LoopUnswitch.cpp revision 205942a4a55d568e93480fc22d25cc7dac525fb7
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/ConstantFolding.h" 36#include "llvm/Analysis/InlineCost.h" 37#include "llvm/Analysis/InstructionSimplify.h" 38#include "llvm/Analysis/LoopInfo.h" 39#include "llvm/Analysis/LoopPass.h" 40#include "llvm/Analysis/Dominators.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> 52using namespace llvm; 53 54STATISTIC(NumBranches, "Number of branches unswitched"); 55STATISTIC(NumSwitches, "Number of switches unswitched"); 56STATISTIC(NumSelects , "Number of selects unswitched"); 57STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 58STATISTIC(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. 62static cl::opt<unsigned> 63Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 64 cl::init(50), cl::Hidden); 65 66namespace { 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 } 115 116 private: 117 118 virtual void releaseMemory() { 119 UnswitchedVals.clear(); 120 } 121 122 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 123 /// remove it. 124 void RemoveLoopFromWorklist(Loop *L) { 125 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 126 LoopProcessWorklist.end(), L); 127 if (I != LoopProcessWorklist.end()) 128 LoopProcessWorklist.erase(I); 129 } 130 131 void initLoopData() { 132 loopHeader = currentLoop->getHeader(); 133 loopPreheader = currentLoop->getLoopPreheader(); 134 } 135 136 /// Split all of the edges from inside the loop to their exit blocks. 137 /// Update the appropriate Phi nodes as we do so. 138 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks); 139 140 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); 141 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 142 BasicBlock *ExitBlock); 143 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 144 145 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 146 Constant *Val, bool isEqual); 147 148 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 149 BasicBlock *TrueDest, 150 BasicBlock *FalseDest, 151 Instruction *InsertPt); 152 153 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 154 void RemoveBlockIfDead(BasicBlock *BB, 155 std::vector<Instruction*> &Worklist, Loop *l); 156 void RemoveLoopFromHierarchy(Loop *L); 157 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, 158 BasicBlock **LoopExit = 0); 159 160 }; 161} 162char LoopUnswitch::ID = 0; 163INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", 164 false, false) 165INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 166INITIALIZE_PASS_DEPENDENCY(LoopInfo) 167INITIALIZE_PASS_DEPENDENCY(LCSSA) 168INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", 169 false, false) 170 171Pass *llvm::createLoopUnswitchPass(bool Os) { 172 return new LoopUnswitch(Os); 173} 174 175/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 176/// invariant in the loop, or has an invariant piece, return the invariant. 177/// Otherwise, return null. 178static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 179 // We can never unswitch on vector conditions. 180 if (Cond->getType()->isVectorTy()) 181 return 0; 182 183 // Constants should be folded, not unswitched on! 184 if (isa<Constant>(Cond)) return 0; 185 186 // TODO: Handle: br (VARIANT|INVARIANT). 187 188 // Hoist simple values out. 189 if (L->makeLoopInvariant(Cond, Changed)) 190 return Cond; 191 192 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 193 if (BO->getOpcode() == Instruction::And || 194 BO->getOpcode() == Instruction::Or) { 195 // If either the left or right side is invariant, we can unswitch on this, 196 // which will cause the branch to go away in one loop and the condition to 197 // simplify in the other one. 198 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 199 return LHS; 200 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 201 return RHS; 202 } 203 204 return 0; 205} 206 207bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 208 LI = &getAnalysis<LoopInfo>(); 209 LPM = &LPM_Ref; 210 DT = getAnalysisIfAvailable<DominatorTree>(); 211 currentLoop = L; 212 Function *F = currentLoop->getHeader()->getParent(); 213 bool Changed = false; 214 do { 215 assert(currentLoop->isLCSSAForm(*DT)); 216 redoLoop = false; 217 Changed |= processCurrentLoop(); 218 } while(redoLoop); 219 220 if (Changed) { 221 // FIXME: Reconstruct dom info, because it is not preserved properly. 222 if (DT) 223 DT->runOnFunction(*F); 224 } 225 return Changed; 226} 227 228/// processCurrentLoop - Do actual work and unswitch loop if possible 229/// and profitable. 230bool LoopUnswitch::processCurrentLoop() { 231 bool Changed = false; 232 LLVMContext &Context = currentLoop->getHeader()->getContext(); 233 234 // Loop over all of the basic blocks in the loop. If we find an interior 235 // block that is branching on a loop-invariant condition, we can unswitch this 236 // loop. 237 for (Loop::block_iterator I = currentLoop->block_begin(), 238 E = currentLoop->block_end(); I != E; ++I) { 239 TerminatorInst *TI = (*I)->getTerminator(); 240 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 241 // If this isn't branching on an invariant condition, we can't unswitch 242 // it. 243 if (BI->isConditional()) { 244 // See if this, or some part of it, is loop invariant. If so, we can 245 // unswitch on it if we desire. 246 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), 247 currentLoop, Changed); 248 if (LoopCond && UnswitchIfProfitable(LoopCond, 249 ConstantInt::getTrue(Context))) { 250 ++NumBranches; 251 return true; 252 } 253 } 254 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 255 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 256 currentLoop, Changed); 257 if (LoopCond && SI->getNumCases() > 1) { 258 // Find a value to unswitch on: 259 // FIXME: this should chose the most expensive case! 260 Constant *UnswitchVal = SI->getCaseValue(1); 261 // Do not process same value again and again. 262 if (!UnswitchedVals.insert(UnswitchVal)) 263 continue; 264 265 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { 266 ++NumSwitches; 267 return true; 268 } 269 } 270 } 271 272 // Scan the instructions to check for unswitchable values. 273 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 274 BBI != E; ++BBI) 275 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 276 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), 277 currentLoop, Changed); 278 if (LoopCond && UnswitchIfProfitable(LoopCond, 279 ConstantInt::getTrue(Context))) { 280 ++NumSelects; 281 return true; 282 } 283 } 284 } 285 return Changed; 286} 287 288/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the 289/// loop with no side effects (including infinite loops). 290/// 291/// If true, we return true and set ExitBB to the block we 292/// exit through. 293/// 294static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 295 BasicBlock *&ExitBB, 296 std::set<BasicBlock*> &Visited) { 297 if (!Visited.insert(BB).second) { 298 // Already visited. Without more analysis, this could indicate an infinte loop. 299 return false; 300 } else if (!L->contains(BB)) { 301 // Otherwise, this is a loop exit, this is fine so long as this is the 302 // first exit. 303 if (ExitBB != 0) return false; 304 ExitBB = BB; 305 return true; 306 } 307 308 // Otherwise, this is an unvisited intra-loop node. Check all successors. 309 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 310 // Check to see if the successor is a trivial loop exit. 311 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 312 return false; 313 } 314 315 // Okay, everything after this looks good, check to make sure that this block 316 // doesn't include any side effects. 317 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 318 if (I->mayHaveSideEffects()) 319 return false; 320 321 return true; 322} 323 324/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 325/// leads to an exit from the specified loop, and has no side-effects in the 326/// process. If so, return the block that is exited to, otherwise return null. 327static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 328 std::set<BasicBlock*> Visited; 329 Visited.insert(L->getHeader()); // Branches to header make infinite loops. 330 BasicBlock *ExitBB = 0; 331 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 332 return ExitBB; 333 return 0; 334} 335 336/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 337/// trivial: that is, that the condition controls whether or not the loop does 338/// anything at all. If this is a trivial condition, unswitching produces no 339/// code duplications (equivalently, it produces a simpler loop and a new empty 340/// loop, which gets deleted). 341/// 342/// If this is a trivial condition, return true, otherwise return false. When 343/// returning true, this sets Cond and Val to the condition that controls the 344/// trivial condition: when Cond dynamically equals Val, the loop is known to 345/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 346/// Cond == Val. 347/// 348bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 349 BasicBlock **LoopExit) { 350 BasicBlock *Header = currentLoop->getHeader(); 351 TerminatorInst *HeaderTerm = Header->getTerminator(); 352 LLVMContext &Context = Header->getContext(); 353 354 BasicBlock *LoopExitBB = 0; 355 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 356 // If the header block doesn't end with a conditional branch on Cond, we 357 // can't handle it. 358 if (!BI->isConditional() || BI->getCondition() != Cond) 359 return false; 360 361 // Check to see if a successor of the branch is guaranteed to 362 // exit through a unique exit block without having any 363 // side-effects. If so, determine the value of Cond that causes it to do 364 // this. 365 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 366 BI->getSuccessor(0)))) { 367 if (Val) *Val = ConstantInt::getTrue(Context); 368 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 369 BI->getSuccessor(1)))) { 370 if (Val) *Val = ConstantInt::getFalse(Context); 371 } 372 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 373 // If this isn't a switch on Cond, we can't handle it. 374 if (SI->getCondition() != Cond) return false; 375 376 // Check to see if a successor of the switch is guaranteed to go to the 377 // latch block or exit through a one exit block without having any 378 // side-effects. If so, determine the value of Cond that causes it to do 379 // this. Note that we can't trivially unswitch on the default case. 380 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 381 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, 382 SI->getSuccessor(i)))) { 383 // Okay, we found a trivial case, remember the value that is trivial. 384 if (Val) *Val = SI->getCaseValue(i); 385 break; 386 } 387 } 388 389 // If we didn't find a single unique LoopExit block, or if the loop exit block 390 // contains phi nodes, this isn't trivial. 391 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 392 return false; // Can't handle this. 393 394 if (LoopExit) *LoopExit = LoopExitBB; 395 396 // We already know that nothing uses any scalar values defined inside of this 397 // loop. As such, we just have to check to see if this loop will execute any 398 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 399 // part of the loop that the code *would* execute. We already checked the 400 // tail, check the header now. 401 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 402 if (I->mayHaveSideEffects()) 403 return false; 404 return true; 405} 406 407/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when 408/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 409/// unswitch the loop, reprocess the pieces, then return true. 410bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { 411 412 initLoopData(); 413 414 // If LoopSimplify was unable to form a preheader, don't do any unswitching. 415 if (!loopPreheader) 416 return false; 417 418 Function *F = loopHeader->getParent(); 419 420 Constant *CondVal = 0; 421 BasicBlock *ExitBlock = 0; 422 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { 423 // If the condition is trivial, always unswitch. There is no code growth 424 // for this case. 425 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); 426 return true; 427 } 428 429 // Check to see if it would be profitable to unswitch current loop. 430 431 // Do not do non-trivial unswitch while optimizing for size. 432 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) 433 return false; 434 435 // FIXME: This is overly conservative because it does not take into 436 // consideration code simplification opportunities and code that can 437 // be shared by the resultant unswitched loops. 438 CodeMetrics Metrics; 439 for (Loop::block_iterator I = currentLoop->block_begin(), 440 E = currentLoop->block_end(); 441 I != E; ++I) 442 Metrics.analyzeBasicBlock(*I); 443 444 // Limit the number of instructions to avoid causing significant code 445 // expansion, and the number of basic blocks, to avoid loops with 446 // large numbers of branches which cause loop unswitching to go crazy. 447 // This is a very ad-hoc heuristic. 448 if (Metrics.NumInsts > Threshold || 449 Metrics.NumBlocks * 5 > Threshold || 450 Metrics.containsIndirectBr || Metrics.isRecursive) { 451 DEBUG(dbgs() << "NOT unswitching loop %" 452 << currentLoop->getHeader()->getName() << ", cost too high: " 453 << currentLoop->getBlocks().size() << "\n"); 454 return false; 455 } 456 457 UnswitchNontrivialCondition(LoopCond, Val, currentLoop); 458 return true; 459} 460 461// RemapInstruction - Convert the instruction operands from referencing the 462// current values into those specified by VMap. 463// 464static inline void RemapInstruction(Instruction *I, 465 ValueToValueMapTy &VMap) { 466 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 467 Value *Op = I->getOperand(op); 468 ValueToValueMapTy::iterator It = VMap.find(Op); 469 if (It != VMap.end()) Op = It->second; 470 I->setOperand(op, Op); 471 } 472} 473 474/// CloneLoop - Recursively clone the specified loop and all of its children, 475/// mapping the blocks with the specified map. 476static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, 477 LoopInfo *LI, LPPassManager *LPM) { 478 Loop *New = new Loop(); 479 LPM->insertLoop(New, PL); 480 481 // Add all of the blocks in L to the new loop. 482 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 483 I != E; ++I) 484 if (LI->getLoopFor(*I) == L) 485 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 486 487 // Add all of the subloops to the new loop. 488 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 489 CloneLoop(*I, New, VM, LI, LPM); 490 491 return New; 492} 493 494/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 495/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 496/// code immediately before InsertPt. 497void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 498 BasicBlock *TrueDest, 499 BasicBlock *FalseDest, 500 Instruction *InsertPt) { 501 // Insert a conditional branch on LIC to the two preheaders. The original 502 // code is the true version and the new code is the false version. 503 Value *BranchVal = LIC; 504 if (!isa<ConstantInt>(Val) || 505 Val->getType() != Type::getInt1Ty(LIC->getContext())) 506 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); 507 else if (Val != ConstantInt::getTrue(Val->getContext())) 508 // We want to enter the new loop when the condition is true. 509 std::swap(TrueDest, FalseDest); 510 511 // Insert the new branch. 512 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 513 514 // If either edge is critical, split it. This helps preserve LoopSimplify 515 // form for enclosing loops. 516 SplitCriticalEdge(BI, 0, this); 517 SplitCriticalEdge(BI, 1, this); 518} 519 520/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 521/// condition in it (a cond branch from its header block to its latch block, 522/// where the path through the loop that doesn't execute its body has no 523/// side-effects), unswitch it. This doesn't involve any code duplication, just 524/// moving the conditional branch outside of the loop and updating loop info. 525void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 526 Constant *Val, 527 BasicBlock *ExitBlock) { 528 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" 529 << loopHeader->getName() << " [" << L->getBlocks().size() 530 << " blocks] in Function " << L->getHeader()->getParent()->getName() 531 << " on cond: " << *Val << " == " << *Cond << "\n"); 532 533 // First step, split the preheader, so that we know that there is a safe place 534 // to insert the conditional branch. We will change loopPreheader to have a 535 // conditional branch on Cond. 536 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); 537 538 // Now that we have a place to insert the conditional branch, create a place 539 // to branch to: this is the exit block out of the loop that we should 540 // short-circuit to. 541 542 // Split this block now, so that the loop maintains its exit block, and so 543 // that the jump from the preheader can execute the contents of the exit block 544 // without actually branching to it (the exit block should be dominated by the 545 // loop header, not the preheader). 546 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 547 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 548 549 // Okay, now we have a position to branch from and a position to branch to, 550 // insert the new conditional branch. 551 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 552 loopPreheader->getTerminator()); 553 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); 554 loopPreheader->getTerminator()->eraseFromParent(); 555 556 // We need to reprocess this loop, it could be unswitched again. 557 redoLoop = true; 558 559 // Now that we know that the loop is never entered when this condition is a 560 // particular value, rewrite the loop with this info. We know that this will 561 // at least eliminate the old branch. 562 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 563 ++NumTrivial; 564} 565 566/// SplitExitEdges - Split all of the edges from inside the loop to their exit 567/// blocks. Update the appropriate Phi nodes as we do so. 568void LoopUnswitch::SplitExitEdges(Loop *L, 569 const SmallVector<BasicBlock *, 8> &ExitBlocks){ 570 571 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 572 BasicBlock *ExitBlock = ExitBlocks[i]; 573 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock), 574 pred_end(ExitBlock)); 575 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), 576 ".us-lcssa", this); 577 } 578} 579 580/// UnswitchNontrivialCondition - We determined that the loop is profitable 581/// to unswitch when LIC equal Val. Split it into loop versions and test the 582/// condition outside of either loop. Return the loops created as Out1/Out2. 583void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 584 Loop *L) { 585 Function *F = loopHeader->getParent(); 586 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" 587 << loopHeader->getName() << " [" << L->getBlocks().size() 588 << " blocks] in Function " << F->getName() 589 << " when '" << *Val << "' == " << *LIC << "\n"); 590 591 LoopBlocks.clear(); 592 NewBlocks.clear(); 593 594 // First step, split the preheader and exit blocks, and add these blocks to 595 // the LoopBlocks list. 596 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); 597 LoopBlocks.push_back(NewPreheader); 598 599 // We want the loop to come after the preheader, but before the exit blocks. 600 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 601 602 SmallVector<BasicBlock*, 8> ExitBlocks; 603 L->getUniqueExitBlocks(ExitBlocks); 604 605 // Split all of the edges from inside the loop to their exit blocks. Update 606 // the appropriate Phi nodes as we do so. 607 SplitExitEdges(L, ExitBlocks); 608 609 // The exit blocks may have been changed due to edge splitting, recompute. 610 ExitBlocks.clear(); 611 L->getUniqueExitBlocks(ExitBlocks); 612 613 // Add exit blocks to the loop blocks. 614 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 615 616 // Next step, clone all of the basic blocks that make up the loop (including 617 // the loop preheader and exit blocks), keeping track of the mapping between 618 // the instructions and blocks. 619 NewBlocks.reserve(LoopBlocks.size()); 620 ValueToValueMapTy VMap; 621 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 622 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); 623 NewBlocks.push_back(NewBB); 624 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. 625 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); 626 } 627 628 // Splice the newly inserted blocks into the function right before the 629 // original preheader. 630 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), 631 NewBlocks[0], F->end()); 632 633 // Now we create the new Loop object for the versioned loop. 634 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); 635 Loop *ParentLoop = L->getParentLoop(); 636 if (ParentLoop) { 637 // Make sure to add the cloned preheader and exit blocks to the parent loop 638 // as well. 639 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 640 } 641 642 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 643 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]); 644 // The new exit block should be in the same loop as the old one. 645 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 646 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 647 648 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 649 "Exit block should have been split to have one successor!"); 650 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 651 652 // If the successor of the exit block had PHI nodes, add an entry for 653 // NewExit. 654 PHINode *PN; 655 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) { 656 PN = cast<PHINode>(I); 657 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 658 ValueToValueMapTy::iterator It = VMap.find(V); 659 if (It != VMap.end()) V = It->second; 660 PN->addIncoming(V, NewExit); 661 } 662 } 663 664 // Rewrite the code to refer to itself. 665 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 666 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 667 E = NewBlocks[i]->end(); I != E; ++I) 668 RemapInstruction(I, VMap); 669 670 // Rewrite the original preheader to select between versions of the loop. 671 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator()); 672 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 673 "Preheader splitting did not work correctly!"); 674 675 // Emit the new branch that selects between the two versions of this loop. 676 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 677 LPM->deleteSimpleAnalysisValue(OldBR, L); 678 OldBR->eraseFromParent(); 679 680 LoopProcessWorklist.push_back(NewLoop); 681 redoLoop = true; 682 683 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody 684 // deletes the instruction (for example by simplifying a PHI that feeds into 685 // the condition that we're unswitching on), we don't rewrite the second 686 // iteration. 687 WeakVH LICHandle(LIC); 688 689 // Now we rewrite the original code to know that the condition is true and the 690 // new code to know that the condition is false. 691 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); 692 693 // It's possible that simplifying one loop could cause the other to be 694 // changed to another value or a constant. If its a constant, don't simplify 695 // it. 696 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop && 697 LICHandle && !isa<Constant>(LICHandle)) 698 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); 699} 700 701/// RemoveFromWorklist - Remove all instances of I from the worklist vector 702/// specified. 703static void RemoveFromWorklist(Instruction *I, 704 std::vector<Instruction*> &Worklist) { 705 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 706 Worklist.end(), I); 707 while (WI != Worklist.end()) { 708 unsigned Offset = WI-Worklist.begin(); 709 Worklist.erase(WI); 710 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 711 } 712} 713 714/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 715/// program, replacing all uses with V and update the worklist. 716static void ReplaceUsesOfWith(Instruction *I, Value *V, 717 std::vector<Instruction*> &Worklist, 718 Loop *L, LPPassManager *LPM) { 719 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); 720 721 // Add uses to the worklist, which may be dead now. 722 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 723 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 724 Worklist.push_back(Use); 725 726 // Add users to the worklist which may be simplified now. 727 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 728 UI != E; ++UI) 729 Worklist.push_back(cast<Instruction>(*UI)); 730 LPM->deleteSimpleAnalysisValue(I, L); 731 RemoveFromWorklist(I, Worklist); 732 I->replaceAllUsesWith(V); 733 I->eraseFromParent(); 734 ++NumSimplify; 735} 736 737/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 738/// information, and remove any dead successors it has. 739/// 740void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 741 std::vector<Instruction*> &Worklist, 742 Loop *L) { 743 if (pred_begin(BB) != pred_end(BB)) { 744 // This block isn't dead, since an edge to BB was just removed, see if there 745 // are any easy simplifications we can do now. 746 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 747 // If it has one pred, fold phi nodes in BB. 748 while (isa<PHINode>(BB->begin())) 749 ReplaceUsesOfWith(BB->begin(), 750 cast<PHINode>(BB->begin())->getIncomingValue(0), 751 Worklist, L, LPM); 752 753 // If this is the header of a loop and the only pred is the latch, we now 754 // have an unreachable loop. 755 if (Loop *L = LI->getLoopFor(BB)) 756 if (loopHeader == BB && L->contains(Pred)) { 757 // Remove the branch from the latch to the header block, this makes 758 // the header dead, which will make the latch dead (because the header 759 // dominates the latch). 760 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 761 Pred->getTerminator()->eraseFromParent(); 762 new UnreachableInst(BB->getContext(), Pred); 763 764 // The loop is now broken, remove it from LI. 765 RemoveLoopFromHierarchy(L); 766 767 // Reprocess the header, which now IS dead. 768 RemoveBlockIfDead(BB, Worklist, L); 769 return; 770 } 771 772 // If pred ends in a uncond branch, add uncond branch to worklist so that 773 // the two blocks will get merged. 774 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 775 if (BI->isUnconditional()) 776 Worklist.push_back(BI); 777 } 778 return; 779 } 780 781 DEBUG(dbgs() << "Nuking dead block: " << *BB); 782 783 // Remove the instructions in the basic block from the worklist. 784 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 785 RemoveFromWorklist(I, Worklist); 786 787 // Anything that uses the instructions in this basic block should have their 788 // uses replaced with undefs. 789 // If I is not void type then replaceAllUsesWith undef. 790 // This allows ValueHandlers and custom metadata to adjust itself. 791 if (!I->getType()->isVoidTy()) 792 I->replaceAllUsesWith(UndefValue::get(I->getType())); 793 } 794 795 // If this is the edge to the header block for a loop, remove the loop and 796 // promote all subloops. 797 if (Loop *BBLoop = LI->getLoopFor(BB)) { 798 if (BBLoop->getLoopLatch() == BB) 799 RemoveLoopFromHierarchy(BBLoop); 800 } 801 802 // Remove the block from the loop info, which removes it from any loops it 803 // was in. 804 LI->removeBlock(BB); 805 806 807 // Remove phi node entries in successors for this block. 808 TerminatorInst *TI = BB->getTerminator(); 809 SmallVector<BasicBlock*, 4> Succs; 810 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 811 Succs.push_back(TI->getSuccessor(i)); 812 TI->getSuccessor(i)->removePredecessor(BB); 813 } 814 815 // Unique the successors, remove anything with multiple uses. 816 array_pod_sort(Succs.begin(), Succs.end()); 817 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 818 819 // Remove the basic block, including all of the instructions contained in it. 820 LPM->deleteSimpleAnalysisValue(BB, L); 821 BB->eraseFromParent(); 822 // Remove successor blocks here that are not dead, so that we know we only 823 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 824 // then getting removed before we revisit them, which is badness. 825 // 826 for (unsigned i = 0; i != Succs.size(); ++i) 827 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 828 // One exception is loop headers. If this block was the preheader for a 829 // loop, then we DO want to visit the loop so the loop gets deleted. 830 // We know that if the successor is a loop header, that this loop had to 831 // be the preheader: the case where this was the latch block was handled 832 // above and headers can only have two predecessors. 833 if (!LI->isLoopHeader(Succs[i])) { 834 Succs.erase(Succs.begin()+i); 835 --i; 836 } 837 } 838 839 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 840 RemoveBlockIfDead(Succs[i], Worklist, L); 841} 842 843/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 844/// become unwrapped, either because the backedge was deleted, or because the 845/// edge into the header was removed. If the edge into the header from the 846/// latch block was removed, the loop is unwrapped but subloops are still alive, 847/// so they just reparent loops. If the loops are actually dead, they will be 848/// removed later. 849void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 850 LPM->deleteLoopFromQueue(L); 851 RemoveLoopFromWorklist(L); 852} 853 854// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 855// the value specified by Val in the specified loop, or we know it does NOT have 856// that value. Rewrite any uses of LIC or of properties correlated to it. 857void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 858 Constant *Val, 859 bool IsEqual) { 860 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 861 862 // FIXME: Support correlated properties, like: 863 // for (...) 864 // if (li1 < li2) 865 // ... 866 // if (li1 > li2) 867 // ... 868 869 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 870 // selects, switches. 871 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 872 std::vector<Instruction*> Worklist; 873 LLVMContext &Context = Val->getContext(); 874 875 876 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 877 // in the loop with the appropriate one directly. 878 if (IsEqual || (isa<ConstantInt>(Val) && 879 Val->getType()->isIntegerTy(1))) { 880 Value *Replacement; 881 if (IsEqual) 882 Replacement = Val; 883 else 884 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), 885 !cast<ConstantInt>(Val)->getZExtValue()); 886 887 for (unsigned i = 0, e = Users.size(); i != e; ++i) 888 if (Instruction *U = cast<Instruction>(Users[i])) { 889 if (!L->contains(U)) 890 continue; 891 U->replaceUsesOfWith(LIC, Replacement); 892 Worklist.push_back(U); 893 } 894 SimplifyCode(Worklist, L); 895 return; 896 } 897 898 // Otherwise, we don't know the precise value of LIC, but we do know that it 899 // is certainly NOT "Val". As such, simplify any uses in the loop that we 900 // can. This case occurs when we unswitch switch statements. 901 for (unsigned i = 0, e = Users.size(); i != e; ++i) { 902 Instruction *U = cast<Instruction>(Users[i]); 903 if (!L->contains(U)) 904 continue; 905 906 Worklist.push_back(U); 907 908 // TODO: We could do other simplifications, for example, turning 909 // 'icmp eq LIC, Val' -> false. 910 911 // If we know that LIC is not Val, use this info to simplify code. 912 SwitchInst *SI = dyn_cast<SwitchInst>(U); 913 if (SI == 0 || !isa<ConstantInt>(Val)) continue; 914 915 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val)); 916 if (DeadCase == 0) continue; // Default case is live for multiple values. 917 918 // Found a dead case value. Don't remove PHI nodes in the 919 // successor if they become single-entry, those PHI nodes may 920 // be in the Users list. 921 922 // FIXME: This is a hack. We need to keep the successor around 923 // and hooked up so as to preserve the loop structure, because 924 // trying to update it is complicated. So instead we preserve the 925 // loop structure and put the block on a dead code path. 926 BasicBlock *Switch = SI->getParent(); 927 SplitEdge(Switch, SI->getSuccessor(DeadCase), this); 928 // Compute the successors instead of relying on the return value 929 // of SplitEdge, since it may have split the switch successor 930 // after PHI nodes. 931 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase); 932 BasicBlock *OldSISucc = *succ_begin(NewSISucc); 933 // Create an "unreachable" destination. 934 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", 935 Switch->getParent(), 936 OldSISucc); 937 new UnreachableInst(Context, Abort); 938 // Force the new case destination to branch to the "unreachable" 939 // block while maintaining a (dead) CFG edge to the old block. 940 NewSISucc->getTerminator()->eraseFromParent(); 941 BranchInst::Create(Abort, OldSISucc, 942 ConstantInt::getTrue(Context), NewSISucc); 943 // Release the PHI operands for this edge. 944 for (BasicBlock::iterator II = NewSISucc->begin(); 945 PHINode *PN = dyn_cast<PHINode>(II); ++II) 946 PN->setIncomingValue(PN->getBasicBlockIndex(Switch), 947 UndefValue::get(PN->getType())); 948 // Tell the domtree about the new block. We don't fully update the 949 // domtree here -- instead we force it to do a full recomputation 950 // after the pass is complete -- but we do need to inform it of 951 // new blocks. 952 if (DT) 953 DT->addNewBlock(Abort, NewSISucc); 954 } 955 956 SimplifyCode(Worklist, L); 957} 958 959/// SimplifyCode - Okay, now that we have simplified some instructions in the 960/// loop, walk over it and constant prop, dce, and fold control flow where 961/// possible. Note that this is effectively a very simple loop-structure-aware 962/// optimizer. During processing of this loop, L could very well be deleted, so 963/// it must not be used. 964/// 965/// FIXME: When the loop optimizer is more mature, separate this out to a new 966/// pass. 967/// 968void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 969 while (!Worklist.empty()) { 970 Instruction *I = Worklist.back(); 971 Worklist.pop_back(); 972 973 // Simple constant folding. 974 if (Constant *C = ConstantFoldInstruction(I)) { 975 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 976 continue; 977 } 978 979 // Simple DCE. 980 if (isInstructionTriviallyDead(I)) { 981 DEBUG(dbgs() << "Remove dead instruction '" << *I); 982 983 // Add uses to the worklist, which may be dead now. 984 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 985 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 986 Worklist.push_back(Use); 987 LPM->deleteSimpleAnalysisValue(I, L); 988 RemoveFromWorklist(I, Worklist); 989 I->eraseFromParent(); 990 ++NumSimplify; 991 continue; 992 } 993 994 // See if instruction simplification can hack this up. This is common for 995 // things like "select false, X, Y" after unswitching made the condition be 996 // 'false'. 997 if (Value *V = SimplifyInstruction(I)) { 998 ReplaceUsesOfWith(I, V, Worklist, L, LPM); 999 continue; 1000 } 1001 1002 // Special case hacks that appear commonly in unswitched code. 1003 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1004 if (BI->isUnconditional()) { 1005 // If BI's parent is the only pred of the successor, fold the two blocks 1006 // together. 1007 BasicBlock *Pred = BI->getParent(); 1008 BasicBlock *Succ = BI->getSuccessor(0); 1009 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1010 if (!SinglePred) continue; // Nothing to do. 1011 assert(SinglePred == Pred && "CFG broken"); 1012 1013 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " 1014 << Succ->getName() << "\n"); 1015 1016 // Resolve any single entry PHI nodes in Succ. 1017 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1018 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1019 1020 // Move all of the successor contents from Succ to Pred. 1021 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1022 Succ->end()); 1023 LPM->deleteSimpleAnalysisValue(BI, L); 1024 BI->eraseFromParent(); 1025 RemoveFromWorklist(BI, Worklist); 1026 1027 // If Succ has any successors with PHI nodes, update them to have 1028 // entries coming from Pred instead of Succ. 1029 Succ->replaceAllUsesWith(Pred); 1030 1031 // Remove Succ from the loop tree. 1032 LI->removeBlock(Succ); 1033 LPM->deleteSimpleAnalysisValue(Succ, L); 1034 Succ->eraseFromParent(); 1035 ++NumSimplify; 1036 continue; 1037 } 1038 1039 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1040 // Conditional branch. Turn it into an unconditional branch, then 1041 // remove dead blocks. 1042 continue; // FIXME: Enable. 1043 1044 DEBUG(dbgs() << "Folded branch: " << *BI); 1045 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1046 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1047 DeadSucc->removePredecessor(BI->getParent(), true); 1048 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1049 LPM->deleteSimpleAnalysisValue(BI, L); 1050 BI->eraseFromParent(); 1051 RemoveFromWorklist(BI, Worklist); 1052 ++NumSimplify; 1053 1054 RemoveBlockIfDead(DeadSucc, Worklist, L); 1055 } 1056 continue; 1057 } 1058 } 1059} 1060