlocks); 185 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(), 186 E = ExitingBlocks.end(); I != E; ++I) 187 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator())) 188 if (BI->isConditional()) { 189 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) { 190 191 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in " 192 << (*I)->getName() << "\n"); 193 194 BI->setCondition(ConstantInt::get(Cond->getType(), 195 !L->contains(BI->getSuccessor(0)))); 196 Changed = true; 197 } 198 } 199 200 // Does the loop already have a preheader? If so, don't insert one. 201 BasicBlock *Preheader = L->getLoopPreheader(); 202 if (!Preheader) { 203 Preheader = InsertPreheaderForLoop(L); 204 if (Preheader) { 205 ++NumInserted; 206 Changed = true; 207 } 208 } 209 210 // Next, check to make sure that all exit nodes of the loop only have 211 // predecessors that are inside of the loop. This check guarantees that the 212 // loop preheader/header will dominate the exit blocks. If the exit block has 213 // predecessors from outside of the loop, split the edge now. 214 SmallVector<BasicBlock*, 8> ExitBlocks; 215 L->getExitBlocks(ExitBlocks); 216 217 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(), 218 ExitBlocks.end()); 219 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(), 220 E = ExitBlockSet.end(); I != E; ++I) { 221 BasicBlock *ExitBlock = *I; 222 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); 223 PI != PE; ++PI) 224 // Must be exactly this loop: no subloops, parent loops, or non-loop preds 225 // allowed. 226 if (!L->contains(*PI)) { 227 if (RewriteLoopExitBlock(L, ExitBlock)) { 228 ++NumInserted; 229 Changed = true; 230 } 231 break; 232 } 233 } 234 235 // If the header has more than two predecessors at this point (from the 236 // preheader and from multiple backedges), we must adjust the loop. 237 BasicBlock *LoopLatch = L->getLoopLatch(); 238 if (!LoopLatch) { 239 // If this is really a nested loop, rip it out into a child loop. Don't do 240 // this for loops with a giant number of backedges, just factor them into a 241 // common backedge instead. 242 if (L->getNumBackEdges() < 8) { 243 if (SeparateNestedLoop(L, LPM)) { 244 ++NumNested; 245 // This is a big restructuring change, reprocess the whole loop. 246 Changed = true; 247 // GCC doesn't tail recursion eliminate this. 248 goto ReprocessLoop; 249 } 250 } 251 252 // If we either couldn't, or didn't want to, identify nesting of the loops, 253 // insert a new block that all backedges target, then make it jump to the 254 // loop header. 255 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader); 256 if (LoopLatch) { 257 ++NumInserted; 258 Changed = true; 259 } 260 } 261 262 // Scan over the PHI nodes in the loop header. Since they now have only two 263 // incoming values (the loop is canonicalized), we may have simplified the PHI 264 // down to 'X = phi [X, Y]', which should be replaced with 'Y'. 265 PHINode *PN; 266 for (BasicBlock::iterator I = L->getHeader()->begin(); 267 (PN = dyn_cast<PHINode>(I++)); ) 268 if (Value *V = SimplifyInstruction(PN, 0, DT)) { 269 if (AA) AA->deleteValue(PN); 270 if (SE) SE->forgetValue(PN); 271 PN->replaceAllUsesWith(V); 272 PN->eraseFromParent(); 273 } 274 275 // If this loop has multiple exits and the exits all go to the same 276 // block, attempt to merge the exits. This helps several passes, such 277 // as LoopRotation, which do not support loops with multiple exits. 278 // SimplifyCFG also does this (and this code uses the same utility 279 // function), however this code is loop-aware, where SimplifyCFG is 280 // not. That gives it the advantage of being able to hoist 281 // loop-invariant instructions out of the way to open up more 282 // opportunities, and the disadvantage of having the responsibility 283 // to preserve dominator information. 284 bool UniqueExit = true; 285 if (!ExitBlocks.empty()) 286 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i) 287 if (ExitBlocks[i] != ExitBlocks[0]) { 288 UniqueExit = false; 289 break; 290 } 291 if (UniqueExit) { 292 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 293 BasicBlock *ExitingBlock = ExitingBlocks[i]; 294 if (!ExitingBlock->getSinglePredecessor()) continue; 295 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 296 if (!BI || !BI->isConditional()) continue; 297 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition()); 298 if (!CI || CI->getParent() != ExitingBlock) continue; 299 300 // Attempt to hoist out all instructions except for the 301 // comparison and the branch. 302 bool AllInvariant = true; 303 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) { 304 Instruction *Inst = I++; 305 // Skip debug info intrinsics. 306 if (isa<DbgInfoIntrinsic>(Inst)) 307 continue; 308 if (Inst == CI) 309 continue; 310 if (!L->makeLoopInvariant(Inst, Changed, 311 Preheader ? Preheader->getTerminator() : 0)) { 312 AllInvariant = false; 313 break; 314 } 315 } 316 if (!AllInvariant) continue; 317 318 // The block has now been cleared of all instructions except for 319 // a comparison and a conditional branch. SimplifyCFG may be able 320 // to fold it now. 321 if (!FoldBranchToCommonDest(BI)) continue; 322 323 // Success. The block is now dead, so remove it from the loop, 324 // update the dominator tree and delete it. 325 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block " 326 << ExitingBlock->getName() << "\n"); 327 328 // If any reachable control flow within this loop has changed, notify 329 // ScalarEvolution. Currently assume the parent loop doesn't change 330 // (spliting edges doesn't count). If blocks, CFG edges, or other values 331 // in the parent loop change, then we need call to forgetLoop() for the 332 // parent instead. 333 if (SE) 334 SE->forgetLoop(L); 335 336 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock)); 337 Changed = true; 338 LI->removeBlock(ExitingBlock); 339 340 DomTreeNode *Node = DT->getNode(ExitingBlock); 341 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children = 342 Node->getChildren(); 343 while (!Children.empty()) { 344 DomTreeNode *Child = Children.front(); 345 DT->changeImmediateDominator(Child, Node->getIDom()); 346 } 347 DT->eraseNode(ExitingBlock); 348 349 BI->getSuccessor(0)->removePredecessor(ExitingBlock); 350 BI->getSuccessor(1)->removePredecessor(ExitingBlock); 351 ExitingBlock->eraseFromParent(); 352 } 353 } 354 355 return Changed; 356 } 357 358 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a 359 /// preheader, this method is called to insert one. This method has two phases: 360 /// preheader insertion and analysis updating. 361 /// 362 BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) { 363 BasicBlock *Header = L->getHeader(); 364 365 // Compute the set of predecessors of the loop that are not in the loop. 366 SmallVector<BasicBlock*, 8> OutsideBlocks; 367 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header); 368 PI != PE; ++PI) { 369 BasicBlock *P = *PI; 370 if (!L->contains(P)) { // Coming in from outside the loop? 371 // If the loop is branched to from an indirect branch, we won't 372 // be able to fully transform the loop, because it prohibits 373 // edge splitting. 374 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 375 376 // Keep track of it. 377 OutsideBlocks.push_back(P); 378 } 379 } 380 381 // Split out the loop pre-header. 382 BasicBlock *NewBB = 383 SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(), 384 ".preheader", this); 385 386 NewBB->getTerminator()->setDebugLoc(Header->getFirstNonPHI()->getDebugLoc()); 387 DEBUG(dbgs() << "LoopSimplify: Creating pre-header " << NewBB->getName() 388 << "\n"); 389 390 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 391 // code layout too horribly. 392 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L); 393 394 return NewBB; 395 } 396 397 /// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit 398 /// blocks. This method is used to split exit blocks that have predecessors 399 /// outside of the loop. 400 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) { 401 SmallVector<BasicBlock*, 8> LoopBlocks; 402 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) { 403 BasicBlock *P = *I; 404 if (L->contains(P)) { 405 // Don't do this if the loop is exited via an indirect branch. 406 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 407 408 LoopBlocks.push_back(P); 409 } 410 } 411 412 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?"); 413 BasicBlock *NewExitBB = 0; 414 415 if (Exit->isLandingPad()) { 416 SmallVector<BasicBlock*, 2> NewBBs; 417 SplitLandingPadPredecessors(Exit, ArrayRef<BasicBlock*>(&LoopBlocks[0], 418 LoopBlocks.size()), 419 ".loopexit", ".nonloopexit", 420 this, NewBBs); 421 NewExitBB = NewBBs[0]; 422 } else { 423 NewExitBB = SplitBlockPredecessors(Exit, &LoopBlocks[0], 424 LoopBlocks.size(), ".loopexit", 425 this); 426 } 427 428 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block " 429 << NewExitBB->getName() << "\n"); 430 return NewExitBB; 431 } 432 433 /// AddBlockAndPredsToSet - Add the specified block, and all of its 434 /// predecessors, to the specified set, if it's not already in there. Stop 435 /// predecessor traversal when we reach StopBlock. 436 static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock, 437 std::set<BasicBlock*> &Blocks) { 438 std::vector<BasicBlock *> WorkList; 439 WorkList.push_back(InputBB); 440 do { 441 BasicBlock *BB = WorkList.back(); WorkList.pop_back(); 442 if (Blocks.insert(BB).second && BB != StopBlock) 443 // If BB is not already processed and it is not a stop block then 444 // insert its predecessor in the work list 445 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) { 446 BasicBlock *WBB = *I; 447 WorkList.push_back(WBB); 448 } 449 } while(!WorkList.empty()); 450 } 451 452 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a 453 /// PHI node that tells us how to partition the loops. 454 static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT, 455 AliasAnalysis *AA, LoopInfo *LI) { 456 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { 457 PHINode *PN = cast<PHINode>(I); 458 ++I; 459 if (Value *V = SimplifyInstruction(PN, 0, DT)) { 460 // This is a degenerate PHI already, don't modify it! 461 PN->replaceAllUsesWith(V); 462 if (AA) AA->deleteValue(PN); 463 PN->eraseFromParent(); 464 continue; 465 } 466 467 // Scan this PHI node looking for a use of the PHI node by itself. 468 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 469 if (PN->getIncomingValue(i) == PN && 470 L->contains(PN->getIncomingBlock(i))) 471 // We found something tasty to remove. 472 return PN; 473 } 474 return 0; 475 } 476 477 // PlaceSplitBlockCarefully - If the block isn't already, move the new block to 478 // right after some 'outside block' block. This prevents the preheader from 479 // being placed inside the loop body, e.g. when the loop hasn't been rotated. 480 void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB, 481 SmallVectorImpl<BasicBlock*> &SplitPreds, 482 Loop *L) { 483 // Check to see if NewBB is already well placed. 484 Function::iterator BBI = NewBB; --BBI; 485 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 486 if (&*BBI == SplitPreds[i]) 487 return; 488 } 489 490 // If it isn't already after an outside block, move it after one. This is 491 // always good as it makes the uncond branch from the outside block into a 492 // fall-through. 493 494 // Figure out *which* outside block to put this after. Prefer an outside 495 // block that neighbors a BB actually in the loop. 496 BasicBlock *FoundBB = 0; 497 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 498 Function::iterator BBI = SplitPreds[i]; 499 if (++BBI != NewBB->getParent()->end() && 500 L->contains(BBI)) { 501 FoundBB = SplitPreds[i]; 502 break; 503 } 504 } 505 506 // If our heuristic for a *good* bb to place this after doesn't find 507 // anything, just pick something. It's likely better than leaving it within 508 // the loop. 509 if (!FoundBB) 510 FoundBB = SplitPreds[0]; 511 NewBB->moveAfter(FoundBB); 512 } 513 514 515 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of 516 /// them out into a nested loop. This is important for code that looks like 517 /// this: 518 /// 519 /// Loop: 520 /// ... 521 /// br cond, Loop, Next 522 /// ... 523 /// br cond2, Loop, Out 524 /// 525 /// To identify this common case, we look at the PHI nodes in the header of the 526 /// loop. PHI nodes with unchanging values on one backedge correspond to values 527 /// that change in the "outer" loop, but not in the "inner" loop. 528 /// 529 /// If we are able to separate out a loop, return the new outer loop that was 530 /// created. 531 /// 532 Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) { 533 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA, LI); 534 if (PN == 0) return 0; // No known way to partition. 535 536 // Pull out all predecessors that have varying values in the loop. This 537 // handles the case when a PHI node has multiple instances of itself as 538 // arguments. 539 SmallVector<BasicBlock*, 8> OuterLoopPreds; 540 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 541 if (PN->getIncomingValue(i) != PN || 542 !L->contains(PN->getIncomingBlock(i))) { 543 // We can't split indirectbr edges. 544 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator())) 545 return 0; 546 547 OuterLoopPreds.push_back(PN->getIncomingBlock(i)); 548 } 549 550 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n"); 551 552 // If ScalarEvolution is around and knows anything about values in 553 // this loop, tell it to forget them, because we're about to 554 // substantially change it. 555 if (SE) 556 SE->forgetLoop(L); 557 558 BasicBlock *Header = L->getHeader(); 559 BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0], 560 OuterLoopPreds.size(), 561 ".outer", this); 562 563 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 564 // code layout too horribly. 565 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L); 566 567 // Create the new outer loop. 568 Loop *NewOuter = new Loop(); 569 570 // Change the parent loop to use the outer loop as its child now. 571 if (Loop *Parent = L->getParentLoop()) 572 Parent->replaceChildLoopWith(L, NewOuter); 573 else 574 LI->changeTopLevelLoop(L, NewOuter); 575 576 // L is now a subloop of our outer loop. 577 NewOuter->addChildLoop(L); 578 579 // Add the new loop to the pass manager queue. 580 LPM.insertLoopIntoQueue(NewOuter); 581 582 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 583 I != E; ++I) 584 NewOuter->addBlockEntry(*I); 585 586 // Now reset the header in L, which had been moved by 587 // SplitBlockPredecessors for the outer loop. 588 L->moveToHeader(Header); 589 590 // Determine which blocks should stay in L and which should be moved out to 591 // the Outer loop now. 592 std::set<BasicBlock*> BlocksInL; 593 for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) { 594 BasicBlock *P = *PI; 595 if (DT->dominates(Header, P)) 596 AddBlockAndPredsToSet(P, Header, BlocksInL); 597 } 598 599 // Scan all of the loop children of L, moving them to OuterLoop if they are 600 // not part of the inner loop. 601 const std::vector<Loop*> &SubLoops = L->getSubLoops(); 602 for (size_t I = 0; I != SubLoops.size(); ) 603 if (BlocksInL.count(SubLoops[I]->getHeader())) 604 ++I; // Loop remains in L 605 else 606 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I)); 607 608 // Now that we know which blocks are in L and which need to be moved to 609 // OuterLoop, move any blocks that need it. 610 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 611 BasicBlock *BB = L->getBlocks()[i]; 612 if (!BlocksInL.count(BB)) { 613 // Move this block to the parent, updating the exit blocks sets 614 L->removeBlockFromLoop(BB); 615 if ((*LI)[BB] == L) 616 LI->changeLoopFor(BB, NewOuter); 617 --i; 618 } 619 } 620 621 return NewOuter; 622 } 623 624 625 626 /// InsertUniqueBackedgeBlock - This method is called when the specified loop 627 /// has more than one backedge in it. If this occurs, revector all of these 628 /// backedges to target a new basic block and have that block branch to the loop 629 /// header. This ensures that loops have exactly one backedge. 630 /// 631 BasicBlock * 632 LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) { 633 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!"); 634 635 // Get information about the loop 636 BasicBlock *Header = L->getHeader(); 637 Function *F = Header->getParent(); 638 639 // Unique backedge insertion currently depends on having a preheader. 640 if (!Preheader) 641 return 0; 642 643 // Figure out which basic blocks contain back-edges to the loop header. 644 std::vector<BasicBlock*> BackedgeBlocks; 645 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){ 646 BasicBlock *P = *I; 647 648 // Indirectbr edges cannot be split, so we must fail if we find one. 649 if (isa<IndirectBrInst>(P->getTerminator())) 650 return 0; 651 652 if (P != Preheader) BackedgeBlocks.push_back(P); 653 } 654 655 // Create and insert the new backedge block... 656 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(), 657 Header->getName()+".backedge", F); 658 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock); 659 660 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block " 661 << BEBlock->getName() << "\n"); 662 663 // Move the new backedge block to right after the last backedge block. 664 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos; 665 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); 666 667 // Now that the block has been inserted into the function, create PHI nodes in 668 // the backedge block which correspond to any PHI nodes in the header block. 669 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 670 PHINode *PN = cast<PHINode>(I); 671 PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(), 672 PN->getName()+".be", BETerminator); 673 if (AA) AA->copyValue(PN, NewPN); 674 675 // Loop over the PHI node, moving all entries except the one for the 676 // preheader over to the new PHI node. 677 unsigned PreheaderIdx = ~0U; 678 bool HasUniqueIncomingValue = true; 679 Value *UniqueValue = 0; 680 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 681 BasicBlock *IBB = PN->getIncomingBlock(i); 682 Value *IV = PN->getIncomingValue(i); 683 if (IBB == Preheader) { 684 PreheaderIdx = i; 685 } else { 686 NewPN->addIncoming(IV, IBB); 687 if (HasUniqueIncomingValue) { 688 if (UniqueValue == 0) 689 UniqueValue = IV; 690 else if (UniqueValue != IV) 691 HasUniqueIncomingValue = false; 692 } 693 } 694 } 695 696 // Delete all of the incoming values from the old PN except the preheader's 697 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??"); 698 if (PreheaderIdx != 0) { 699 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); 700 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); 701 } 702 // Nuke all entries except the zero'th. 703 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) 704 PN->removeIncomingValue(e-i, false); 705 706 // Finally, add the newly constructed PHI node as the entry for the BEBlock. 707 PN->addIncoming(NewPN, BEBlock); 708 709 // As an optimization, if all incoming values in the new PhiNode (which is a 710 // subset of the incoming values of the old PHI node) have the same value, 711 // eliminate the PHI Node. 712 if (HasUniqueIncomingValue) { 713 NewPN->replaceAllUsesWith(UniqueValue); 714 if (AA) AA->deleteValue(NewPN); 715 BEBlock->getInstList().erase(NewPN); 716 } 717 } 718 719 // Now that all of the PHI nodes have been inserted and adjusted, modify the 720 // backedge blocks to just to the BEBlock instead of the header. 721 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { 722 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator(); 723 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op) 724 if (TI->getSuccessor(Op) == Header) 725 TI->setSuccessor(Op, BEBlock); 726 } 727 728 //===--- Update all analyses which we must preserve now -----------------===// 729 730 // Update Loop Information - we know that this block is now in the current 731 // loop and all parent loops. 732 L->addBasicBlockToLoop(BEBlock, LI->getBase()); 733 734 // Update dominator information 735 DT->splitBlock(BEBlock); 736 737 return BEBlock; 738 } 739 740 void LoopSimplify::verifyAnalysis() const { 741 // It used to be possible to just assert L->isLoopSimplifyForm(), however 742 // with the introduction of indirectbr, there are now cases where it's 743 // not possible to transform a loop as necessary. We can at least check 744 // that there is an indirectbr near any time there's trouble. 745 746 // Indirectbr can interfere with preheader and unique backedge insertion. 747 if (!L->getLoopPreheader() || !L->getLoopLatch()) { 748 bool HasIndBrPred = false; 749 for (pred_iterator PI = pred_begin(L->getHeader()), 750 PE = pred_end(L->getHeader()); PI != PE; ++PI) 751 if (isa<IndirectBrInst>((*PI)->getTerminator())) { 752 HasIndBrPred = true; 753 break; 754 } 755 assert(HasIndBrPred && 756 "LoopSimplify has no excuse for missing loop header info!"); 757 (void)HasIndBrPred; 758 } 759 760 // Indirectbr can interfere with exit block canonicalization. 761 if (!L->hasDedicatedExits()) { 762 bool HasIndBrExiting = false; 763 SmallVector<BasicBlock*, 8> ExitingBlocks; 764 L->getExitingBlocks(ExitingBlocks); 765 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 766 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) { 767 HasIndBrExiting = true; 768 break; 769 } 770 } 771 772 assert(HasIndBrExiting && 773 "LoopSimplify has no excuse for missing exit block info!"); 774 (void)HasIndBrExiting; 775 } 776 } 777