1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===// 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 transforms loops that contain branches on loop-invariant conditions 11 // to have multiple loops. For example, it turns the left into the right code: 12 // 13 // for (...) if (lic) 14 // A for (...) 15 // if (lic) A; B; C 16 // B else 17 // C for (...) 18 // A; C 19 // 20 // This can increase the size of the code exponentially (doubling it every time 21 // a loop is unswitched) so we only unswitch if the resultant code will be 22 // smaller than a threshold. 23 // 24 // This pass expects LICM to be run before it to hoist invariant conditions out 25 // of the loop, to make the unswitching opportunity obvious. 26 // 27 //===----------------------------------------------------------------------===// 28 29 #define DEBUG_TYPE "loop-unswitch" 30 #include "llvm/Transforms/Scalar.h" 31 #include "llvm/Constants.h" 32 #include "llvm/DerivedTypes.h" 33 #include "llvm/Function.h" 34 #include "llvm/Instructions.h" 35 #include "llvm/Analysis/InlineCost.h" 36 #include "llvm/Analysis/InstructionSimplify.h" 37 #include "llvm/Analysis/LoopInfo.h" 38 #include "llvm/Analysis/LoopPass.h" 39 #include "llvm/Analysis/Dominators.h" 40 #include "llvm/Analysis/ScalarEvolution.h" 41 #include "llvm/Transforms/Utils/Cloning.h" 42 #include "llvm/Transforms/Utils/Local.h" 43 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 44 #include "llvm/ADT/Statistic.h" 45 #include "llvm/ADT/SmallPtrSet.h" 46 #include "llvm/ADT/STLExtras.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Debug.h" 49 #include "llvm/Support/raw_ostream.h" 50 #include <algorithm> 51 #include <set> 52 using namespace llvm; 53 54 STATISTIC(NumBranches, "Number of branches unswitched"); 55 STATISTIC(NumSwitches, "Number of switches unswitched"); 56 STATISTIC(NumSelects , "Number of selects unswitched"); 57 STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 58 STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 59 60 // The specific value of 50 here was chosen based only on intuition and a 61 // few specific examples. 62 static cl::opt<unsigned> 63 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 64 cl::init(50), cl::Hidden); 65 66 namespace { 67 class LoopUnswitch : public LoopPass { 68 LoopInfo *LI; // Loop information 69 LPPassManager *LPM; 70 71 // LoopProcessWorklist - Used to check if second loop needs processing 72 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 73 std::vector<Loop*> LoopProcessWorklist; 74 SmallPtrSet<Value *,8> UnswitchedVals; 75 76 bool OptimizeForSize; 77 bool redoLoop; 78 79 Loop *currentLoop; 80 DominatorTree *DT; 81 BasicBlock *loopHeader; 82 BasicBlock *loopPreheader; 83 84 // LoopBlocks contains all of the basic blocks of the loop, including the 85 // preheader of the loop, the body of the loop, and the exit blocks of the 86 // loop, in that order. 87 std::vector<BasicBlock*> LoopBlocks; 88 // NewBlocks contained cloned copy of basic blocks from LoopBlocks. 89 std::vector<BasicBlock*> NewBlocks; 90 91 public: 92 static char ID; // Pass ID, replacement for typeid 93 explicit LoopUnswitch(bool Os = false) : 94 LoopPass(ID), OptimizeForSize(Os), redoLoop(false), 95 currentLoop(NULL), DT(NULL), loopHeader(NULL), 96 loopPreheader(NULL) { 97 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry()); 98 } 99 100 bool runOnLoop(Loop *L, LPPassManager &LPM); 101 bool processCurrentLoop(); 102 103 /// This transformation requires natural loop information & requires that 104 /// loop preheaders be inserted into the CFG. 105 /// 106 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 107 AU.addRequiredID(LoopSimplifyID); 108 AU.addPreservedID(LoopSimplifyID); 109 AU.addRequired<LoopInfo>(); 110 AU.addPreserved<LoopInfo>(); 111 AU.addRequiredID(LCSSAID); 112 AU.addPreservedID(LCSSAID); 113 AU.addPreserved<DominatorTree>(); 114 AU.addPreserved<ScalarEvolution>(); 115 } 116 117 private: 118 119 virtual void releaseMemory() { 120 UnswitchedVals.clear(); 121 } 122 123 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 124 /// remove it. 125 void RemoveLoopFromWorklist(Loop *L) { 126 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 127 LoopProcessWorklist.end(), L); 128 if (I != LoopProcessWorklist.end()) 129 LoopProcessWorklist.erase(I); 130 } 131 132 void initLoopData() { 133 loopHeader = currentLoop->getHeader(); 134 loopPreheader = currentLoop->getLoopPreheader(); 135 } 136 137 /// Split all of the edges from inside the loop to their exit blocks. 138 /// Update the appropriate Phi nodes as we do so. 139 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 140 141 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 142 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 143 BasicBlock *ExitBlock); 144 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 145 146 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 147 Constant *Val, bool isEqual); 148 149 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 150 BasicBlock *TrueDest, 151 BasicBlock *FalseDest, 152 Instruction *InsertPt); 153 154 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 155 void RemoveBlockIfDead(BasicBlock *BB, 156 std::vector<Instruction*> &Worklist, Loop *l); 157 void RemoveLoopFromHierarchy(Loop *L); 158 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 159 BasicBlock **LoopExit = 0); 160 161 }; 162 } 163 char LoopUnswitch::ID = 0; 164 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", 165 false, false) 166 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 167 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 168 INITIALIZE_PASS_DEPENDENCY(LCSSA) 169 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", 170 false, false) 171 172 Pass *llvm::createLoopUnswitchPass(bool Os) { 173 return new LoopUnswitch(Os); 174 } 175 176 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 177 /// invariant in the loop, or has an invariant piece, return the invariant. 178 /// Otherwise, return null. 179 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 180 // We can never unswitch on vector conditions. 181 if (Cond->getType()->isVectorTy()) 182 return 0; 183 184 // Constants should be folded, not unswitched on! 185 if (isa<Constant>(Cond)) return 0; 186 187 // TODO: Handle: br (VARIANT|INVARIANT). 188 189 // Hoist simple values out. 190 if (L->makeLoopInvariant(Cond, Changed)) 191 return Cond; 192 193 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 194 if (BO->getOpcode() == Instruction::And || 195 BO->getOpcode() == Instruction::Or) { 196 // If either the left or right side is invariant, we can unswitch on this, 197 // which will cause the branch to go away in one loop and the condition to 198 // simplify in the other one. 199 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 200 return LHS; 201 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 202 return RHS; 203 } 204 205 return 0; 206 } 207 208 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 209 LI = &getAnalysis<LoopInfo>(); 210 LPM = &LPM_Ref; 211 DT = getAnalysisIfAvailable<DominatorTree>(); 212 currentLoop = L; 213 Function *F = currentLoop->getHeader()->getParent(); 214 bool Changed = false; 215 do { 216 assert(currentLoop->isLCSSAForm(*DT)); 217 redoLoop = false; 218 Changed |= processCurrentLoop(); 219 } while(redoLoop); 220 221 if (Changed) { 222 // FIXME: Reconstruct dom info, because it is not preserved properly. 223 if (DT) 224 DT->runOnFunction(*F); 225 } 226 return Changed; 227 } 228 229 /// processCurrentLoop - Do actual work and unswitch loop if possible 230 /// and profitable. 231 bool LoopUnswitch::processCurrentLoop() { 232 bool Changed = false; 233 LLVMContext &Context = currentLoop->getHeader()->getContext(); 234 235 // Loop over all of the basic blocks in the loop. If we find an interior 236 // block that is branching on a loop-invariant condition, we can unswitch this 237 // loop. 238 for (Loop::block_iterator I = currentLoop->block_begin(), 239 E = currentLoop->block_end(); I != E; ++I) { 240 TerminatorInst *TI = (*I)->getTerminator(); 241 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 242 // If this isn't branching on an invariant condition, we can't unswitch 243 // it. 244 if (BI->isConditional()) { 245 // See if this, or some part of it, is loop invariant. If so, we can 246 // unswitch on it if we desire. 247 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 248 currentLoop, Changed); 249 if (LoopCond && UnswitchIfProfitable(LoopCond, 250 ConstantInt::getTrue(Context))) { 251 ++NumBranches; 252 return true; 253 } 254 } 255 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 256 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 257 currentLoop, Changed); 258 if (LoopCond && SI->getNumCases() > 1) { 259 // Find a value to unswitch on: 260 // FIXME: this should chose the most expensive case! 261 // FIXME: scan for a case with a non-critical edge? 262 Constant *UnswitchVal = SI->getCaseValue(1); 263 // Do not process same value again and again. 264 if (!UnswitchedVals.insert(UnswitchVal)) 265 continue; 266 267 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 268 ++NumSwitches; 269 return true; 270 } 271 } 272 } 273 274 // Scan the instructions to check for unswitchable values. 275 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 276 BBI != E; ++BBI) 277 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 278 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 279 currentLoop, Changed); 280 if (LoopCond && UnswitchIfProfitable(LoopCond, 281 ConstantInt::getTrue(Context))) { 282 ++NumSelects; 283 return true; 284 } 285 } 286 } 287 return Changed; 288 } 289 290 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the 291 /// loop with no side effects (including infinite loops). 292 /// 293 /// If true, we return true and set ExitBB to the block we 294 /// exit through. 295 /// 296 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 297 BasicBlock *&ExitBB, 298 std::set<BasicBlock*> &Visited) { 299 if (!Visited.insert(BB).second) { 300 // Already visited. Without more analysis, this could indicate an infinte loop. 301 return false; 302 } else if (!L->contains(BB)) { 303 // Otherwise, this is a loop exit, this is fine so long as this is the 304 // first exit. 305 if (ExitBB != 0) return false; 306 ExitBB = BB; 307 return true; 308 } 309 310 // Otherwise, this is an unvisited intra-loop node. Check all successors. 311 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 312 // Check to see if the successor is a trivial loop exit. 313 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 314 return false; 315 } 316 317 // Okay, everything after this looks good, check to make sure that this block 318 // doesn't include any side effects. 319 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 320 if (I->mayHaveSideEffects()) 321 return false; 322 323 return true; 324 } 325 326 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally 327 /// leads to an exit from the specified loop, and has no side-effects in the 328 /// process. If so, return the block that is exited to, otherwise return null. 329 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 330 std::set<BasicBlock*> Visited; 331 Visited.insert(L->getHeader()); // Branches to header make infinite loops. 332 BasicBlock *ExitBB = 0; 333 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 334 return ExitBB; 335 return 0; 336 } 337 338 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 339 /// trivial: that is, that the condition controls whether or not the loop does 340 /// anything at all. If this is a trivial condition, unswitching produces no 341 /// code duplications (equivalently, it produces a simpler loop and a new empty 342 /// loop, which gets deleted). 343 /// 344 /// If this is a trivial condition, return true, otherwise return false. When 345 /// returning true, this sets Cond and Val to the condition that controls the 346 /// trivial condition: when Cond dynamically equals Val, the loop is known to 347 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when 348 /// Cond == Val. 349 /// 350 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 351 BasicBlock **LoopExit) { 352 BasicBlock *Header = currentLoop->getHeader(); 353 TerminatorInst *HeaderTerm = Header->getTerminator(); 354 LLVMContext &Context = Header->getContext(); 355 356 BasicBlock *LoopExitBB = 0; 357 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 358 // If the header block doesn't end with a conditional branch on Cond, we 359 // can't handle it. 360 if (!BI->isConditional() || BI->getCondition() != Cond) 361 return false; 362 363 // Check to see if a successor of the branch is guaranteed to 364 // exit through a unique exit block without having any 365 // side-effects. If so, determine the value of Cond that causes it to do 366 // this. 367 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 368 BI->getSuccessor(0)))) { 369 if (Val) *Val = ConstantInt::getTrue(Context); 370 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 371 BI->getSuccessor(1)))) { 372 if (Val) *Val = ConstantInt::getFalse(Context); 373 } 374 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 375 // If this isn't a switch on Cond, we can't handle it. 376 if (SI->getCondition() != Cond) return false; 377 378 // Check to see if a successor of the switch is guaranteed to go to the 379 // latch block or exit through a one exit block without having any 380 // side-effects. If so, determine the value of Cond that causes it to do 381 // this. Note that we can't trivially unswitch on the default case. 382 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 383 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 384 SI->getSuccessor(i)))) { 385 // Okay, we found a trivial case, remember the value that is trivial. 386 if (Val) *Val = SI->getCaseValue(i); 387 break; 388 } 389 } 390 391 // If we didn't find a single unique LoopExit block, or if the loop exit block 392 // contains phi nodes, this isn't trivial. 393 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 394 return false; // Can't handle this. 395 396 if (LoopExit) *LoopExit = LoopExitBB; 397 398 // We already know that nothing uses any scalar values defined inside of this 399 // loop. As such, we just have to check to see if this loop will execute any 400 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 401 // part of the loop that the code *would* execute. We already checked the 402 // tail, check the header now. 403 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 404 if (I->mayHaveSideEffects()) 405 return false; 406 return true; 407 } 408 409 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 410 /// LoopCond == Val to simplify the loop. If we decide that this is profitable, 411 /// unswitch the loop, reprocess the pieces, then return true. 412 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { 413 414 initLoopData(); 415 416 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 417 if (!loopPreheader) 418 return false; 419 420 Function *F = loopHeader->getParent(); 421 422 Constant *CondVal = 0; 423 BasicBlock *ExitBlock = 0; 424 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 425 // If the condition is trivial, always unswitch. There is no code growth 426 // for this case. 427 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 428 return true; 429 } 430 431 // Check to see if it would be profitable to unswitch current loop. 432 433 // Do not do non-trivial unswitch while optimizing for size. 434 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) 435 return false; 436 437 // FIXME: This is overly conservative because it does not take into 438 // consideration code simplification opportunities and code that can 439 // be shared by the resultant unswitched loops. 440 CodeMetrics Metrics; 441 for (Loop::block_iterator I = currentLoop->block_begin(), 442 E = currentLoop->block_end(); 443 I != E; ++I) 444 Metrics.analyzeBasicBlock(*I); 445 446 // Limit the number of instructions to avoid causing significant code 447 // expansion, and the number of basic blocks, to avoid loops with 448 // large numbers of branches which cause loop unswitching to go crazy. 449 // This is a very ad-hoc heuristic. 450 if (Metrics.NumInsts > Threshold || 451 Metrics.NumBlocks * 5 > Threshold || 452 Metrics.containsIndirectBr || Metrics.isRecursive) { 453 DEBUG(dbgs() << "NOT unswitching loop %" 454 << currentLoop->getHeader()->getName() << ", cost too high: " 455 << currentLoop->getBlocks().size() << "\n"); 456 return false; 457 } 458 459 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 460 return true; 461 } 462 463 /// CloneLoop - Recursively clone the specified loop and all of its children, 464 /// mapping the blocks with the specified map. 465 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 466 LoopInfo *LI, LPPassManager *LPM) { 467 Loop *New = new Loop(); 468 LPM->insertLoop(New, PL); 469 470 // Add all of the blocks in L to the new loop. 471 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 472 I != E; ++I) 473 if (LI->getLoopFor(*I) == L) 474 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 475 476 // Add all of the subloops to the new loop. 477 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 478 CloneLoop(*I, New, VM, LI, LPM); 479 480 return New; 481 } 482 483 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 484 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 485 /// code immediately before InsertPt. 486 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 487 BasicBlock *TrueDest, 488 BasicBlock *FalseDest, 489 Instruction *InsertPt) { 490 // Insert a conditional branch on LIC to the two preheaders. The original 491 // code is the true version and the new code is the false version. 492 Value *BranchVal = LIC; 493 if (!isa<ConstantInt>(Val) || 494 Val->getType() != Type::getInt1Ty(LIC->getContext())) 495 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); 496 else if (Val != ConstantInt::getTrue(Val->getContext())) 497 // We want to enter the new loop when the condition is true. 498 std::swap(TrueDest, FalseDest); 499 500 // Insert the new branch. 501 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 502 503 // If either edge is critical, split it. This helps preserve LoopSimplify 504 // form for enclosing loops. 505 SplitCriticalEdge(BI, 0, this); 506 SplitCriticalEdge(BI, 1, this); 507 } 508 509 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 510 /// condition in it (a cond branch from its header block to its latch block, 511 /// where the path through the loop that doesn't execute its body has no 512 /// side-effects), unswitch it. This doesn't involve any code duplication, just 513 /// moving the conditional branch outside of the loop and updating loop info. 514 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 515 Constant *Val, 516 BasicBlock *ExitBlock) { 517 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 518 << loopHeader->getName() << " [" << L->getBlocks().size() 519 << " blocks] in Function " << L->getHeader()->getParent()->getName() 520 << " on cond: " << *Val << " == " << *Cond << "\n"); 521 522 // First step, split the preheader, so that we know that there is a safe place 523 // to insert the conditional branch. We will change loopPreheader to have a 524 // conditional branch on Cond. 525 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 526 527 // Now that we have a place to insert the conditional branch, create a place 528 // to branch to: this is the exit block out of the loop that we should 529 // short-circuit to. 530 531 // Split this block now, so that the loop maintains its exit block, and so 532 // that the jump from the preheader can execute the contents of the exit block 533 // without actually branching to it (the exit block should be dominated by the 534 // loop header, not the preheader). 535 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 536 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 537 538 // Okay, now we have a position to branch from and a position to branch to, 539 // insert the new conditional branch. 540 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 541 loopPreheader->getTerminator()); 542 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 543 loopPreheader->getTerminator()->eraseFromParent(); 544 545 // We need to reprocess this loop, it could be unswitched again. 546 redoLoop = true; 547 548 // Now that we know that the loop is never entered when this condition is a 549 // particular value, rewrite the loop with this info. We know that this will 550 // at least eliminate the old branch. 551 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 552 ++NumTrivial; 553 } 554 555 /// SplitExitEdges - Split all of the edges from inside the loop to their exit 556 /// blocks. Update the appropriate Phi nodes as we do so. 557 void LoopUnswitch::SplitExitEdges(Loop *L, 558 const SmallVector<BasicBlock *, 8> &ExitBlocks){ 559 560 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 561 BasicBlock *ExitBlock = ExitBlocks[i]; 562 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 563 pred_end(ExitBlock)); 564 565 // Although SplitBlockPredecessors doesn't preserve loop-simplify in 566 // general, if we call it on all predecessors of all exits then it does. 567 if (!ExitBlock->isLandingPad()) { 568 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), 569 ".us-lcssa", this); 570 } else { 571 SmallVector<BasicBlock*, 2> NewBBs; 572 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa", 573 this, NewBBs); 574 } 575 } 576 } 577 578 /// UnswitchNontrivialCondition - We determined that the loop is profitable 579 /// to unswitch when LIC equal Val. Split it into loop versions and test the 580 /// condition outside of either loop. Return the loops created as Out1/Out2. 581 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 582 Loop *L) { 583 Function *F = loopHeader->getParent(); 584 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 585 << loopHeader->getName() << " [" << L->getBlocks().size() 586 << " blocks] in Function " << F->getName() 587 << " when '" << *Val << "' == " << *LIC << "\n"); 588 589 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>()) 590 SE->forgetLoop(L); 591 592 LoopBlocks.clear(); 593 NewBlocks.clear(); 594 595 // First step, split the preheader and exit blocks, and add these blocks to 596 // the LoopBlocks list. 597 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 598 LoopBlocks.push_back(NewPreheader); 599 600 // We want the loop to come after the preheader, but before the exit blocks. 601 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 602 603 SmallVector<BasicBlock*, 8> ExitBlocks; 604 L->getUniqueExitBlocks(ExitBlocks); 605 606 // Split all of the edges from inside the loop to their exit blocks. Update 607 // the appropriate Phi nodes as we do so. 608 SplitExitEdges(L, ExitBlocks); 609 610 // The exit blocks may have been changed due to edge splitting, recompute. 611 ExitBlocks.clear(); 612 L->getUniqueExitBlocks(ExitBlocks); 613 614 // Add exit blocks to the loop blocks. 615 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 616 617 // Next step, clone all of the basic blocks that make up the loop (including 618 // the loop preheader and exit blocks), keeping track of the mapping between 619 // the instructions and blocks. 620 NewBlocks.reserve(LoopBlocks.size()); 621 ValueToValueMapTy VMap; 622 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 623 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 624 NewBlocks.push_back(NewBB); 625 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 626 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 627 } 628 629 // Splice the newly inserted blocks into the function right before the 630 // original preheader. 631 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), 632 NewBlocks[0], F->end()); 633 634 // Now we create the new Loop object for the versioned loop. 635 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 636 Loop *ParentLoop = L->getParentLoop(); 637 if (ParentLoop) { 638 // Make sure to add the cloned preheader and exit blocks to the parent loop 639 // as well. 640 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 641 } 642 643 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 644 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 645 // The new exit block should be in the same loop as the old one. 646 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 647 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 648 649 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 650 "Exit block should have been split to have one successor!"); 651 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 652 653 // If the successor of the exit block had PHI nodes, add an entry for 654 // NewExit. 655 PHINode *PN; 656 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) { 657 PN = cast<PHINode>(I); 658 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 659 ValueToValueMapTy::iterator It = VMap.find(V); 660 if (It != VMap.end()) V = It->second; 661 PN->addIncoming(V, NewExit); 662 } 663 664 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { 665 PN = PHINode::Create(LPad->getType(), 0, "", 666 ExitSucc->getFirstInsertionPt()); 667 668 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); 669 I != E; ++I) { 670 BasicBlock *BB = *I; 671 LandingPadInst *LPI = BB->getLandingPadInst(); 672 LPI->replaceAllUsesWith(PN); 673 PN->addIncoming(LPI, BB); 674 } 675 } 676 } 677 678 // Rewrite the code to refer to itself. 679 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 680 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 681 E = NewBlocks[i]->end(); I != E; ++I) 682 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); 683 684 // Rewrite the original preheader to select between versions of the loop. 685 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 686 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 687 "Preheader splitting did not work correctly!"); 688 689 // Emit the new branch that selects between the two versions of this loop. 690 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 691 LPM->deleteSimpleAnalysisValue(OldBR, L); 692 OldBR->eraseFromParent(); 693 694 LoopProcessWorklist.push_back(NewLoop); 695 redoLoop = true; 696 697 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 698 // deletes the instruction (for example by simplifying a PHI that feeds into 699 // the condition that we're unswitching on), we don't rewrite the second 700 // iteration. 701 WeakVH LICHandle(LIC); 702 703 // Now we rewrite the original code to know that the condition is true and the 704 // new code to know that the condition is false. 705 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 706 707 // It's possible that simplifying one loop could cause the other to be 708 // changed to another value or a constant. If its a constant, don't simplify 709 // it. 710 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 711 LICHandle && !isa<Constant>(LICHandle)) 712 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 713 } 714 715 /// RemoveFromWorklist - Remove all instances of I from the worklist vector 716 /// specified. 717 static void RemoveFromWorklist(Instruction *I, 718 std::vector<Instruction*> &Worklist) { 719 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 720 Worklist.end(), I); 721 while (WI != Worklist.end()) { 722 unsigned Offset = WI-Worklist.begin(); 723 Worklist.erase(WI); 724 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 725 } 726 } 727 728 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 729 /// program, replacing all uses with V and update the worklist. 730 static void ReplaceUsesOfWith(Instruction *I, Value *V, 731 std::vector<Instruction*> &Worklist, 732 Loop *L, LPPassManager *LPM) { 733 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 734 735 // Add uses to the worklist, which may be dead now. 736 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 737 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 738 Worklist.push_back(Use); 739 740 // Add users to the worklist which may be simplified now. 741 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 742 UI != E; ++UI) 743 Worklist.push_back(cast<Instruction>(*UI)); 744 LPM->deleteSimpleAnalysisValue(I, L); 745 RemoveFromWorklist(I, Worklist); 746 I->replaceAllUsesWith(V); 747 I->eraseFromParent(); 748 ++NumSimplify; 749 } 750 751 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 752 /// information, and remove any dead successors it has. 753 /// 754 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 755 std::vector<Instruction*> &Worklist, 756 Loop *L) { 757 if (pred_begin(BB) != pred_end(BB)) { 758 // This block isn't dead, since an edge to BB was just removed, see if there 759 // are any easy simplifications we can do now. 760 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 761 // If it has one pred, fold phi nodes in BB. 762 while (isa<PHINode>(BB->begin())) 763 ReplaceUsesOfWith(BB->begin(), 764 cast<PHINode>(BB->begin())->getIncomingValue(0), 765 Worklist, L, LPM); 766 767 // If this is the header of a loop and the only pred is the latch, we now 768 // have an unreachable loop. 769 if (Loop *L = LI->getLoopFor(BB)) 770 if (loopHeader == BB && L->contains(Pred)) { 771 // Remove the branch from the latch to the header block, this makes 772 // the header dead, which will make the latch dead (because the header 773 // dominates the latch). 774 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 775 Pred->getTerminator()->eraseFromParent(); 776 new UnreachableInst(BB->getContext(), Pred); 777 778 // The loop is now broken, remove it from LI. 779 RemoveLoopFromHierarchy(L); 780 781 // Reprocess the header, which now IS dead. 782 RemoveBlockIfDead(BB, Worklist, L); 783 return; 784 } 785 786 // If pred ends in a uncond branch, add uncond branch to worklist so that 787 // the two blocks will get merged. 788 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 789 if (BI->isUnconditional()) 790 Worklist.push_back(BI); 791 } 792 return; 793 } 794 795 DEBUG(dbgs() << "Nuking dead block: " << *BB); 796 797 // Remove the instructions in the basic block from the worklist. 798 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 799 RemoveFromWorklist(I, Worklist); 800 801 // Anything that uses the instructions in this basic block should have their 802 // uses replaced with undefs. 803 // If I is not void type then replaceAllUsesWith undef. 804 // This allows ValueHandlers and custom metadata to adjust itself. 805 if (!I->getType()->isVoidTy()) 806 I->replaceAllUsesWith(UndefValue::get(I->getType())); 807 } 808 809 // If this is the edge to the header block for a loop, remove the loop and 810 // promote all subloops. 811 if (Loop *BBLoop = LI->getLoopFor(BB)) { 812 if (BBLoop->getLoopLatch() == BB) { 813 RemoveLoopFromHierarchy(BBLoop); 814 if (currentLoop == BBLoop) { 815 currentLoop = 0; 816 redoLoop = false; 817 } 818 } 819 } 820 821 // Remove the block from the loop info, which removes it from any loops it 822 // was in. 823 LI->removeBlock(BB); 824 825 826 // Remove phi node entries in successors for this block. 827 TerminatorInst *TI = BB->getTerminator(); 828 SmallVector<BasicBlock*, 4> Succs; 829 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 830 Succs.push_back(TI->getSuccessor(i)); 831 TI->getSuccessor(i)->removePredecessor(BB); 832 } 833 834 // Unique the successors, remove anything with multiple uses. 835 array_pod_sort(Succs.begin(), Succs.end()); 836 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 837 838 // Remove the basic block, including all of the instructions contained in it. 839 LPM->deleteSimpleAnalysisValue(BB, L); 840 BB->eraseFromParent(); 841 // Remove successor blocks here that are not dead, so that we know we only 842 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 843 // then getting removed before we revisit them, which is badness. 844 // 845 for (unsigned i = 0; i != Succs.size(); ++i) 846 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 847 // One exception is loop headers. If this block was the preheader for a 848 // loop, then we DO want to visit the loop so the loop gets deleted. 849 // We know that if the successor is a loop header, that this loop had to 850 // be the preheader: the case where this was the latch block was handled 851 // above and headers can only have two predecessors. 852 if (!LI->isLoopHeader(Succs[i])) { 853 Succs.erase(Succs.begin()+i); 854 --i; 855 } 856 } 857 858 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 859 RemoveBlockIfDead(Succs[i], Worklist, L); 860 } 861 862 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has 863 /// become unwrapped, either because the backedge was deleted, or because the 864 /// edge into the header was removed. If the edge into the header from the 865 /// latch block was removed, the loop is unwrapped but subloops are still alive, 866 /// so they just reparent loops. If the loops are actually dead, they will be 867 /// removed later. 868 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 869 LPM->deleteLoopFromQueue(L); 870 RemoveLoopFromWorklist(L); 871 } 872 873 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 874 // the value specified by Val in the specified loop, or we know it does NOT have 875 // that value. Rewrite any uses of LIC or of properties correlated to it. 876 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 877 Constant *Val, 878 bool IsEqual) { 879 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 880 881 // FIXME: Support correlated properties, like: 882 // for (...) 883 // if (li1 < li2) 884 // ... 885 // if (li1 > li2) 886 // ... 887 888 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 889 // selects, switches. 890 std::vector<Instruction*> Worklist; 891 LLVMContext &Context = Val->getContext(); 892 893 894 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 895 // in the loop with the appropriate one directly. 896 if (IsEqual || (isa<ConstantInt>(Val) && 897 Val->getType()->isIntegerTy(1))) { 898 Value *Replacement; 899 if (IsEqual) 900 Replacement = Val; 901 else 902 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 903 !cast<ConstantInt>(Val)->getZExtValue()); 904 905 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 906 UI != E; ++UI) { 907 Instruction *U = dyn_cast<Instruction>(*UI); 908 if (!U || !L->contains(U)) 909 continue; 910 U->replaceUsesOfWith(LIC, Replacement); 911 Worklist.push_back(U); 912 } 913 SimplifyCode(Worklist, L); 914 return; 915 } 916 917 // Otherwise, we don't know the precise value of LIC, but we do know that it 918 // is certainly NOT "Val". As such, simplify any uses in the loop that we 919 // can. This case occurs when we unswitch switch statements. 920 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); 921 UI != E; ++UI) { 922 Instruction *U = dyn_cast<Instruction>(*UI); 923 if (!U || !L->contains(U)) 924 continue; 925 926 Worklist.push_back(U); 927 928 // TODO: We could do other simplifications, for example, turning 929 // 'icmp eq LIC, Val' -> false. 930 931 // If we know that LIC is not Val, use this info to simplify code. 932 SwitchInst *SI = dyn_cast<SwitchInst>(U); 933 if (SI == 0 || !isa<ConstantInt>(Val)) continue; 934 935 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 936 if (DeadCase == 0) continue; // Default case is live for multiple values. 937 938 // Found a dead case value. Don't remove PHI nodes in the 939 // successor if they become single-entry, those PHI nodes may 940 // be in the Users list. 941 942 BasicBlock *Switch = SI->getParent(); 943 BasicBlock *SISucc = SI->getSuccessor(DeadCase); 944 BasicBlock *Latch = L->getLoopLatch(); 945 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical. 946 // If the DeadCase successor dominates the loop latch, then the 947 // transformation isn't safe since it will delete the sole predecessor edge 948 // to the latch. 949 if (Latch && DT->dominates(SISucc, Latch)) 950 continue; 951 952 // FIXME: This is a hack. We need to keep the successor around 953 // and hooked up so as to preserve the loop structure, because 954 // trying to update it is complicated. So instead we preserve the 955 // loop structure and put the block on a dead code path. 956 SplitEdge(Switch, SISucc, this); 957 // Compute the successors instead of relying on the return value 958 // of SplitEdge, since it may have split the switch successor 959 // after PHI nodes. 960 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase); 961 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 962 // Create an "unreachable" destination. 963 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 964 Switch->getParent(), 965 OldSISucc); 966 new UnreachableInst(Context, Abort); 967 // Force the new case destination to branch to the "unreachable" 968 // block while maintaining a (dead) CFG edge to the old block. 969 NewSISucc->getTerminator()->eraseFromParent(); 970 BranchInst::Create(Abort, OldSISucc, 971 ConstantInt::getTrue(Context), NewSISucc); 972 // Release the PHI operands for this edge. 973 for (BasicBlock::iterator II = NewSISucc->begin(); 974 PHINode *PN = dyn_cast<PHINode>(II); ++II) 975 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 976 UndefValue::get(PN->getType())); 977 // Tell the domtree about the new block. We don't fully update the 978 // domtree here -- instead we force it to do a full recomputation 979 // after the pass is complete -- but we do need to inform it of 980 // new blocks. 981 if (DT) 982 DT->addNewBlock(Abort, NewSISucc); 983 } 984 985 SimplifyCode(Worklist, L); 986 } 987 988 /// SimplifyCode - Okay, now that we have simplified some instructions in the 989 /// loop, walk over it and constant prop, dce, and fold control flow where 990 /// possible. Note that this is effectively a very simple loop-structure-aware 991 /// optimizer. During processing of this loop, L could very well be deleted, so 992 /// it must not be used. 993 /// 994 /// FIXME: When the loop optimizer is more mature, separate this out to a new 995 /// pass. 996 /// 997 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 998 while (!Worklist.empty()) { 999 Instruction *I = Worklist.back(); 1000 Worklist.pop_back(); 1001 1002 // Simple DCE. 1003 if (isInstructionTriviallyDead(I)) { 1004 DEBUG(dbgs() << "Remove dead instruction '" << *I); 1005 1006 // Add uses to the worklist, which may be dead now. 1007 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1008 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1009 Worklist.push_back(Use); 1010 LPM->deleteSimpleAnalysisValue(I, L); 1011 RemoveFromWorklist(I, Worklist); 1012 I->eraseFromParent(); 1013 ++NumSimplify; 1014 continue; 1015 } 1016 1017 // See if instruction simplification can hack this up. This is common for 1018 // things like "select false, X, Y" after unswitching made the condition be 1019 // 'false'. 1020 if (Value *V = SimplifyInstruction(I, 0, DT)) 1021 if (LI->replacementPreservesLCSSAForm(I, V)) { 1022 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 1023 continue; 1024 } 1025 1026 // Special case hacks that appear commonly in unswitched code. 1027 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1028 if (BI->isUnconditional()) { 1029 // If BI's parent is the only pred of the successor, fold the two blocks 1030 // together. 1031 BasicBlock *Pred = BI->getParent(); 1032 BasicBlock *Succ = BI->getSuccessor(0); 1033 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1034 if (!SinglePred) continue; // Nothing to do. 1035 assert(SinglePred == Pred && "CFG broken"); 1036 1037 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1038 << Succ->getName() << "\n"); 1039 1040 // Resolve any single entry PHI nodes in Succ. 1041 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1042 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1043 1044 // If Succ has any successors with PHI nodes, update them to have 1045 // entries coming from Pred instead of Succ. 1046 Succ->replaceAllUsesWith(Pred); 1047 1048 // Move all of the successor contents from Succ to Pred. 1049 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1050 Succ->end()); 1051 LPM->deleteSimpleAnalysisValue(BI, L); 1052 BI->eraseFromParent(); 1053 RemoveFromWorklist(BI, Worklist); 1054 1055 // Remove Succ from the loop tree. 1056 LI->removeBlock(Succ); 1057 LPM->deleteSimpleAnalysisValue(Succ, L); 1058 Succ->eraseFromParent(); 1059 ++NumSimplify; 1060 continue; 1061 } 1062 1063 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1064 // Conditional branch. Turn it into an unconditional branch, then 1065 // remove dead blocks. 1066 continue; // FIXME: Enable. 1067 1068 DEBUG(dbgs() << "Folded branch: " << *BI); 1069 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1070 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1071 DeadSucc->removePredecessor(BI->getParent(), true); 1072 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1073 LPM->deleteSimpleAnalysisValue(BI, L); 1074 BI->eraseFromParent(); 1075 RemoveFromWorklist(BI, Worklist); 1076 ++NumSimplify; 1077 1078 RemoveBlockIfDead(DeadSucc, Worklist, L); 1079 } 1080 continue; 1081 } 1082 } 1083 } 1084