LoopUnswitch.cpp revision 02dea8b39f3acad5de1df36273444d149145e7fc
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/LoopInfo.h" 37#include "llvm/Analysis/LoopPass.h" 38#include "llvm/Analysis/Dominators.h" 39#include "llvm/Transforms/Utils/Cloning.h" 40#include "llvm/Transforms/Utils/Local.h" 41#include "llvm/Transforms/Utils/BasicBlockUtils.h" 42#include "llvm/ADT/Statistic.h" 43#include "llvm/ADT/SmallPtrSet.h" 44#include "llvm/Support/CommandLine.h" 45#include "llvm/Support/Compiler.h" 46#include "llvm/Support/Debug.h" 47#include <algorithm> 48#include <set> 49using namespace llvm; 50 51STATISTIC(NumBranches, "Number of branches unswitched"); 52STATISTIC(NumSwitches, "Number of switches unswitched"); 53STATISTIC(NumSelects , "Number of selects unswitched"); 54STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 55STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 56 57static cl::opt<unsigned> 58Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 59 cl::init(10), cl::Hidden); 60 61namespace { 62 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass { 63 LoopInfo *LI; // Loop information 64 LPPassManager *LPM; 65 66 // LoopProcessWorklist - Used to check if second loop needs processing 67 // after RewriteLoopBodyWithConditionConstant rewrites first loop. 68 std::vector<Loop*> LoopProcessWorklist; 69 SmallPtrSet<Value *,8> UnswitchedVals; 70 71 bool OptimizeForSize; 72 bool redoLoop; 73 74 DominanceFrontier *DF; 75 DominatorTree *DT; 76 77 /// LoopDF - Loop's dominance frontier. This set is a collection of 78 /// loop exiting blocks' DF member blocks. However this does set does not 79 /// includes basic blocks that are inside loop. 80 SmallPtrSet<BasicBlock *, 8> LoopDF; 81 82 /// OrigLoopExitMap - This is used to map loop exiting block with 83 /// corresponding loop exit block, before updating CFG. 84 DenseMap<BasicBlock *, BasicBlock *> OrigLoopExitMap; 85 public: 86 static char ID; // Pass ID, replacement for typeid 87 explicit LoopUnswitch(bool Os = false) : 88 LoopPass((intptr_t)&ID), OptimizeForSize(Os), redoLoop(false) {} 89 90 bool runOnLoop(Loop *L, LPPassManager &LPM); 91 bool processLoop(Loop *L); 92 93 /// This transformation requires natural loop information & requires that 94 /// loop preheaders be inserted into the CFG... 95 /// 96 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 97 AU.addRequiredID(LoopSimplifyID); 98 AU.addPreservedID(LoopSimplifyID); 99 AU.addRequired<LoopInfo>(); 100 AU.addPreserved<LoopInfo>(); 101 AU.addRequiredID(LCSSAID); 102 AU.addPreservedID(LCSSAID); 103 AU.addPreserved<DominatorTree>(); 104 AU.addPreserved<DominanceFrontier>(); 105 } 106 107 private: 108 109 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 110 /// remove it. 111 void RemoveLoopFromWorklist(Loop *L) { 112 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 113 LoopProcessWorklist.end(), L); 114 if (I != LoopProcessWorklist.end()) 115 LoopProcessWorklist.erase(I); 116 } 117 118 /// Split all of the edges from inside the loop to their exit blocks. 119 /// Update the appropriate Phi nodes as we do so. 120 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks, 121 SmallVector<BasicBlock *, 8> &MiddleBlocks); 122 123 /// If BB's dominance frontier has a member that is not part of loop L then 124 /// remove it. Add NewDFMember in BB's dominance frontier. 125 void ReplaceLoopExternalDFMember(Loop *L, BasicBlock *BB, 126 BasicBlock *NewDFMember); 127 128 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L); 129 unsigned getLoopUnswitchCost(Loop *L, Value *LIC); 130 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 131 BasicBlock *ExitBlock); 132 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 133 134 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 135 Constant *Val, bool isEqual); 136 137 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 138 BasicBlock *TrueDest, 139 BasicBlock *FalseDest, 140 Instruction *InsertPt); 141 142 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L); 143 void RemoveBlockIfDead(BasicBlock *BB, 144 std::vector<Instruction*> &Worklist, Loop *l); 145 void RemoveLoopFromHierarchy(Loop *L); 146 }; 147} 148char LoopUnswitch::ID = 0; 149static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 150 151LoopPass *llvm::createLoopUnswitchPass(bool Os) { 152 return new LoopUnswitch(Os); 153} 154 155/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 156/// invariant in the loop, or has an invariant piece, return the invariant. 157/// Otherwise, return null. 158static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 159 // Constants should be folded, not unswitched on! 160 if (isa<Constant>(Cond)) return false; 161 162 // TODO: Handle: br (VARIANT|INVARIANT). 163 // TODO: Hoist simple expressions out of loops. 164 if (L->isLoopInvariant(Cond)) return Cond; 165 166 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 167 if (BO->getOpcode() == Instruction::And || 168 BO->getOpcode() == Instruction::Or) { 169 // If either the left or right side is invariant, we can unswitch on this, 170 // which will cause the branch to go away in one loop and the condition to 171 // simplify in the other one. 172 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 173 return LHS; 174 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 175 return RHS; 176 } 177 178 return 0; 179} 180 181bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 182 LI = &getAnalysis<LoopInfo>(); 183 LPM = &LPM_Ref; 184 DF = getAnalysisToUpdate<DominanceFrontier>(); 185 DT = getAnalysisToUpdate<DominatorTree>(); 186 187 bool Changed = false; 188 189 do { 190 redoLoop = false; 191 Changed |= processLoop(L); 192 } while(redoLoop); 193 194 return Changed; 195} 196 197/// processLoop - Do actual work and unswitch loop if possible and profitable. 198bool LoopUnswitch::processLoop(Loop *L) { 199 assert(L->isLCSSAForm()); 200 bool Changed = false; 201 202 // Loop over all of the basic blocks in the loop. If we find an interior 203 // block that is branching on a loop-invariant condition, we can unswitch this 204 // loop. 205 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 206 I != E; ++I) { 207 TerminatorInst *TI = (*I)->getTerminator(); 208 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 209 // If this isn't branching on an invariant condition, we can't unswitch 210 // it. 211 if (BI->isConditional()) { 212 // See if this, or some part of it, is loop invariant. If so, we can 213 // unswitch on it if we desire. 214 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); 215 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 216 L)) { 217 ++NumBranches; 218 return true; 219 } 220 } 221 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 222 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 223 if (LoopCond && SI->getNumCases() > 1) { 224 // Find a value to unswitch on: 225 // FIXME: this should chose the most expensive case! 226 Constant *UnswitchVal = SI->getCaseValue(1); 227 // Do not process same value again and again. 228 if (!UnswitchedVals.insert(UnswitchVal)) 229 continue; 230 231 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { 232 ++NumSwitches; 233 return true; 234 } 235 } 236 } 237 238 // Scan the instructions to check for unswitchable values. 239 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 240 BBI != E; ++BBI) 241 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 242 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 243 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 244 L)) { 245 ++NumSelects; 246 return true; 247 } 248 } 249 } 250 251 assert(L->isLCSSAForm()); 252 253 return Changed; 254} 255 256/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 257/// 1. Exit the loop with no side effects. 258/// 2. Branch to the latch block with no side-effects. 259/// 260/// If these conditions are true, we return true and set ExitBB to the block we 261/// exit through. 262/// 263static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 264 BasicBlock *&ExitBB, 265 std::set<BasicBlock*> &Visited) { 266 if (!Visited.insert(BB).second) { 267 // Already visited and Ok, end of recursion. 268 return true; 269 } else if (!L->contains(BB)) { 270 // Otherwise, this is a loop exit, this is fine so long as this is the 271 // first exit. 272 if (ExitBB != 0) return false; 273 ExitBB = BB; 274 return true; 275 } 276 277 // Otherwise, this is an unvisited intra-loop node. Check all successors. 278 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 279 // Check to see if the successor is a trivial loop exit. 280 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 281 return false; 282 } 283 284 // Okay, everything after this looks good, check to make sure that this block 285 // doesn't include any side effects. 286 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 287 if (I->mayWriteToMemory()) 288 return false; 289 290 return true; 291} 292 293/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 294/// leads to an exit from the specified loop, and has no side-effects in the 295/// process. If so, return the block that is exited to, otherwise return null. 296static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 297 std::set<BasicBlock*> Visited; 298 Visited.insert(L->getHeader()); // Branches to header are ok. 299 BasicBlock *ExitBB = 0; 300 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 301 return ExitBB; 302 return 0; 303} 304 305/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 306/// trivial: that is, that the condition controls whether or not the loop does 307/// anything at all. If this is a trivial condition, unswitching produces no 308/// code duplications (equivalently, it produces a simpler loop and a new empty 309/// loop, which gets deleted). 310/// 311/// If this is a trivial condition, return true, otherwise return false. When 312/// returning true, this sets Cond and Val to the condition that controls the 313/// trivial condition: when Cond dynamically equals Val, the loop is known to 314/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 315/// Cond == Val. 316/// 317static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, 318 BasicBlock **LoopExit = 0) { 319 BasicBlock *Header = L->getHeader(); 320 TerminatorInst *HeaderTerm = Header->getTerminator(); 321 322 BasicBlock *LoopExitBB = 0; 323 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 324 // If the header block doesn't end with a conditional branch on Cond, we 325 // can't handle it. 326 if (!BI->isConditional() || BI->getCondition() != Cond) 327 return false; 328 329 // Check to see if a successor of the branch is guaranteed to go to the 330 // latch block or exit through a one exit block without having any 331 // side-effects. If so, determine the value of Cond that causes it to do 332 // this. 333 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { 334 if (Val) *Val = ConstantInt::getTrue(); 335 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { 336 if (Val) *Val = ConstantInt::getFalse(); 337 } 338 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 339 // If this isn't a switch on Cond, we can't handle it. 340 if (SI->getCondition() != Cond) return false; 341 342 // Check to see if a successor of the switch is guaranteed to go to the 343 // latch block or exit through a one exit block without having any 344 // side-effects. If so, determine the value of Cond that causes it to do 345 // this. Note that we can't trivially unswitch on the default case. 346 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 347 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) { 348 // Okay, we found a trivial case, remember the value that is trivial. 349 if (Val) *Val = SI->getCaseValue(i); 350 break; 351 } 352 } 353 354 // If we didn't find a single unique LoopExit block, or if the loop exit block 355 // contains phi nodes, this isn't trivial. 356 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 357 return false; // Can't handle this. 358 359 if (LoopExit) *LoopExit = LoopExitBB; 360 361 // We already know that nothing uses any scalar values defined inside of this 362 // loop. As such, we just have to check to see if this loop will execute any 363 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 364 // part of the loop that the code *would* execute. We already checked the 365 // tail, check the header now. 366 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 367 if (I->mayWriteToMemory()) 368 return false; 369 return true; 370} 371 372/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 373/// we choose to unswitch the specified loop on the specified value. 374/// 375unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { 376 // If the condition is trivial, always unswitch. There is no code growth for 377 // this case. 378 if (IsTrivialUnswitchCondition(L, LIC)) 379 return 0; 380 381 // FIXME: This is really overly conservative. However, more liberal 382 // estimations have thus far resulted in excessive unswitching, which is bad 383 // both in compile time and in code size. This should be replaced once 384 // someone figures out how a good estimation. 385 return L->getBlocks().size(); 386 387 unsigned Cost = 0; 388 // FIXME: this is brain dead. It should take into consideration code 389 // shrinkage. 390 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 391 I != E; ++I) { 392 BasicBlock *BB = *I; 393 // Do not include empty blocks in the cost calculation. This happen due to 394 // loop canonicalization and will be removed. 395 if (BB->begin() == BasicBlock::iterator(BB->getTerminator())) 396 continue; 397 398 // Count basic blocks. 399 ++Cost; 400 } 401 402 return Cost; 403} 404 405/// UnswitchIfProfitable - We have found that we can unswitch L when 406/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 407/// unswitch the loop, reprocess the pieces, then return true. 408bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ 409 // Check to see if it would be profitable to unswitch this loop. 410 unsigned Cost = getLoopUnswitchCost(L, LoopCond); 411 412 // Do not do non-trivial unswitch while optimizing for size. 413 if (Cost && OptimizeForSize) 414 return false; 415 416 if (Cost > Threshold) { 417 // FIXME: this should estimate growth by the amount of code shared by the 418 // resultant unswitched loops. 419 // 420 DOUT << "NOT unswitching loop %" 421 << L->getHeader()->getName() << ", cost too high: " 422 << L->getBlocks().size() << "\n"; 423 return false; 424 } 425 426 // If this is a trivial condition to unswitch (which results in no code 427 // duplication), do it now. 428 Constant *CondVal; 429 BasicBlock *ExitBlock; 430 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) { 431 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock); 432 } else { 433 UnswitchNontrivialCondition(LoopCond, Val, L); 434 } 435 436 return true; 437} 438 439// RemapInstruction - Convert the instruction operands from referencing the 440// current values into those specified by ValueMap. 441// 442static inline void RemapInstruction(Instruction *I, 443 DenseMap<const Value *, Value*> &ValueMap) { 444 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 445 Value *Op = I->getOperand(op); 446 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 447 if (It != ValueMap.end()) Op = It->second; 448 I->setOperand(op, Op); 449 } 450} 451 452// CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator 453// Info. 454// 455// If Orig block's immediate dominator is mapped in VM then use corresponding 456// immediate dominator from the map. Otherwise Orig block's dominator is also 457// NewBB's dominator. 458// 459// OrigPreheader is loop pre-header before this pass started 460// updating CFG. NewPrehader is loops new pre-header. However, after CFG 461// manipulation, loop L may not exist. So rely on input parameter NewPreheader. 462static void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig, 463 BasicBlock *NewPreheader, BasicBlock *OrigPreheader, 464 BasicBlock *OrigHeader, 465 DominatorTree *DT, DominanceFrontier *DF, 466 DenseMap<const Value*, Value*> &VM) { 467 468 // If NewBB alreay has found its place in domiantor tree then no need to do 469 // anything. 470 if (DT->getNode(NewBB)) 471 return; 472 473 // If Orig does not have any immediate domiantor then its clone, NewBB, does 474 // not need any immediate dominator. 475 DomTreeNode *OrigNode = DT->getNode(Orig); 476 if (!OrigNode) 477 return; 478 DomTreeNode *OrigIDomNode = OrigNode->getIDom(); 479 if (!OrigIDomNode) 480 return; 481 482 BasicBlock *OrigIDom = NULL; 483 484 // If Orig is original loop header then its immediate dominator is 485 // NewPreheader. 486 if (Orig == OrigHeader) 487 OrigIDom = NewPreheader; 488 489 // If Orig is new pre-header then its immediate dominator is 490 // original pre-header. 491 else if (Orig == NewPreheader) 492 OrigIDom = OrigPreheader; 493 494 // Other as DT to find Orig's immediate dominator. 495 else 496 OrigIDom = OrigIDomNode->getBlock(); 497 498 // Initially use Orig's immediate dominator as NewBB's immediate dominator. 499 BasicBlock *NewIDom = OrigIDom; 500 DenseMap<const Value*, Value*>::iterator I = VM.find(OrigIDom); 501 if (I != VM.end()) { 502 NewIDom = cast<BasicBlock>(I->second); 503 504 // If NewIDom does not have corresponding dominatore tree node then 505 // get one. 506 if (!DT->getNode(NewIDom)) 507 CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader, 508 OrigHeader, DT, DF, VM); 509 } 510 511 DT->addNewBlock(NewBB, NewIDom); 512 513 // Copy cloned dominance frontiner set 514 DominanceFrontier::DomSetType NewDFSet; 515 if (DF) { 516 DominanceFrontier::iterator DFI = DF->find(Orig); 517 if ( DFI != DF->end()) { 518 DominanceFrontier::DomSetType S = DFI->second; 519 for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end(); 520 I != E; ++I) { 521 BasicBlock *BB = *I; 522 DenseMap<const Value*, Value*>::iterator IDM = VM.find(BB); 523 if (IDM != VM.end()) 524 NewDFSet.insert(cast<BasicBlock>(IDM->second)); 525 else 526 NewDFSet.insert(BB); 527 } 528 } 529 DF->addBasicBlock(NewBB, NewDFSet); 530 } 531} 532 533/// CloneLoop - Recursively clone the specified loop and all of its children, 534/// mapping the blocks with the specified map. 535static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 536 LoopInfo *LI, LPPassManager *LPM) { 537 Loop *New = new Loop(); 538 539 LPM->insertLoop(New, PL); 540 541 // Add all of the blocks in L to the new loop. 542 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 543 I != E; ++I) 544 if (LI->getLoopFor(*I) == L) 545 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase()); 546 547 // Add all of the subloops to the new loop. 548 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 549 CloneLoop(*I, New, VM, LI, LPM); 550 551 return New; 552} 553 554/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 555/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 556/// code immediately before InsertPt. 557void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 558 BasicBlock *TrueDest, 559 BasicBlock *FalseDest, 560 Instruction *InsertPt) { 561 // Insert a conditional branch on LIC to the two preheaders. The original 562 // code is the true version and the new code is the false version. 563 Value *BranchVal = LIC; 564 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty) 565 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); 566 else if (Val != ConstantInt::getTrue()) 567 // We want to enter the new loop when the condition is true. 568 std::swap(TrueDest, FalseDest); 569 570 // Insert the new branch. 571 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); 572 573} 574 575 576/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 577/// condition in it (a cond branch from its header block to its latch block, 578/// where the path through the loop that doesn't execute its body has no 579/// side-effects), unswitch it. This doesn't involve any code duplication, just 580/// moving the conditional branch outside of the loop and updating loop info. 581void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 582 Constant *Val, 583 BasicBlock *ExitBlock) { 584 DOUT << "loop-unswitch: Trivial-Unswitch loop %" 585 << L->getHeader()->getName() << " [" << L->getBlocks().size() 586 << " blocks] in Function " << L->getHeader()->getParent()->getName() 587 << " on cond: " << *Val << " == " << *Cond << "\n"; 588 589 // First step, split the preheader, so that we know that there is a safe place 590 // to insert the conditional branch. We will change 'OrigPH' to have a 591 // conditional branch on Cond. 592 BasicBlock *OrigPH = L->getLoopPreheader(); 593 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader(), this); 594 595 // Now that we have a place to insert the conditional branch, create a place 596 // to branch to: this is the exit block out of the loop that we should 597 // short-circuit to. 598 599 // Split this block now, so that the loop maintains its exit block, and so 600 // that the jump from the preheader can execute the contents of the exit block 601 // without actually branching to it (the exit block should be dominated by the 602 // loop header, not the preheader). 603 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 604 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); 605 606 // Okay, now we have a position to branch from and a position to branch to, 607 // insert the new conditional branch. 608 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 609 OrigPH->getTerminator()); 610 LPM->deleteSimpleAnalysisValue(OrigPH->getTerminator(), L); 611 OrigPH->getTerminator()->eraseFromParent(); 612 613 // We need to reprocess this loop, it could be unswitched again. 614 redoLoop = true; 615 616 // Now that we know that the loop is never entered when this condition is a 617 // particular value, rewrite the loop with this info. We know that this will 618 // at least eliminate the old branch. 619 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 620 ++NumTrivial; 621} 622 623/// ReplaceLoopExternalDFMember - 624/// If BB's dominance frontier has a member that is not part of loop L then 625/// remove it. Add NewDFMember in BB's dominance frontier. 626void LoopUnswitch::ReplaceLoopExternalDFMember(Loop *L, BasicBlock *BB, 627 BasicBlock *NewDFMember) { 628 629 DominanceFrontier::iterator DFI = DF->find(BB); 630 if (DFI == DF->end()) 631 return; 632 633 DominanceFrontier::DomSetType &DFSet = DFI->second; 634 for (DominanceFrontier::DomSetType::iterator DI = DFSet.begin(), 635 DE = DFSet.end(); DI != DE;) { 636 BasicBlock *B = *DI++; 637 if (L->contains(B)) 638 continue; 639 640 DF->removeFromFrontier(DFI, B); 641 LoopDF.insert(B); 642 } 643 644 DF->addToFrontier(DFI, NewDFMember); 645} 646 647/// SplitExitEdges - Split all of the edges from inside the loop to their exit 648/// blocks. Update the appropriate Phi nodes as we do so. 649void LoopUnswitch::SplitExitEdges(Loop *L, 650 const SmallVector<BasicBlock *, 8> &ExitBlocks, 651 SmallVector<BasicBlock *, 8> &MiddleBlocks) { 652 653 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 654 BasicBlock *ExitBlock = ExitBlocks[i]; 655 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 656 657 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 658 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock, this); 659 MiddleBlocks.push_back(MiddleBlock); 660 BasicBlock* StartBlock = Preds[j]; 661 BasicBlock* EndBlock; 662 if (MiddleBlock->getSinglePredecessor() == ExitBlock) { 663 EndBlock = MiddleBlock; 664 MiddleBlock = EndBlock->getSinglePredecessor();; 665 } else { 666 EndBlock = ExitBlock; 667 } 668 669 OrigLoopExitMap[StartBlock] = EndBlock; 670 671 std::set<PHINode*> InsertedPHIs; 672 PHINode* OldLCSSA = 0; 673 for (BasicBlock::iterator I = EndBlock->begin(); 674 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) { 675 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock); 676 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(), 677 OldLCSSA->getName() + ".us-lcssa", 678 MiddleBlock->getTerminator()); 679 NewLCSSA->addIncoming(OldValue, StartBlock); 680 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock), 681 NewLCSSA); 682 InsertedPHIs.insert(NewLCSSA); 683 } 684 685 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI(); 686 for (BasicBlock::iterator I = MiddleBlock->begin(); 687 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0; 688 ++I) { 689 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(), 690 OldLCSSA->getName() + ".us-lcssa", 691 InsertPt); 692 OldLCSSA->replaceAllUsesWith(NewLCSSA); 693 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); 694 } 695 696 if (DF && DT) { 697 // StartBlock -- > MiddleBlock -- > EndBlock 698 // StartBlock is loop exiting block. EndBlock will become merge point 699 // of two loop exits after loop unswitch. 700 701 // If StartBlock's DF member includes a block that is not loop member 702 // then replace that DF member with EndBlock. 703 704 // If MiddleBlock's DF member includes a block that is not loop member 705 // tnen replace that DF member with EndBlock. 706 707 ReplaceLoopExternalDFMember(L, StartBlock, EndBlock); 708 ReplaceLoopExternalDFMember(L, MiddleBlock, EndBlock); 709 } 710 } 711 } 712 713} 714 715/// UnswitchNontrivialCondition - We determined that the loop is profitable 716/// to unswitch when LIC equal Val. Split it into loop versions and test the 717/// condition outside of either loop. Return the loops created as Out1/Out2. 718void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 719 Loop *L) { 720 Function *F = L->getHeader()->getParent(); 721 DOUT << "loop-unswitch: Unswitching loop %" 722 << L->getHeader()->getName() << " [" << L->getBlocks().size() 723 << " blocks] in Function " << F->getName() 724 << " when '" << *Val << "' == " << *LIC << "\n"; 725 726 // LoopBlocks contains all of the basic blocks of the loop, including the 727 // preheader of the loop, the body of the loop, and the exit blocks of the 728 // loop, in that order. 729 std::vector<BasicBlock*> LoopBlocks; 730 731 // First step, split the preheader and exit blocks, and add these blocks to 732 // the LoopBlocks list. 733 BasicBlock *OrigHeader = L->getHeader(); 734 BasicBlock *OrigPreheader = L->getLoopPreheader(); 735 BasicBlock *NewPreheader = SplitEdge(OrigPreheader, L->getHeader(), this); 736 LoopBlocks.push_back(NewPreheader); 737 738 // We want the loop to come after the preheader, but before the exit blocks. 739 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 740 741 SmallVector<BasicBlock*, 8> ExitBlocks; 742 L->getUniqueExitBlocks(ExitBlocks); 743 744 // Split all of the edges from inside the loop to their exit blocks. Update 745 // the appropriate Phi nodes as we do so. 746 SmallVector<BasicBlock *,8> MiddleBlocks; 747 SplitExitEdges(L, ExitBlocks, MiddleBlocks); 748 749 // The exit blocks may have been changed due to edge splitting, recompute. 750 ExitBlocks.clear(); 751 L->getUniqueExitBlocks(ExitBlocks); 752 753 // Add exit blocks to the loop blocks. 754 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 755 756 // Next step, clone all of the basic blocks that make up the loop (including 757 // the loop preheader and exit blocks), keeping track of the mapping between 758 // the instructions and blocks. 759 std::vector<BasicBlock*> NewBlocks; 760 NewBlocks.reserve(LoopBlocks.size()); 761 DenseMap<const Value*, Value*> ValueMap; 762 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 763 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 764 NewBlocks.push_back(New); 765 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 766 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L); 767 } 768 769 // OutSiders are basic block that are dominated by original header and 770 // at the same time they are not part of loop. 771 SmallPtrSet<BasicBlock *, 8> OutSiders; 772 if (DT) { 773 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); 774 for(std::vector<DomTreeNode*>::iterator DI = OrigHeaderNode->begin(), 775 DE = OrigHeaderNode->end(); DI != DE; ++DI) { 776 BasicBlock *B = (*DI)->getBlock(); 777 778 DenseMap<const Value*, Value*>::iterator VI = ValueMap.find(B); 779 if (VI == ValueMap.end()) 780 OutSiders.insert(B); 781 } 782 } 783 784 // Splice the newly inserted blocks into the function right before the 785 // original preheader. 786 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 787 NewBlocks[0], F->end()); 788 789 // Now we create the new Loop object for the versioned loop. 790 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 791 Loop *ParentLoop = L->getParentLoop(); 792 if (ParentLoop) { 793 // Make sure to add the cloned preheader and exit blocks to the parent loop 794 // as well. 795 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); 796 } 797 798 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 799 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 800 // The new exit block should be in the same loop as the old one. 801 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 802 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); 803 804 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 805 "Exit block should have been split to have one successor!"); 806 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 807 808 // If the successor of the exit block had PHI nodes, add an entry for 809 // NewExit. 810 PHINode *PN; 811 for (BasicBlock::iterator I = ExitSucc->begin(); 812 (PN = dyn_cast<PHINode>(I)); ++I) { 813 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 814 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 815 if (It != ValueMap.end()) V = It->second; 816 PN->addIncoming(V, NewExit); 817 } 818 } 819 820 // Rewrite the code to refer to itself. 821 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 822 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 823 E = NewBlocks[i]->end(); I != E; ++I) 824 RemapInstruction(I, ValueMap); 825 826 // Rewrite the original preheader to select between versions of the loop. 827 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 828 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 829 "Preheader splitting did not work correctly!"); 830 831 // Emit the new branch that selects between the two versions of this loop. 832 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 833 LPM->deleteSimpleAnalysisValue(OldBR, L); 834 OldBR->eraseFromParent(); 835 836 // Update dominator info 837 if (DF && DT) { 838 839 SmallVector<BasicBlock *,4> ExitingBlocks; 840 L->getExitingBlocks(ExitingBlocks); 841 842 // Clone dominator info for all cloned basic block. 843 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 844 BasicBlock *LBB = LoopBlocks[i]; 845 BasicBlock *NBB = NewBlocks[i]; 846 CloneDomInfo(NBB, LBB, NewPreheader, OrigPreheader, 847 OrigHeader, DT, DF, ValueMap); 848 849 // If LBB's dominance frontier includes DFMember 850 // such that DFMember is also a member of LoopDF then 851 // - Remove DFMember from LBB's dominance frontier 852 // - Copy loop exiting blocks', that are dominated by BB, 853 // dominance frontier member in BB's dominance frontier 854 855 DominanceFrontier::iterator LBBI = DF->find(LBB); 856 DominanceFrontier::iterator NBBI = DF->find(NBB); 857 if (LBBI == DF->end()) 858 continue; 859 860 DominanceFrontier::DomSetType &LBSet = LBBI->second; 861 for (DominanceFrontier::DomSetType::iterator LI = LBSet.begin(), 862 LE = LBSet.end(); LI != LE; /* NULL */) { 863 BasicBlock *B = *LI++; 864 if (B == LBB && B == L->getHeader()) 865 continue; 866 bool removeB = false; 867 if (!LoopDF.count(B)) 868 continue; 869 870 // If LBB dominates loop exits then insert loop exit block's DF 871 // into B's DF. 872 for(SmallVector<BasicBlock *, 4>::iterator 873 LExitI = ExitingBlocks.begin(), 874 LExitE = ExitingBlocks.end(); LExitI != LExitE; ++LExitI) { 875 BasicBlock *E = *LExitI; 876 877 if (!DT->dominates(LBB,E)) 878 continue; 879 880 DenseMap<BasicBlock *, BasicBlock *>::iterator DFBI = 881 OrigLoopExitMap.find(E); 882 if (DFBI == OrigLoopExitMap.end()) 883 continue; 884 885 BasicBlock *DFB = DFBI->second; 886 DF->addToFrontier(LBBI, DFB); 887 DF->addToFrontier(NBBI, DFB); 888 removeB = true; 889 } 890 891 // If B's replacement is inserted in DF then now is the time to remove 892 // B. 893 if (removeB) { 894 DF->removeFromFrontier(LBBI, B); 895 if (L->contains(B)) 896 DF->removeFromFrontier(NBBI, cast<BasicBlock>(ValueMap[B])); 897 else 898 DF->removeFromFrontier(NBBI, B); 899 } 900 } 901 902 } 903 904 // MiddleBlocks are dominated by original pre header. SplitEdge updated 905 // MiddleBlocks' dominance frontier appropriately. 906 for (unsigned i = 0, e = MiddleBlocks.size(); i != e; ++i) { 907 BasicBlock *MBB = MiddleBlocks[i]; 908 if (!MBB->getSinglePredecessor()) 909 DT->changeImmediateDominator(MBB, OrigPreheader); 910 } 911 912 // All Outsiders are now dominated by original pre header. 913 for (SmallPtrSet<BasicBlock *, 8>::iterator OI = OutSiders.begin(), 914 OE = OutSiders.end(); OI != OE; ++OI) { 915 BasicBlock *OB = *OI; 916 DT->changeImmediateDominator(OB, OrigPreheader); 917 } 918 919 // New loop headers are dominated by original preheader 920 DT->changeImmediateDominator(NewBlocks[0], OrigPreheader); 921 DT->changeImmediateDominator(LoopBlocks[0], OrigPreheader); 922 } 923 924 LoopProcessWorklist.push_back(NewLoop); 925 redoLoop = true; 926 927 // Now we rewrite the original code to know that the condition is true and the 928 // new code to know that the condition is false. 929 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 930 931 // It's possible that simplifying one loop could cause the other to be 932 // deleted. If so, don't simplify it. 933 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 934 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 935} 936 937/// RemoveFromWorklist - Remove all instances of I from the worklist vector 938/// specified. 939static void RemoveFromWorklist(Instruction *I, 940 std::vector<Instruction*> &Worklist) { 941 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 942 Worklist.end(), I); 943 while (WI != Worklist.end()) { 944 unsigned Offset = WI-Worklist.begin(); 945 Worklist.erase(WI); 946 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 947 } 948} 949 950/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 951/// program, replacing all uses with V and update the worklist. 952static void ReplaceUsesOfWith(Instruction *I, Value *V, 953 std::vector<Instruction*> &Worklist, 954 Loop *L, LPPassManager *LPM) { 955 DOUT << "Replace with '" << *V << "': " << *I; 956 957 // Add uses to the worklist, which may be dead now. 958 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 959 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 960 Worklist.push_back(Use); 961 962 // Add users to the worklist which may be simplified now. 963 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 964 UI != E; ++UI) 965 Worklist.push_back(cast<Instruction>(*UI)); 966 LPM->deleteSimpleAnalysisValue(I, L); 967 RemoveFromWorklist(I, Worklist); 968 I->replaceAllUsesWith(V); 969 I->eraseFromParent(); 970 ++NumSimplify; 971} 972 973/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 974/// information, and remove any dead successors it has. 975/// 976void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 977 std::vector<Instruction*> &Worklist, 978 Loop *L) { 979 if (pred_begin(BB) != pred_end(BB)) { 980 // This block isn't dead, since an edge to BB was just removed, see if there 981 // are any easy simplifications we can do now. 982 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 983 // If it has one pred, fold phi nodes in BB. 984 while (isa<PHINode>(BB->begin())) 985 ReplaceUsesOfWith(BB->begin(), 986 cast<PHINode>(BB->begin())->getIncomingValue(0), 987 Worklist, L, LPM); 988 989 // If this is the header of a loop and the only pred is the latch, we now 990 // have an unreachable loop. 991 if (Loop *L = LI->getLoopFor(BB)) 992 if (L->getHeader() == BB && L->contains(Pred)) { 993 // Remove the branch from the latch to the header block, this makes 994 // the header dead, which will make the latch dead (because the header 995 // dominates the latch). 996 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); 997 Pred->getTerminator()->eraseFromParent(); 998 new UnreachableInst(Pred); 999 1000 // The loop is now broken, remove it from LI. 1001 RemoveLoopFromHierarchy(L); 1002 1003 // Reprocess the header, which now IS dead. 1004 RemoveBlockIfDead(BB, Worklist, L); 1005 return; 1006 } 1007 1008 // If pred ends in a uncond branch, add uncond branch to worklist so that 1009 // the two blocks will get merged. 1010 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 1011 if (BI->isUnconditional()) 1012 Worklist.push_back(BI); 1013 } 1014 return; 1015 } 1016 1017 DOUT << "Nuking dead block: " << *BB; 1018 1019 // Remove the instructions in the basic block from the worklist. 1020 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 1021 RemoveFromWorklist(I, Worklist); 1022 1023 // Anything that uses the instructions in this basic block should have their 1024 // uses replaced with undefs. 1025 if (!I->use_empty()) 1026 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1027 } 1028 1029 // If this is the edge to the header block for a loop, remove the loop and 1030 // promote all subloops. 1031 if (Loop *BBLoop = LI->getLoopFor(BB)) { 1032 if (BBLoop->getLoopLatch() == BB) 1033 RemoveLoopFromHierarchy(BBLoop); 1034 } 1035 1036 // Remove the block from the loop info, which removes it from any loops it 1037 // was in. 1038 LI->removeBlock(BB); 1039 1040 1041 // Remove phi node entries in successors for this block. 1042 TerminatorInst *TI = BB->getTerminator(); 1043 std::vector<BasicBlock*> Succs; 1044 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 1045 Succs.push_back(TI->getSuccessor(i)); 1046 TI->getSuccessor(i)->removePredecessor(BB); 1047 } 1048 1049 // Unique the successors, remove anything with multiple uses. 1050 std::sort(Succs.begin(), Succs.end()); 1051 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 1052 1053 // Remove the basic block, including all of the instructions contained in it. 1054 LPM->deleteSimpleAnalysisValue(BB, L); 1055 BB->eraseFromParent(); 1056 // Remove successor blocks here that are not dead, so that we know we only 1057 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 1058 // then getting removed before we revisit them, which is badness. 1059 // 1060 for (unsigned i = 0; i != Succs.size(); ++i) 1061 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 1062 // One exception is loop headers. If this block was the preheader for a 1063 // loop, then we DO want to visit the loop so the loop gets deleted. 1064 // We know that if the successor is a loop header, that this loop had to 1065 // be the preheader: the case where this was the latch block was handled 1066 // above and headers can only have two predecessors. 1067 if (!LI->isLoopHeader(Succs[i])) { 1068 Succs.erase(Succs.begin()+i); 1069 --i; 1070 } 1071 } 1072 1073 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 1074 RemoveBlockIfDead(Succs[i], Worklist, L); 1075} 1076 1077/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 1078/// become unwrapped, either because the backedge was deleted, or because the 1079/// edge into the header was removed. If the edge into the header from the 1080/// latch block was removed, the loop is unwrapped but subloops are still alive, 1081/// so they just reparent loops. If the loops are actually dead, they will be 1082/// removed later. 1083void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 1084 LPM->deleteLoopFromQueue(L); 1085 RemoveLoopFromWorklist(L); 1086} 1087 1088 1089 1090// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 1091// the value specified by Val in the specified loop, or we know it does NOT have 1092// that value. Rewrite any uses of LIC or of properties correlated to it. 1093void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 1094 Constant *Val, 1095 bool IsEqual) { 1096 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 1097 1098 // FIXME: Support correlated properties, like: 1099 // for (...) 1100 // if (li1 < li2) 1101 // ... 1102 // if (li1 > li2) 1103 // ... 1104 1105 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 1106 // selects, switches. 1107 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 1108 std::vector<Instruction*> Worklist; 1109 1110 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 1111 // in the loop with the appropriate one directly. 1112 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) { 1113 Value *Replacement; 1114 if (IsEqual) 1115 Replacement = Val; 1116 else 1117 Replacement = ConstantInt::get(Type::Int1Ty, 1118 !cast<ConstantInt>(Val)->getZExtValue()); 1119 1120 for (unsigned i = 0, e = Users.size(); i != e; ++i) 1121 if (Instruction *U = cast<Instruction>(Users[i])) { 1122 if (!L->contains(U->getParent())) 1123 continue; 1124 U->replaceUsesOfWith(LIC, Replacement); 1125 Worklist.push_back(U); 1126 } 1127 } else { 1128 // Otherwise, we don't know the precise value of LIC, but we do know that it 1129 // is certainly NOT "Val". As such, simplify any uses in the loop that we 1130 // can. This case occurs when we unswitch switch statements. 1131 for (unsigned i = 0, e = Users.size(); i != e; ++i) 1132 if (Instruction *U = cast<Instruction>(Users[i])) { 1133 if (!L->contains(U->getParent())) 1134 continue; 1135 1136 Worklist.push_back(U); 1137 1138 // If we know that LIC is not Val, use this info to simplify code. 1139 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 1140 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 1141 if (SI->getCaseValue(i) == Val) { 1142 // Found a dead case value. Don't remove PHI nodes in the 1143 // successor if they become single-entry, those PHI nodes may 1144 // be in the Users list. 1145 1146 // FIXME: This is a hack. We need to keep the successor around 1147 // and hooked up so as to preserve the loop structure, because 1148 // trying to update it is complicated. So instead we preserve the 1149 // loop structure and put the block on an dead code path. 1150 1151 BasicBlock* Old = SI->getParent(); 1152 BasicBlock* Split = SplitBlock(Old, SI, this); 1153 1154 Instruction* OldTerm = Old->getTerminator(); 1155 BranchInst::Create(Split, SI->getSuccessor(i), 1156 ConstantInt::getTrue(), OldTerm); 1157 1158 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L); 1159 Old->getTerminator()->eraseFromParent(); 1160 1161 PHINode *PN; 1162 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); 1163 (PN = dyn_cast<PHINode>(II)); ++II) { 1164 Value *InVal = PN->removeIncomingValue(Split, false); 1165 PN->addIncoming(InVal, Old); 1166 } 1167 1168 SI->removeCase(i); 1169 break; 1170 } 1171 } 1172 } 1173 1174 // TODO: We could do other simplifications, for example, turning 1175 // LIC == Val -> false. 1176 } 1177 } 1178 1179 SimplifyCode(Worklist, L); 1180} 1181 1182/// SimplifyCode - Okay, now that we have simplified some instructions in the 1183/// loop, walk over it and constant prop, dce, and fold control flow where 1184/// possible. Note that this is effectively a very simple loop-structure-aware 1185/// optimizer. During processing of this loop, L could very well be deleted, so 1186/// it must not be used. 1187/// 1188/// FIXME: When the loop optimizer is more mature, separate this out to a new 1189/// pass. 1190/// 1191void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) { 1192 while (!Worklist.empty()) { 1193 Instruction *I = Worklist.back(); 1194 Worklist.pop_back(); 1195 1196 // Simple constant folding. 1197 if (Constant *C = ConstantFoldInstruction(I)) { 1198 ReplaceUsesOfWith(I, C, Worklist, L, LPM); 1199 continue; 1200 } 1201 1202 // Simple DCE. 1203 if (isInstructionTriviallyDead(I)) { 1204 DOUT << "Remove dead instruction '" << *I; 1205 1206 // Add uses to the worklist, which may be dead now. 1207 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1208 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1209 Worklist.push_back(Use); 1210 LPM->deleteSimpleAnalysisValue(I, L); 1211 RemoveFromWorklist(I, Worklist); 1212 I->eraseFromParent(); 1213 ++NumSimplify; 1214 continue; 1215 } 1216 1217 // Special case hacks that appear commonly in unswitched code. 1218 switch (I->getOpcode()) { 1219 case Instruction::Select: 1220 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 1221 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, 1222 LPM); 1223 continue; 1224 } 1225 break; 1226 case Instruction::And: 1227 if (isa<ConstantInt>(I->getOperand(0)) && 1228 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1229 cast<BinaryOperator>(I)->swapOperands(); 1230 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1231 if (CB->getType() == Type::Int1Ty) { 1232 if (CB->isOne()) // X & 1 -> X 1233 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1234 else // X & 0 -> 0 1235 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1236 continue; 1237 } 1238 break; 1239 case Instruction::Or: 1240 if (isa<ConstantInt>(I->getOperand(0)) && 1241 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1242 cast<BinaryOperator>(I)->swapOperands(); 1243 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1244 if (CB->getType() == Type::Int1Ty) { 1245 if (CB->isOne()) // X | 1 -> 1 1246 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); 1247 else // X | 0 -> X 1248 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); 1249 continue; 1250 } 1251 break; 1252 case Instruction::Br: { 1253 BranchInst *BI = cast<BranchInst>(I); 1254 if (BI->isUnconditional()) { 1255 // If BI's parent is the only pred of the successor, fold the two blocks 1256 // together. 1257 BasicBlock *Pred = BI->getParent(); 1258 BasicBlock *Succ = BI->getSuccessor(0); 1259 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1260 if (!SinglePred) continue; // Nothing to do. 1261 assert(SinglePred == Pred && "CFG broken"); 1262 1263 DOUT << "Merging blocks: " << Pred->getName() << " <- " 1264 << Succ->getName() << "\n"; 1265 1266 // Resolve any single entry PHI nodes in Succ. 1267 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1268 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); 1269 1270 // Move all of the successor contents from Succ to Pred. 1271 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1272 Succ->end()); 1273 LPM->deleteSimpleAnalysisValue(BI, L); 1274 BI->eraseFromParent(); 1275 RemoveFromWorklist(BI, Worklist); 1276 1277 // If Succ has any successors with PHI nodes, update them to have 1278 // entries coming from Pred instead of Succ. 1279 Succ->replaceAllUsesWith(Pred); 1280 1281 // Remove Succ from the loop tree. 1282 LI->removeBlock(Succ); 1283 LPM->deleteSimpleAnalysisValue(Succ, L); 1284 Succ->eraseFromParent(); 1285 ++NumSimplify; 1286 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1287 // Conditional branch. Turn it into an unconditional branch, then 1288 // remove dead blocks. 1289 break; // FIXME: Enable. 1290 1291 DOUT << "Folded branch: " << *BI; 1292 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1293 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1294 DeadSucc->removePredecessor(BI->getParent(), true); 1295 Worklist.push_back(BranchInst::Create(LiveSucc, BI)); 1296 LPM->deleteSimpleAnalysisValue(BI, L); 1297 BI->eraseFromParent(); 1298 RemoveFromWorklist(BI, Worklist); 1299 ++NumSimplify; 1300 1301 RemoveBlockIfDead(DeadSucc, Worklist, L); 1302 } 1303 break; 1304 } 1305 } 1306 } 1307} 1308