LoopUnswitch.cpp revision 743f7e8839dc02f2e808e977803368c431279189
1//===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass 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/Transforms/Utils/Cloning.h" 39#include "llvm/Transforms/Utils/Local.h" 40#include "llvm/Transforms/Utils/BasicBlockUtils.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/ADT/SmallPtrSet.h" 43#include "llvm/ADT/PostOrderIterator.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 57namespace { 58 cl::opt<unsigned> 59 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 60 cl::init(10), cl::Hidden); 61 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 public: 73 static char ID; // Pass ID, replacement for typeid 74 LoopUnswitch(bool Os = false) : 75 LoopPass((intptr_t)&ID), OptimizeForSize(Os) {} 76 77 bool runOnLoop(Loop *L, LPPassManager &LPM); 78 79 /// This transformation requires natural loop information & requires that 80 /// loop preheaders be inserted into the CFG... 81 /// 82 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 83 AU.addRequiredID(LoopSimplifyID); 84 AU.addPreservedID(LoopSimplifyID); 85 AU.addRequired<LoopInfo>(); 86 AU.addPreserved<LoopInfo>(); 87 AU.addRequiredID(LCSSAID); 88 AU.addPreservedID(LCSSAID); 89 } 90 91 private: 92 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 93 /// remove it. 94 void RemoveLoopFromWorklist(Loop *L) { 95 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 96 LoopProcessWorklist.end(), L); 97 if (I != LoopProcessWorklist.end()) 98 LoopProcessWorklist.erase(I); 99 } 100 101 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L); 102 unsigned getLoopUnswitchCost(Loop *L, Value *LIC); 103 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 104 BasicBlock *ExitBlock); 105 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 106 BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To); 107 BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt); 108 109 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 110 Constant *Val, bool isEqual); 111 112 void SimplifyCode(std::vector<Instruction*> &Worklist); 113 void RemoveBlockIfDead(BasicBlock *BB, 114 std::vector<Instruction*> &Worklist); 115 void RemoveLoopFromHierarchy(Loop *L); 116 }; 117 char LoopUnswitch::ID = 0; 118 RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 119} 120 121LoopPass *llvm::createLoopUnswitchPass(bool Os) { 122 return new LoopUnswitch(Os); 123} 124 125/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 126/// invariant in the loop, or has an invariant piece, return the invariant. 127/// Otherwise, return null. 128static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 129 // Constants should be folded, not unswitched on! 130 if (isa<Constant>(Cond)) return false; 131 132 // TODO: Handle: br (VARIANT|INVARIANT). 133 // TODO: Hoist simple expressions out of loops. 134 if (L->isLoopInvariant(Cond)) return Cond; 135 136 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 137 if (BO->getOpcode() == Instruction::And || 138 BO->getOpcode() == Instruction::Or) { 139 // If either the left or right side is invariant, we can unswitch on this, 140 // which will cause the branch to go away in one loop and the condition to 141 // simplify in the other one. 142 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 143 return LHS; 144 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 145 return RHS; 146 } 147 148 return 0; 149} 150 151bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 152 assert(L->isLCSSAForm()); 153 LI = &getAnalysis<LoopInfo>(); 154 LPM = &LPM_Ref; 155 bool Changed = false; 156 157 // Loop over all of the basic blocks in the loop. If we find an interior 158 // block that is branching on a loop-invariant condition, we can unswitch this 159 // loop. 160 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 161 I != E; ++I) { 162 TerminatorInst *TI = (*I)->getTerminator(); 163 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 164 // If this isn't branching on an invariant condition, we can't unswitch 165 // it. 166 if (BI->isConditional()) { 167 // See if this, or some part of it, is loop invariant. If so, we can 168 // unswitch on it if we desire. 169 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); 170 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 171 L)) { 172 ++NumBranches; 173 return true; 174 } 175 } 176 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 177 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 178 if (LoopCond && SI->getNumCases() > 1) { 179 // Find a value to unswitch on: 180 // FIXME: this should chose the most expensive case! 181 Constant *UnswitchVal = SI->getCaseValue(1); 182 // Do not process same value again and again. 183 if (!UnswitchedVals.insert(UnswitchVal)) 184 continue; 185 186 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { 187 ++NumSwitches; 188 return true; 189 } 190 } 191 } 192 193 // Scan the instructions to check for unswitchable values. 194 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 195 BBI != E; ++BBI) 196 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 197 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 198 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 199 L)) { 200 ++NumSelects; 201 return true; 202 } 203 } 204 } 205 206 assert(L->isLCSSAForm()); 207 208 return Changed; 209} 210 211/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 212/// 1. Exit the loop with no side effects. 213/// 2. Branch to the latch block with no side-effects. 214/// 215/// If these conditions are true, we return true and set ExitBB to the block we 216/// exit through. 217/// 218static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 219 BasicBlock *&ExitBB, 220 std::set<BasicBlock*> &Visited) { 221 if (!Visited.insert(BB).second) { 222 // Already visited and Ok, end of recursion. 223 return true; 224 } else if (!L->contains(BB)) { 225 // Otherwise, this is a loop exit, this is fine so long as this is the 226 // first exit. 227 if (ExitBB != 0) return false; 228 ExitBB = BB; 229 return true; 230 } 231 232 // Otherwise, this is an unvisited intra-loop node. Check all successors. 233 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 234 // Check to see if the successor is a trivial loop exit. 235 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 236 return false; 237 } 238 239 // Okay, everything after this looks good, check to make sure that this block 240 // doesn't include any side effects. 241 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 242 if (I->mayWriteToMemory()) 243 return false; 244 245 return true; 246} 247 248/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 249/// leads to an exit from the specified loop, and has no side-effects in the 250/// process. If so, return the block that is exited to, otherwise return null. 251static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 252 std::set<BasicBlock*> Visited; 253 Visited.insert(L->getHeader()); // Branches to header are ok. 254 BasicBlock *ExitBB = 0; 255 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 256 return ExitBB; 257 return 0; 258} 259 260/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 261/// trivial: that is, that the condition controls whether or not the loop does 262/// anything at all. If this is a trivial condition, unswitching produces no 263/// code duplications (equivalently, it produces a simpler loop and a new empty 264/// loop, which gets deleted). 265/// 266/// If this is a trivial condition, return true, otherwise return false. When 267/// returning true, this sets Cond and Val to the condition that controls the 268/// trivial condition: when Cond dynamically equals Val, the loop is known to 269/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 270/// Cond == Val. 271/// 272static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, 273 BasicBlock **LoopExit = 0) { 274 BasicBlock *Header = L->getHeader(); 275 TerminatorInst *HeaderTerm = Header->getTerminator(); 276 277 BasicBlock *LoopExitBB = 0; 278 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 279 // If the header block doesn't end with a conditional branch on Cond, we 280 // can't handle it. 281 if (!BI->isConditional() || BI->getCondition() != Cond) 282 return false; 283 284 // Check to see if a successor of the branch is guaranteed to go to the 285 // latch block or exit through a one exit block without having any 286 // side-effects. If so, determine the value of Cond that causes it to do 287 // this. 288 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { 289 if (Val) *Val = ConstantInt::getTrue(); 290 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { 291 if (Val) *Val = ConstantInt::getFalse(); 292 } 293 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 294 // If this isn't a switch on Cond, we can't handle it. 295 if (SI->getCondition() != Cond) return false; 296 297 // Check to see if a successor of the switch is guaranteed to go to the 298 // latch block or exit through a one exit block without having any 299 // side-effects. If so, determine the value of Cond that causes it to do 300 // this. Note that we can't trivially unswitch on the default case. 301 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 302 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) { 303 // Okay, we found a trivial case, remember the value that is trivial. 304 if (Val) *Val = SI->getCaseValue(i); 305 break; 306 } 307 } 308 309 // If we didn't find a single unique LoopExit block, or if the loop exit block 310 // contains phi nodes, this isn't trivial. 311 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 312 return false; // Can't handle this. 313 314 if (LoopExit) *LoopExit = LoopExitBB; 315 316 // We already know that nothing uses any scalar values defined inside of this 317 // loop. As such, we just have to check to see if this loop will execute any 318 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 319 // part of the loop that the code *would* execute. We already checked the 320 // tail, check the header now. 321 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 322 if (I->mayWriteToMemory()) 323 return false; 324 return true; 325} 326 327/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 328/// we choose to unswitch the specified loop on the specified value. 329/// 330unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { 331 // If the condition is trivial, always unswitch. There is no code growth for 332 // this case. 333 if (IsTrivialUnswitchCondition(L, LIC)) 334 return 0; 335 336 // FIXME: This is really overly conservative. However, more liberal 337 // estimations have thus far resulted in excessive unswitching, which is bad 338 // both in compile time and in code size. This should be replaced once 339 // someone figures out how a good estimation. 340 return L->getBlocks().size(); 341 342 unsigned Cost = 0; 343 // FIXME: this is brain dead. It should take into consideration code 344 // shrinkage. 345 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 346 I != E; ++I) { 347 BasicBlock *BB = *I; 348 // Do not include empty blocks in the cost calculation. This happen due to 349 // loop canonicalization and will be removed. 350 if (BB->begin() == BasicBlock::iterator(BB->getTerminator())) 351 continue; 352 353 // Count basic blocks. 354 ++Cost; 355 } 356 357 return Cost; 358} 359 360/// UnswitchIfProfitable - We have found that we can unswitch L when 361/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 362/// unswitch the loop, reprocess the pieces, then return true. 363bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ 364 // Check to see if it would be profitable to unswitch this loop. 365 unsigned Cost = getLoopUnswitchCost(L, LoopCond); 366 367 // Do not do non-trivial unswitch while optimizing for size. 368 if (Cost && OptimizeForSize) 369 return false; 370 371 if (Cost > Threshold) { 372 // FIXME: this should estimate growth by the amount of code shared by the 373 // resultant unswitched loops. 374 // 375 DOUT << "NOT unswitching loop %" 376 << L->getHeader()->getName() << ", cost too high: " 377 << L->getBlocks().size() << "\n"; 378 return false; 379 } 380 381 // If this is a trivial condition to unswitch (which results in no code 382 // duplication), do it now. 383 Constant *CondVal; 384 BasicBlock *ExitBlock; 385 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) { 386 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock); 387 } else { 388 UnswitchNontrivialCondition(LoopCond, Val, L); 389 } 390 391 return true; 392} 393 394/// SplitBlock - Split the specified block at the specified instruction - every 395/// thing before SplitPt stays in Old and everything starting with SplitPt moves 396/// to a new block. The two blocks are joined by an unconditional branch and 397/// the loop info is updated. 398/// 399BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) { 400 BasicBlock::iterator SplitIt = SplitPt; 401 while (isa<PHINode>(SplitIt)) 402 ++SplitIt; 403 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split"); 404 405 // The new block lives in whichever loop the old one did. 406 if (Loop *L = LI->getLoopFor(Old)) 407 L->addBasicBlockToLoop(New, *LI); 408 409 return New; 410} 411 412 413BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) { 414 TerminatorInst *LatchTerm = BB->getTerminator(); 415 unsigned SuccNum = 0; 416 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) { 417 assert(i != e && "Didn't find edge?"); 418 if (LatchTerm->getSuccessor(i) == Succ) { 419 SuccNum = i; 420 break; 421 } 422 } 423 424 // If this is a critical edge, let SplitCriticalEdge do it. 425 Loop *OrigDestBBL = LI->getLoopFor(BB->getTerminator()->getSuccessor(SuccNum)); 426 if (SplitCriticalEdge(BB->getTerminator(), SuccNum)) { 427 BasicBlock *NewBB = LatchTerm->getSuccessor(SuccNum); 428 429 Loop *BBL = LI->getLoopFor(BB); 430 if (!BBL || !OrigDestBBL) 431 return NewBB; 432 433 // If edge is inside a loop then NewBB is part of same loop. 434 if (BBL == OrigDestBBL) 435 BBL->addBasicBlockToLoop(NewBB, *LI); 436 // If edge is entering loop then NewBB is part of outer loop. 437 else if (BBL->contains(OrigDestBBL->getHeader())) 438 BBL->addBasicBlockToLoop(NewBB, *LI); 439 // If edge is from an inner loop to outer loop then NewBB is part 440 // of outer loop. 441 else if (OrigDestBBL->contains(BBL->getHeader())) 442 OrigDestBBL->addBasicBlockToLoop(NewBB, *LI); 443 // Else edge is connecting two loops and NewBB is part of their parent loop 444 else if (Loop *PL = OrigDestBBL->getParentLoop()) 445 PL->addBasicBlockToLoop(NewBB, *LI); 446 447 return NewBB; 448 } 449 450 // If the edge isn't critical, then BB has a single successor or Succ has a 451 // single pred. Split the block. 452 BasicBlock::iterator SplitPoint; 453 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 454 // If the successor only has a single pred, split the top of the successor 455 // block. 456 assert(SP == BB && "CFG broken"); 457 return SplitBlock(Succ, Succ->begin()); 458 } else { 459 // Otherwise, if BB has a single successor, split it at the bottom of the 460 // block. 461 assert(BB->getTerminator()->getNumSuccessors() == 1 && 462 "Should have a single succ!"); 463 return SplitBlock(BB, BB->getTerminator()); 464 } 465} 466 467 468 469// RemapInstruction - Convert the instruction operands from referencing the 470// current values into those specified by ValueMap. 471// 472static inline void RemapInstruction(Instruction *I, 473 DenseMap<const Value *, Value*> &ValueMap) { 474 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 475 Value *Op = I->getOperand(op); 476 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 477 if (It != ValueMap.end()) Op = It->second; 478 I->setOperand(op, Op); 479 } 480} 481 482/// CloneLoop - Recursively clone the specified loop and all of its children, 483/// mapping the blocks with the specified map. 484static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 485 LoopInfo *LI, LPPassManager *LPM) { 486 Loop *New = new Loop(); 487 488 LPM->insertLoop(New, PL); 489 490 // Add all of the blocks in L to the new loop. 491 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 492 I != E; ++I) 493 if (LI->getLoopFor(*I) == L) 494 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 495 496 // Add all of the subloops to the new loop. 497 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 498 CloneLoop(*I, New, VM, LI, LPM); 499 500 return New; 501} 502 503/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 504/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 505/// code immediately before InsertPt. 506static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 507 BasicBlock *TrueDest, 508 BasicBlock *FalseDest, 509 Instruction *InsertPt) { 510 // Insert a conditional branch on LIC to the two preheaders. The original 511 // code is the true version and the new code is the false version. 512 Value *BranchVal = LIC; 513 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty) 514 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); 515 else if (Val != ConstantInt::getTrue()) 516 // We want to enter the new loop when the condition is true. 517 std::swap(TrueDest, FalseDest); 518 519 // Insert the new branch. 520 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); 521} 522 523 524/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 525/// condition in it (a cond branch from its header block to its latch block, 526/// where the path through the loop that doesn't execute its body has no 527/// side-effects), unswitch it. This doesn't involve any code duplication, just 528/// moving the conditional branch outside of the loop and updating loop info. 529void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 530 Constant *Val, 531 BasicBlock *ExitBlock) { 532 DOUT << "loop-unswitch: Trivial-Unswitch loop %" 533 << L->getHeader()->getName() << " [" << L->getBlocks().size() 534 << " blocks] in Function " << L->getHeader()->getParent()->getName() 535 << " on cond: " << *Val << " == " << *Cond << "\n"; 536 537 // First step, split the preheader, so that we know that there is a safe place 538 // to insert the conditional branch. We will change 'OrigPH' to have a 539 // conditional branch on Cond. 540 BasicBlock *OrigPH = L->getLoopPreheader(); 541 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); 542 543 // Now that we have a place to insert the conditional branch, create a place 544 // to branch to: this is the exit block out of the loop that we should 545 // short-circuit to. 546 547 // Split this block now, so that the loop maintains its exit block, and so 548 // that the jump from the preheader can execute the contents of the exit block 549 // without actually branching to it (the exit block should be dominated by the 550 // loop header, not the preheader). 551 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 552 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); 553 554 // Okay, now we have a position to branch from and a position to branch to, 555 // insert the new conditional branch. 556 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 557 OrigPH->getTerminator()); 558 OrigPH->getTerminator()->eraseFromParent(); 559 560 // We need to reprocess this loop, it could be unswitched again. 561 LPM->redoLoop(L); 562 563 // Now that we know that the loop is never entered when this condition is a 564 // particular value, rewrite the loop with this info. We know that this will 565 // at least eliminate the old branch. 566 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 567 ++NumTrivial; 568} 569 570 571/// VersionLoop - We determined that the loop is profitable to unswitch when LIC 572/// equal Val. Split it into loop versions and test the condition outside of 573/// either loop. Return the loops created as Out1/Out2. 574void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 575 Loop *L) { 576 Function *F = L->getHeader()->getParent(); 577 DOUT << "loop-unswitch: Unswitching loop %" 578 << L->getHeader()->getName() << " [" << L->getBlocks().size() 579 << " blocks] in Function " << F->getName() 580 << " when '" << *Val << "' == " << *LIC << "\n"; 581 582 // LoopBlocks contains all of the basic blocks of the loop, including the 583 // preheader of the loop, the body of the loop, and the exit blocks of the 584 // loop, in that order. 585 std::vector<BasicBlock*> LoopBlocks; 586 587 // First step, split the preheader and exit blocks, and add these blocks to 588 // the LoopBlocks list. 589 BasicBlock *OrigPreheader = L->getLoopPreheader(); 590 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); 591 592 // We want the loop to come after the preheader, but before the exit blocks. 593 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 594 595 std::vector<BasicBlock*> ExitBlocks; 596 L->getUniqueExitBlocks(ExitBlocks); 597 598 // Split all of the edges from inside the loop to their exit blocks. Update 599 // the appropriate Phi nodes as we do so. 600 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 601 BasicBlock *ExitBlock = ExitBlocks[i]; 602 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 603 604 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 605 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock); 606 BasicBlock* StartBlock = Preds[j]; 607 BasicBlock* EndBlock; 608 if (MiddleBlock->getSinglePredecessor() == ExitBlock) { 609 EndBlock = MiddleBlock; 610 MiddleBlock = EndBlock->getSinglePredecessor();; 611 } else { 612 EndBlock = ExitBlock; 613 } 614 615 std::set<PHINode*> InsertedPHIs; 616 PHINode* OldLCSSA = 0; 617 for (BasicBlock::iterator I = EndBlock->begin(); 618 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) { 619 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock); 620 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(), 621 OldLCSSA->getName() + ".us-lcssa", 622 MiddleBlock->getTerminator()); 623 NewLCSSA->addIncoming(OldValue, StartBlock); 624 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock), 625 NewLCSSA); 626 InsertedPHIs.insert(NewLCSSA); 627 } 628 629 BasicBlock::iterator InsertPt = EndBlock->begin(); 630 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt; 631 for (BasicBlock::iterator I = MiddleBlock->begin(); 632 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0; 633 ++I) { 634 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(), 635 OldLCSSA->getName() + ".us-lcssa", 636 InsertPt); 637 OldLCSSA->replaceAllUsesWith(NewLCSSA); 638 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); 639 } 640 } 641 } 642 643 // The exit blocks may have been changed due to edge splitting, recompute. 644 ExitBlocks.clear(); 645 L->getUniqueExitBlocks(ExitBlocks); 646 647 // Add exit blocks to the loop blocks. 648 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 649 650 // Next step, clone all of the basic blocks that make up the loop (including 651 // the loop preheader and exit blocks), keeping track of the mapping between 652 // the instructions and blocks. 653 std::vector<BasicBlock*> NewBlocks; 654 NewBlocks.reserve(LoopBlocks.size()); 655 DenseMap<const Value*, Value*> ValueMap; 656 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 657 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 658 NewBlocks.push_back(New); 659 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 660 } 661 662 // Splice the newly inserted blocks into the function right before the 663 // original preheader. 664 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 665 NewBlocks[0], F->end()); 666 667 // Now we create the new Loop object for the versioned loop. 668 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 669 Loop *ParentLoop = L->getParentLoop(); 670 if (ParentLoop) { 671 // Make sure to add the cloned preheader and exit blocks to the parent loop 672 // as well. 673 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 674 } 675 676 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 677 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 678 // The new exit block should be in the same loop as the old one. 679 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 680 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 681 682 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 683 "Exit block should have been split to have one successor!"); 684 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 685 686 // If the successor of the exit block had PHI nodes, add an entry for 687 // NewExit. 688 PHINode *PN; 689 for (BasicBlock::iterator I = ExitSucc->begin(); 690 (PN = dyn_cast<PHINode>(I)); ++I) { 691 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 692 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 693 if (It != ValueMap.end()) V = It->second; 694 PN->addIncoming(V, NewExit); 695 } 696 } 697 698 // Rewrite the code to refer to itself. 699 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 700 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 701 E = NewBlocks[i]->end(); I != E; ++I) 702 RemapInstruction(I, ValueMap); 703 704 // Rewrite the original preheader to select between versions of the loop. 705 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 706 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 707 "Preheader splitting did not work correctly!"); 708 709 // Emit the new branch that selects between the two versions of this loop. 710 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 711 OldBR->eraseFromParent(); 712 713 LoopProcessWorklist.push_back(NewLoop); 714 LPM->redoLoop(L); 715 716 // Now we rewrite the original code to know that the condition is true and the 717 // new code to know that the condition is false. 718 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 719 720 // It's possible that simplifying one loop could cause the other to be 721 // deleted. If so, don't simplify it. 722 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 723 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 724} 725 726/// RemoveFromWorklist - Remove all instances of I from the worklist vector 727/// specified. 728static void RemoveFromWorklist(Instruction *I, 729 std::vector<Instruction*> &Worklist) { 730 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 731 Worklist.end(), I); 732 while (WI != Worklist.end()) { 733 unsigned Offset = WI-Worklist.begin(); 734 Worklist.erase(WI); 735 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 736 } 737} 738 739/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 740/// program, replacing all uses with V and update the worklist. 741static void ReplaceUsesOfWith(Instruction *I, Value *V, 742 std::vector<Instruction*> &Worklist) { 743 DOUT << "Replace with '" << *V << "': " << *I; 744 745 // Add uses to the worklist, which may be dead now. 746 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 747 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 748 Worklist.push_back(Use); 749 750 // Add users to the worklist which may be simplified now. 751 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 752 UI != E; ++UI) 753 Worklist.push_back(cast<Instruction>(*UI)); 754 I->replaceAllUsesWith(V); 755 I->eraseFromParent(); 756 RemoveFromWorklist(I, Worklist); 757 ++NumSimplify; 758} 759 760/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 761/// information, and remove any dead successors it has. 762/// 763void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 764 std::vector<Instruction*> &Worklist) { 765 if (pred_begin(BB) != pred_end(BB)) { 766 // This block isn't dead, since an edge to BB was just removed, see if there 767 // are any easy simplifications we can do now. 768 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 769 // If it has one pred, fold phi nodes in BB. 770 while (isa<PHINode>(BB->begin())) 771 ReplaceUsesOfWith(BB->begin(), 772 cast<PHINode>(BB->begin())->getIncomingValue(0), 773 Worklist); 774 775 // If this is the header of a loop and the only pred is the latch, we now 776 // have an unreachable loop. 777 if (Loop *L = LI->getLoopFor(BB)) 778 if (L->getHeader() == BB && L->contains(Pred)) { 779 // Remove the branch from the latch to the header block, this makes 780 // the header dead, which will make the latch dead (because the header 781 // dominates the latch). 782 Pred->getTerminator()->eraseFromParent(); 783 new UnreachableInst(Pred); 784 785 // The loop is now broken, remove it from LI. 786 RemoveLoopFromHierarchy(L); 787 788 // Reprocess the header, which now IS dead. 789 RemoveBlockIfDead(BB, Worklist); 790 return; 791 } 792 793 // If pred ends in a uncond branch, add uncond branch to worklist so that 794 // the two blocks will get merged. 795 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 796 if (BI->isUnconditional()) 797 Worklist.push_back(BI); 798 } 799 return; 800 } 801 802 DOUT << "Nuking dead block: " << *BB; 803 804 // Remove the instructions in the basic block from the worklist. 805 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 806 RemoveFromWorklist(I, Worklist); 807 808 // Anything that uses the instructions in this basic block should have their 809 // uses replaced with undefs. 810 if (!I->use_empty()) 811 I->replaceAllUsesWith(UndefValue::get(I->getType())); 812 } 813 814 // If this is the edge to the header block for a loop, remove the loop and 815 // promote all subloops. 816 if (Loop *BBLoop = LI->getLoopFor(BB)) { 817 if (BBLoop->getLoopLatch() == BB) 818 RemoveLoopFromHierarchy(BBLoop); 819 } 820 821 // Remove the block from the loop info, which removes it from any loops it 822 // was in. 823 LI->removeBlock(BB); 824 825 826 // Remove phi node entries in successors for this block. 827 TerminatorInst *TI = BB->getTerminator(); 828 std::vector<BasicBlock*> Succs; 829 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 830 Succs.push_back(TI->getSuccessor(i)); 831 TI->getSuccessor(i)->removePredecessor(BB); 832 } 833 834 // Unique the successors, remove anything with multiple uses. 835 std::sort(Succs.begin(), Succs.end()); 836 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 837 838 // Remove the basic block, including all of the instructions contained in it. 839 BB->eraseFromParent(); 840 841 // Remove successor blocks here that are not dead, so that we know we only 842 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 843 // then getting removed before we revisit them, which is badness. 844 // 845 for (unsigned i = 0; i != Succs.size(); ++i) 846 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 847 // One exception is loop headers. If this block was the preheader for a 848 // loop, then we DO want to visit the loop so the loop gets deleted. 849 // We know that if the successor is a loop header, that this loop had to 850 // be the preheader: the case where this was the latch block was handled 851 // above and headers can only have two predecessors. 852 if (!LI->isLoopHeader(Succs[i])) { 853 Succs.erase(Succs.begin()+i); 854 --i; 855 } 856 } 857 858 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 859 RemoveBlockIfDead(Succs[i], Worklist); 860} 861 862/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 863/// become unwrapped, either because the backedge was deleted, or because the 864/// edge into the header was removed. If the edge into the header from the 865/// latch block was removed, the loop is unwrapped but subloops are still alive, 866/// so they just reparent loops. If the loops are actually dead, they will be 867/// removed later. 868void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 869 LPM->deleteLoopFromQueue(L); 870 RemoveLoopFromWorklist(L); 871} 872 873 874 875// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 876// the value specified by Val in the specified loop, or we know it does NOT have 877// that value. Rewrite any uses of LIC or of properties correlated to it. 878void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 879 Constant *Val, 880 bool IsEqual) { 881 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 882 883 // FIXME: Support correlated properties, like: 884 // for (...) 885 // if (li1 < li2) 886 // ... 887 // if (li1 > li2) 888 // ... 889 890 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 891 // selects, switches. 892 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 893 std::vector<Instruction*> Worklist; 894 895 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 896 // in the loop with the appropriate one directly. 897 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) { 898 Value *Replacement; 899 if (IsEqual) 900 Replacement = Val; 901 else 902 Replacement = ConstantInt::get(Type::Int1Ty, 903 !cast<ConstantInt>(Val)->getZExtValue()); 904 905 for (unsigned i = 0, e = Users.size(); i != e; ++i) 906 if (Instruction *U = cast<Instruction>(Users[i])) { 907 if (!L->contains(U->getParent())) 908 continue; 909 U->replaceUsesOfWith(LIC, Replacement); 910 Worklist.push_back(U); 911 } 912 } else { 913 // Otherwise, we don't know the precise value of LIC, but we do know that it 914 // is certainly NOT "Val". As such, simplify any uses in the loop that we 915 // can. This case occurs when we unswitch switch statements. 916 for (unsigned i = 0, e = Users.size(); i != e; ++i) 917 if (Instruction *U = cast<Instruction>(Users[i])) { 918 if (!L->contains(U->getParent())) 919 continue; 920 921 Worklist.push_back(U); 922 923 // If we know that LIC is not Val, use this info to simplify code. 924 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 925 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 926 if (SI->getCaseValue(i) == Val) { 927 // Found a dead case value. Don't remove PHI nodes in the 928 // successor if they become single-entry, those PHI nodes may 929 // be in the Users list. 930 931 // FIXME: This is a hack. We need to keep the successor around 932 // and hooked up so as to preserve the loop structure, because 933 // trying to update it is complicated. So instead we preserve the 934 // loop structure and put the block on an dead code path. 935 936 BasicBlock* Old = SI->getParent(); 937 BasicBlock* Split = SplitBlock(Old, SI); 938 939 Instruction* OldTerm = Old->getTerminator(); 940 new BranchInst(Split, SI->getSuccessor(i), 941 ConstantInt::getTrue(), OldTerm); 942 943 Old->getTerminator()->eraseFromParent(); 944 945 946 PHINode *PN; 947 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); 948 (PN = dyn_cast<PHINode>(II)); ++II) { 949 Value *InVal = PN->removeIncomingValue(Split, false); 950 PN->addIncoming(InVal, Old); 951 } 952 953 SI->removeCase(i); 954 break; 955 } 956 } 957 } 958 959 // TODO: We could do other simplifications, for example, turning 960 // LIC == Val -> false. 961 } 962 } 963 964 SimplifyCode(Worklist); 965} 966 967/// SimplifyCode - Okay, now that we have simplified some instructions in the 968/// loop, walk over it and constant prop, dce, and fold control flow where 969/// possible. Note that this is effectively a very simple loop-structure-aware 970/// optimizer. During processing of this loop, L could very well be deleted, so 971/// it must not be used. 972/// 973/// FIXME: When the loop optimizer is more mature, separate this out to a new 974/// pass. 975/// 976void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) { 977 while (!Worklist.empty()) { 978 Instruction *I = Worklist.back(); 979 Worklist.pop_back(); 980 981 // Simple constant folding. 982 if (Constant *C = ConstantFoldInstruction(I)) { 983 ReplaceUsesOfWith(I, C, Worklist); 984 continue; 985 } 986 987 // Simple DCE. 988 if (isInstructionTriviallyDead(I)) { 989 DOUT << "Remove dead instruction '" << *I; 990 991 // Add uses to the worklist, which may be dead now. 992 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 993 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 994 Worklist.push_back(Use); 995 I->eraseFromParent(); 996 RemoveFromWorklist(I, Worklist); 997 ++NumSimplify; 998 continue; 999 } 1000 1001 // Special case hacks that appear commonly in unswitched code. 1002 switch (I->getOpcode()) { 1003 case Instruction::Select: 1004 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 1005 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist); 1006 continue; 1007 } 1008 break; 1009 case Instruction::And: 1010 if (isa<ConstantInt>(I->getOperand(0)) && 1011 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1012 cast<BinaryOperator>(I)->swapOperands(); 1013 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1014 if (CB->getType() == Type::Int1Ty) { 1015 if (CB->isOne()) // X & 1 -> X 1016 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1017 else // X & 0 -> 0 1018 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1019 continue; 1020 } 1021 break; 1022 case Instruction::Or: 1023 if (isa<ConstantInt>(I->getOperand(0)) && 1024 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1025 cast<BinaryOperator>(I)->swapOperands(); 1026 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1027 if (CB->getType() == Type::Int1Ty) { 1028 if (CB->isOne()) // X | 1 -> 1 1029 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1030 else // X | 0 -> X 1031 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1032 continue; 1033 } 1034 break; 1035 case Instruction::Br: { 1036 BranchInst *BI = cast<BranchInst>(I); 1037 if (BI->isUnconditional()) { 1038 // If BI's parent is the only pred of the successor, fold the two blocks 1039 // together. 1040 BasicBlock *Pred = BI->getParent(); 1041 BasicBlock *Succ = BI->getSuccessor(0); 1042 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1043 if (!SinglePred) continue; // Nothing to do. 1044 assert(SinglePred == Pred && "CFG broken"); 1045 1046 DOUT << "Merging blocks: " << Pred->getName() << " <- " 1047 << Succ->getName() << "\n"; 1048 1049 // Resolve any single entry PHI nodes in Succ. 1050 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1051 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); 1052 1053 // Move all of the successor contents from Succ to Pred. 1054 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1055 Succ->end()); 1056 BI->eraseFromParent(); 1057 RemoveFromWorklist(BI, Worklist); 1058 1059 // If Succ has any successors with PHI nodes, update them to have 1060 // entries coming from Pred instead of Succ. 1061 Succ->replaceAllUsesWith(Pred); 1062 1063 // Remove Succ from the loop tree. 1064 LI->removeBlock(Succ); 1065 Succ->eraseFromParent(); 1066 ++NumSimplify; 1067 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1068 // Conditional branch. Turn it into an unconditional branch, then 1069 // remove dead blocks. 1070 break; // FIXME: Enable. 1071 1072 DOUT << "Folded branch: " << *BI; 1073 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1074 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1075 DeadSucc->removePredecessor(BI->getParent(), true); 1076 Worklist.push_back(new BranchInst(LiveSucc, BI)); 1077 BI->eraseFromParent(); 1078 RemoveFromWorklist(BI, Worklist); 1079 ++NumSimplify; 1080 1081 RemoveBlockIfDead(DeadSucc, Worklist); 1082 } 1083 break; 1084 } 1085 } 1086 } 1087} 1088