LoopUnswitch.cpp revision 1997473cf72957d0e70322e2fe6fe2ab141c58a6
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 public: 72 static char ID; // Pass ID, replacement for typeid 73 LoopUnswitch() : LoopPass((intptr_t)&ID) {} 74 75 bool runOnLoop(Loop *L, LPPassManager &LPM); 76 77 /// This transformation requires natural loop information & requires that 78 /// loop preheaders be inserted into the CFG... 79 /// 80 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 81 AU.addRequiredID(LoopSimplifyID); 82 AU.addPreservedID(LoopSimplifyID); 83 AU.addRequired<LoopInfo>(); 84 AU.addPreserved<LoopInfo>(); 85 AU.addRequiredID(LCSSAID); 86 AU.addPreservedID(LCSSAID); 87 } 88 89 private: 90 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 91 /// remove it. 92 void RemoveLoopFromWorklist(Loop *L) { 93 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 94 LoopProcessWorklist.end(), L); 95 if (I != LoopProcessWorklist.end()) 96 LoopProcessWorklist.erase(I); 97 } 98 99 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L); 100 unsigned getLoopUnswitchCost(Loop *L, Value *LIC); 101 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 102 BasicBlock *ExitBlock); 103 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 104 BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To); 105 BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt); 106 107 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 108 Constant *Val, bool isEqual); 109 110 void SimplifyCode(std::vector<Instruction*> &Worklist); 111 void RemoveBlockIfDead(BasicBlock *BB, 112 std::vector<Instruction*> &Worklist); 113 void RemoveLoopFromHierarchy(Loop *L); 114 }; 115 char LoopUnswitch::ID = 0; 116 RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 117} 118 119LoopPass *llvm::createLoopUnswitchPass() { return new LoopUnswitch(); } 120 121/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 122/// invariant in the loop, or has an invariant piece, return the invariant. 123/// Otherwise, return null. 124static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 125 // Constants should be folded, not unswitched on! 126 if (isa<Constant>(Cond)) return false; 127 128 // TODO: Handle: br (VARIANT|INVARIANT). 129 // TODO: Hoist simple expressions out of loops. 130 if (L->isLoopInvariant(Cond)) return Cond; 131 132 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 133 if (BO->getOpcode() == Instruction::And || 134 BO->getOpcode() == Instruction::Or) { 135 // If either the left or right side is invariant, we can unswitch on this, 136 // which will cause the branch to go away in one loop and the condition to 137 // simplify in the other one. 138 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 139 return LHS; 140 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 141 return RHS; 142 } 143 144 return 0; 145} 146 147bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { 148 assert(L->isLCSSAForm()); 149 LI = &getAnalysis<LoopInfo>(); 150 LPM = &LPM_Ref; 151 bool Changed = false; 152 153 // Loop over all of the basic blocks in the loop. If we find an interior 154 // block that is branching on a loop-invariant condition, we can unswitch this 155 // loop. 156 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 157 I != E; ++I) { 158 TerminatorInst *TI = (*I)->getTerminator(); 159 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 160 // If this isn't branching on an invariant condition, we can't unswitch 161 // it. 162 if (BI->isConditional()) { 163 // See if this, or some part of it, is loop invariant. If so, we can 164 // unswitch on it if we desire. 165 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); 166 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 167 L)) { 168 ++NumBranches; 169 return true; 170 } 171 } 172 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 173 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 174 if (LoopCond && SI->getNumCases() > 1) { 175 // Find a value to unswitch on: 176 // FIXME: this should chose the most expensive case! 177 Constant *UnswitchVal = SI->getCaseValue(1); 178 // Do not process same value again and again. 179 if (!UnswitchedVals.insert(UnswitchVal)) 180 continue; 181 182 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { 183 ++NumSwitches; 184 return true; 185 } 186 } 187 } 188 189 // Scan the instructions to check for unswitchable values. 190 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 191 BBI != E; ++BBI) 192 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 193 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 194 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 195 L)) { 196 ++NumSelects; 197 return true; 198 } 199 } 200 } 201 202 assert(L->isLCSSAForm()); 203 204 return Changed; 205} 206 207/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 208/// 1. Exit the loop with no side effects. 209/// 2. Branch to the latch block with no side-effects. 210/// 211/// If these conditions are true, we return true and set ExitBB to the block we 212/// exit through. 213/// 214static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 215 BasicBlock *&ExitBB, 216 std::set<BasicBlock*> &Visited) { 217 if (!Visited.insert(BB).second) { 218 // Already visited and Ok, end of recursion. 219 return true; 220 } else if (!L->contains(BB)) { 221 // Otherwise, this is a loop exit, this is fine so long as this is the 222 // first exit. 223 if (ExitBB != 0) return false; 224 ExitBB = BB; 225 return true; 226 } 227 228 // Otherwise, this is an unvisited intra-loop node. Check all successors. 229 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 230 // Check to see if the successor is a trivial loop exit. 231 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 232 return false; 233 } 234 235 // Okay, everything after this looks good, check to make sure that this block 236 // doesn't include any side effects. 237 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 238 if (I->mayWriteToMemory()) 239 return false; 240 241 return true; 242} 243 244/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 245/// leads to an exit from the specified loop, and has no side-effects in the 246/// process. If so, return the block that is exited to, otherwise return null. 247static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 248 std::set<BasicBlock*> Visited; 249 Visited.insert(L->getHeader()); // Branches to header are ok. 250 BasicBlock *ExitBB = 0; 251 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 252 return ExitBB; 253 return 0; 254} 255 256/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 257/// trivial: that is, that the condition controls whether or not the loop does 258/// anything at all. If this is a trivial condition, unswitching produces no 259/// code duplications (equivalently, it produces a simpler loop and a new empty 260/// loop, which gets deleted). 261/// 262/// If this is a trivial condition, return true, otherwise return false. When 263/// returning true, this sets Cond and Val to the condition that controls the 264/// trivial condition: when Cond dynamically equals Val, the loop is known to 265/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 266/// Cond == Val. 267/// 268static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, 269 BasicBlock **LoopExit = 0) { 270 BasicBlock *Header = L->getHeader(); 271 TerminatorInst *HeaderTerm = Header->getTerminator(); 272 273 BasicBlock *LoopExitBB = 0; 274 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 275 // If the header block doesn't end with a conditional branch on Cond, we 276 // can't handle it. 277 if (!BI->isConditional() || BI->getCondition() != Cond) 278 return false; 279 280 // Check to see if a successor of the branch is guaranteed to go to the 281 // latch block or exit through a one exit block without having any 282 // side-effects. If so, determine the value of Cond that causes it to do 283 // this. 284 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { 285 if (Val) *Val = ConstantInt::getTrue(); 286 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { 287 if (Val) *Val = ConstantInt::getFalse(); 288 } 289 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 290 // If this isn't a switch on Cond, we can't handle it. 291 if (SI->getCondition() != Cond) return false; 292 293 // Check to see if a successor of the switch is guaranteed to go to the 294 // latch block or exit through a one exit block without having any 295 // side-effects. If so, determine the value of Cond that causes it to do 296 // this. Note that we can't trivially unswitch on the default case. 297 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 298 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) { 299 // Okay, we found a trivial case, remember the value that is trivial. 300 if (Val) *Val = SI->getCaseValue(i); 301 break; 302 } 303 } 304 305 // If we didn't find a single unique LoopExit block, or if the loop exit block 306 // contains phi nodes, this isn't trivial. 307 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 308 return false; // Can't handle this. 309 310 if (LoopExit) *LoopExit = LoopExitBB; 311 312 // We already know that nothing uses any scalar values defined inside of this 313 // loop. As such, we just have to check to see if this loop will execute any 314 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 315 // part of the loop that the code *would* execute. We already checked the 316 // tail, check the header now. 317 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 318 if (I->mayWriteToMemory()) 319 return false; 320 return true; 321} 322 323/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 324/// we choose to unswitch the specified loop on the specified value. 325/// 326unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { 327 // If the condition is trivial, always unswitch. There is no code growth for 328 // this case. 329 if (IsTrivialUnswitchCondition(L, LIC)) 330 return 0; 331 332 // FIXME: This is really overly conservative. However, more liberal 333 // estimations have thus far resulted in excessive unswitching, which is bad 334 // both in compile time and in code size. This should be replaced once 335 // someone figures out how a good estimation. 336 return L->getBlocks().size(); 337 338 unsigned Cost = 0; 339 // FIXME: this is brain dead. It should take into consideration code 340 // shrinkage. 341 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 342 I != E; ++I) { 343 BasicBlock *BB = *I; 344 // Do not include empty blocks in the cost calculation. This happen due to 345 // loop canonicalization and will be removed. 346 if (BB->begin() == BasicBlock::iterator(BB->getTerminator())) 347 continue; 348 349 // Count basic blocks. 350 ++Cost; 351 } 352 353 return Cost; 354} 355 356/// UnswitchIfProfitable - We have found that we can unswitch L when 357/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 358/// unswitch the loop, reprocess the pieces, then return true. 359bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ 360 // Check to see if it would be profitable to unswitch this loop. 361 unsigned Cost = getLoopUnswitchCost(L, LoopCond); 362 if (Cost > Threshold) { 363 // FIXME: this should estimate growth by the amount of code shared by the 364 // resultant unswitched loops. 365 // 366 DOUT << "NOT unswitching loop %" 367 << L->getHeader()->getName() << ", cost too high: " 368 << L->getBlocks().size() << "\n"; 369 return false; 370 } 371 372 // If this is a trivial condition to unswitch (which results in no code 373 // duplication), do it now. 374 Constant *CondVal; 375 BasicBlock *ExitBlock; 376 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) { 377 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock); 378 } else { 379 UnswitchNontrivialCondition(LoopCond, Val, L); 380 } 381 382 return true; 383} 384 385/// SplitBlock - Split the specified block at the specified instruction - every 386/// thing before SplitPt stays in Old and everything starting with SplitPt moves 387/// to a new block. The two blocks are joined by an unconditional branch and 388/// the loop info is updated. 389/// 390BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) { 391 BasicBlock::iterator SplitIt = SplitPt; 392 while (isa<PHINode>(SplitIt)) 393 ++SplitIt; 394 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split"); 395 396 // The new block lives in whichever loop the old one did. 397 if (Loop *L = LI->getLoopFor(Old)) 398 L->addBasicBlockToLoop(New, *LI); 399 400 return New; 401} 402 403 404BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) { 405 TerminatorInst *LatchTerm = BB->getTerminator(); 406 unsigned SuccNum = 0; 407 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) { 408 assert(i != e && "Didn't find edge?"); 409 if (LatchTerm->getSuccessor(i) == Succ) { 410 SuccNum = i; 411 break; 412 } 413 } 414 415 // If this is a critical edge, let SplitCriticalEdge do it. 416 if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this)) 417 return LatchTerm->getSuccessor(SuccNum); 418 419 // If the edge isn't critical, then BB has a single successor or Succ has a 420 // single pred. Split the block. 421 BasicBlock::iterator SplitPoint; 422 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 423 // If the successor only has a single pred, split the top of the successor 424 // block. 425 assert(SP == BB && "CFG broken"); 426 return SplitBlock(Succ, Succ->begin()); 427 } else { 428 // Otherwise, if BB has a single successor, split it at the bottom of the 429 // block. 430 assert(BB->getTerminator()->getNumSuccessors() == 1 && 431 "Should have a single succ!"); 432 return SplitBlock(BB, BB->getTerminator()); 433 } 434} 435 436 437 438// RemapInstruction - Convert the instruction operands from referencing the 439// current values into those specified by ValueMap. 440// 441static inline void RemapInstruction(Instruction *I, 442 DenseMap<const Value *, Value*> &ValueMap) { 443 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 444 Value *Op = I->getOperand(op); 445 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 446 if (It != ValueMap.end()) Op = It->second; 447 I->setOperand(op, Op); 448 } 449} 450 451/// CloneLoop - Recursively clone the specified loop and all of its children, 452/// mapping the blocks with the specified map. 453static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 454 LoopInfo *LI, LPPassManager *LPM) { 455 Loop *New = new Loop(); 456 457 LPM->insertLoop(New, PL); 458 459 // Add all of the blocks in L to the new loop. 460 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 461 I != E; ++I) 462 if (LI->getLoopFor(*I) == L) 463 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 464 465 // Add all of the subloops to the new loop. 466 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 467 CloneLoop(*I, New, VM, LI, LPM); 468 469 return New; 470} 471 472/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 473/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 474/// code immediately before InsertPt. 475static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 476 BasicBlock *TrueDest, 477 BasicBlock *FalseDest, 478 Instruction *InsertPt) { 479 // Insert a conditional branch on LIC to the two preheaders. The original 480 // code is the true version and the new code is the false version. 481 Value *BranchVal = LIC; 482 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty) 483 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); 484 else if (Val != ConstantInt::getTrue()) 485 // We want to enter the new loop when the condition is true. 486 std::swap(TrueDest, FalseDest); 487 488 // Insert the new branch. 489 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); 490} 491 492 493/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 494/// condition in it (a cond branch from its header block to its latch block, 495/// where the path through the loop that doesn't execute its body has no 496/// side-effects), unswitch it. This doesn't involve any code duplication, just 497/// moving the conditional branch outside of the loop and updating loop info. 498void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 499 Constant *Val, 500 BasicBlock *ExitBlock) { 501 DOUT << "loop-unswitch: Trivial-Unswitch loop %" 502 << L->getHeader()->getName() << " [" << L->getBlocks().size() 503 << " blocks] in Function " << L->getHeader()->getParent()->getName() 504 << " on cond: " << *Val << " == " << *Cond << "\n"; 505 506 // First step, split the preheader, so that we know that there is a safe place 507 // to insert the conditional branch. We will change 'OrigPH' to have a 508 // conditional branch on Cond. 509 BasicBlock *OrigPH = L->getLoopPreheader(); 510 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); 511 512 // Now that we have a place to insert the conditional branch, create a place 513 // to branch to: this is the exit block out of the loop that we should 514 // short-circuit to. 515 516 // Split this block now, so that the loop maintains its exit block, and so 517 // that the jump from the preheader can execute the contents of the exit block 518 // without actually branching to it (the exit block should be dominated by the 519 // loop header, not the preheader). 520 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 521 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); 522 523 // Okay, now we have a position to branch from and a position to branch to, 524 // insert the new conditional branch. 525 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 526 OrigPH->getTerminator()); 527 OrigPH->getTerminator()->eraseFromParent(); 528 529 // We need to reprocess this loop, it could be unswitched again. 530 LPM->redoLoop(L); 531 532 // Now that we know that the loop is never entered when this condition is a 533 // particular value, rewrite the loop with this info. We know that this will 534 // at least eliminate the old branch. 535 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 536 ++NumTrivial; 537} 538 539 540/// VersionLoop - We determined that the loop is profitable to unswitch when LIC 541/// equal Val. Split it into loop versions and test the condition outside of 542/// either loop. Return the loops created as Out1/Out2. 543void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 544 Loop *L) { 545 Function *F = L->getHeader()->getParent(); 546 DOUT << "loop-unswitch: Unswitching loop %" 547 << L->getHeader()->getName() << " [" << L->getBlocks().size() 548 << " blocks] in Function " << F->getName() 549 << " when '" << *Val << "' == " << *LIC << "\n"; 550 551 // LoopBlocks contains all of the basic blocks of the loop, including the 552 // preheader of the loop, the body of the loop, and the exit blocks of the 553 // loop, in that order. 554 std::vector<BasicBlock*> LoopBlocks; 555 556 // First step, split the preheader and exit blocks, and add these blocks to 557 // the LoopBlocks list. 558 BasicBlock *OrigPreheader = L->getLoopPreheader(); 559 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); 560 561 // We want the loop to come after the preheader, but before the exit blocks. 562 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 563 564 std::vector<BasicBlock*> ExitBlocks; 565 L->getUniqueExitBlocks(ExitBlocks); 566 567 // Split all of the edges from inside the loop to their exit blocks. Update 568 // the appropriate Phi nodes as we do so. 569 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 570 BasicBlock *ExitBlock = ExitBlocks[i]; 571 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 572 573 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 574 assert(L->contains(Preds[j]) && 575 "All preds of loop exit blocks must be the same loop!"); 576 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock); 577 BasicBlock* StartBlock = Preds[j]; 578 BasicBlock* EndBlock; 579 if (MiddleBlock->getSinglePredecessor() == ExitBlock) { 580 EndBlock = MiddleBlock; 581 MiddleBlock = EndBlock->getSinglePredecessor();; 582 } else { 583 EndBlock = ExitBlock; 584 } 585 586 std::set<PHINode*> InsertedPHIs; 587 PHINode* OldLCSSA = 0; 588 for (BasicBlock::iterator I = EndBlock->begin(); 589 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) { 590 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock); 591 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(), 592 OldLCSSA->getName() + ".us-lcssa", 593 MiddleBlock->getTerminator()); 594 NewLCSSA->addIncoming(OldValue, StartBlock); 595 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock), 596 NewLCSSA); 597 InsertedPHIs.insert(NewLCSSA); 598 } 599 600 BasicBlock::iterator InsertPt = EndBlock->begin(); 601 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt; 602 for (BasicBlock::iterator I = MiddleBlock->begin(); 603 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0; 604 ++I) { 605 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(), 606 OldLCSSA->getName() + ".us-lcssa", 607 InsertPt); 608 OldLCSSA->replaceAllUsesWith(NewLCSSA); 609 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); 610 } 611 } 612 } 613 614 // The exit blocks may have been changed due to edge splitting, recompute. 615 ExitBlocks.clear(); 616 L->getUniqueExitBlocks(ExitBlocks); 617 618 // Add exit blocks to the loop blocks. 619 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 620 621 // Next step, clone all of the basic blocks that make up the loop (including 622 // the loop preheader and exit blocks), keeping track of the mapping between 623 // the instructions and blocks. 624 std::vector<BasicBlock*> NewBlocks; 625 NewBlocks.reserve(LoopBlocks.size()); 626 DenseMap<const Value*, Value*> ValueMap; 627 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 628 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 629 NewBlocks.push_back(New); 630 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 631 } 632 633 // Splice the newly inserted blocks into the function right before the 634 // original preheader. 635 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 636 NewBlocks[0], F->end()); 637 638 // Now we create the new Loop object for the versioned loop. 639 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 640 Loop *ParentLoop = L->getParentLoop(); 641 if (ParentLoop) { 642 // Make sure to add the cloned preheader and exit blocks to the parent loop 643 // as well. 644 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 645 } 646 647 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 648 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 649 // The new exit block should be in the same loop as the old one. 650 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 651 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 652 653 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 654 "Exit block should have been split to have one successor!"); 655 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 656 657 // If the successor of the exit block had PHI nodes, add an entry for 658 // NewExit. 659 PHINode *PN; 660 for (BasicBlock::iterator I = ExitSucc->begin(); 661 (PN = dyn_cast<PHINode>(I)); ++I) { 662 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 663 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 664 if (It != ValueMap.end()) V = It->second; 665 PN->addIncoming(V, NewExit); 666 } 667 } 668 669 // Rewrite the code to refer to itself. 670 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 671 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 672 E = NewBlocks[i]->end(); I != E; ++I) 673 RemapInstruction(I, ValueMap); 674 675 // Rewrite the original preheader to select between versions of the loop. 676 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 677 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 678 "Preheader splitting did not work correctly!"); 679 680 // Emit the new branch that selects between the two versions of this loop. 681 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 682 OldBR->eraseFromParent(); 683 684 LoopProcessWorklist.push_back(NewLoop); 685 LPM->redoLoop(L); 686 687 // Now we rewrite the original code to know that the condition is true and the 688 // new code to know that the condition is false. 689 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 690 691 // It's possible that simplifying one loop could cause the other to be 692 // deleted. If so, don't simplify it. 693 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 694 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 695} 696 697/// RemoveFromWorklist - Remove all instances of I from the worklist vector 698/// specified. 699static void RemoveFromWorklist(Instruction *I, 700 std::vector<Instruction*> &Worklist) { 701 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 702 Worklist.end(), I); 703 while (WI != Worklist.end()) { 704 unsigned Offset = WI-Worklist.begin(); 705 Worklist.erase(WI); 706 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 707 } 708} 709 710/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 711/// program, replacing all uses with V and update the worklist. 712static void ReplaceUsesOfWith(Instruction *I, Value *V, 713 std::vector<Instruction*> &Worklist) { 714 DOUT << "Replace with '" << *V << "': " << *I; 715 716 // Add uses to the worklist, which may be dead now. 717 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 718 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 719 Worklist.push_back(Use); 720 721 // Add users to the worklist which may be simplified now. 722 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 723 UI != E; ++UI) 724 Worklist.push_back(cast<Instruction>(*UI)); 725 I->replaceAllUsesWith(V); 726 I->eraseFromParent(); 727 RemoveFromWorklist(I, Worklist); 728 ++NumSimplify; 729} 730 731/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 732/// information, and remove any dead successors it has. 733/// 734void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 735 std::vector<Instruction*> &Worklist) { 736 if (pred_begin(BB) != pred_end(BB)) { 737 // This block isn't dead, since an edge to BB was just removed, see if there 738 // are any easy simplifications we can do now. 739 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 740 // If it has one pred, fold phi nodes in BB. 741 while (isa<PHINode>(BB->begin())) 742 ReplaceUsesOfWith(BB->begin(), 743 cast<PHINode>(BB->begin())->getIncomingValue(0), 744 Worklist); 745 746 // If this is the header of a loop and the only pred is the latch, we now 747 // have an unreachable loop. 748 if (Loop *L = LI->getLoopFor(BB)) 749 if (L->getHeader() == BB && L->contains(Pred)) { 750 // Remove the branch from the latch to the header block, this makes 751 // the header dead, which will make the latch dead (because the header 752 // dominates the latch). 753 Pred->getTerminator()->eraseFromParent(); 754 new UnreachableInst(Pred); 755 756 // The loop is now broken, remove it from LI. 757 RemoveLoopFromHierarchy(L); 758 759 // Reprocess the header, which now IS dead. 760 RemoveBlockIfDead(BB, Worklist); 761 return; 762 } 763 764 // If pred ends in a uncond branch, add uncond branch to worklist so that 765 // the two blocks will get merged. 766 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 767 if (BI->isUnconditional()) 768 Worklist.push_back(BI); 769 } 770 return; 771 } 772 773 DOUT << "Nuking dead block: " << *BB; 774 775 // Remove the instructions in the basic block from the worklist. 776 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 777 RemoveFromWorklist(I, Worklist); 778 779 // Anything that uses the instructions in this basic block should have their 780 // uses replaced with undefs. 781 if (!I->use_empty()) 782 I->replaceAllUsesWith(UndefValue::get(I->getType())); 783 } 784 785 // If this is the edge to the header block for a loop, remove the loop and 786 // promote all subloops. 787 if (Loop *BBLoop = LI->getLoopFor(BB)) { 788 if (BBLoop->getLoopLatch() == BB) 789 RemoveLoopFromHierarchy(BBLoop); 790 } 791 792 // Remove the block from the loop info, which removes it from any loops it 793 // was in. 794 LI->removeBlock(BB); 795 796 797 // Remove phi node entries in successors for this block. 798 TerminatorInst *TI = BB->getTerminator(); 799 std::vector<BasicBlock*> Succs; 800 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 801 Succs.push_back(TI->getSuccessor(i)); 802 TI->getSuccessor(i)->removePredecessor(BB); 803 } 804 805 // Unique the successors, remove anything with multiple uses. 806 std::sort(Succs.begin(), Succs.end()); 807 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 808 809 // Remove the basic block, including all of the instructions contained in it. 810 BB->eraseFromParent(); 811 812 // Remove successor blocks here that are not dead, so that we know we only 813 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 814 // then getting removed before we revisit them, which is badness. 815 // 816 for (unsigned i = 0; i != Succs.size(); ++i) 817 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 818 // One exception is loop headers. If this block was the preheader for a 819 // loop, then we DO want to visit the loop so the loop gets deleted. 820 // We know that if the successor is a loop header, that this loop had to 821 // be the preheader: the case where this was the latch block was handled 822 // above and headers can only have two predecessors. 823 if (!LI->isLoopHeader(Succs[i])) { 824 Succs.erase(Succs.begin()+i); 825 --i; 826 } 827 } 828 829 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 830 RemoveBlockIfDead(Succs[i], Worklist); 831} 832 833/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 834/// become unwrapped, either because the backedge was deleted, or because the 835/// edge into the header was removed. If the edge into the header from the 836/// latch block was removed, the loop is unwrapped but subloops are still alive, 837/// so they just reparent loops. If the loops are actually dead, they will be 838/// removed later. 839void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 840 LPM->deleteLoopFromQueue(L); 841 RemoveLoopFromWorklist(L); 842} 843 844 845 846// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 847// the value specified by Val in the specified loop, or we know it does NOT have 848// that value. Rewrite any uses of LIC or of properties correlated to it. 849void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 850 Constant *Val, 851 bool IsEqual) { 852 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 853 854 // FIXME: Support correlated properties, like: 855 // for (...) 856 // if (li1 < li2) 857 // ... 858 // if (li1 > li2) 859 // ... 860 861 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 862 // selects, switches. 863 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 864 std::vector<Instruction*> Worklist; 865 866 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 867 // in the loop with the appropriate one directly. 868 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) { 869 Value *Replacement; 870 if (IsEqual) 871 Replacement = Val; 872 else 873 Replacement = ConstantInt::get(Type::Int1Ty, 874 !cast<ConstantInt>(Val)->getZExtValue()); 875 876 for (unsigned i = 0, e = Users.size(); i != e; ++i) 877 if (Instruction *U = cast<Instruction>(Users[i])) { 878 if (!L->contains(U->getParent())) 879 continue; 880 U->replaceUsesOfWith(LIC, Replacement); 881 Worklist.push_back(U); 882 } 883 } else { 884 // Otherwise, we don't know the precise value of LIC, but we do know that it 885 // is certainly NOT "Val". As such, simplify any uses in the loop that we 886 // can. This case occurs when we unswitch switch statements. 887 for (unsigned i = 0, e = Users.size(); i != e; ++i) 888 if (Instruction *U = cast<Instruction>(Users[i])) { 889 if (!L->contains(U->getParent())) 890 continue; 891 892 Worklist.push_back(U); 893 894 // If we know that LIC is not Val, use this info to simplify code. 895 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 896 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 897 if (SI->getCaseValue(i) == Val) { 898 // Found a dead case value. Don't remove PHI nodes in the 899 // successor if they become single-entry, those PHI nodes may 900 // be in the Users list. 901 902 // FIXME: This is a hack. We need to keep the successor around 903 // and hooked up so as to preserve the loop structure, because 904 // trying to update it is complicated. So instead we preserve the 905 // loop structure and put the block on an dead code path. 906 907 BasicBlock* Old = SI->getParent(); 908 BasicBlock* Split = SplitBlock(Old, SI); 909 910 Instruction* OldTerm = Old->getTerminator(); 911 new BranchInst(Split, SI->getSuccessor(i), 912 ConstantInt::getTrue(), OldTerm); 913 914 Old->getTerminator()->eraseFromParent(); 915 916 917 PHINode *PN; 918 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); 919 (PN = dyn_cast<PHINode>(II)); ++II) { 920 Value *InVal = PN->removeIncomingValue(Split, false); 921 PN->addIncoming(InVal, Old); 922 } 923 924 SI->removeCase(i); 925 break; 926 } 927 } 928 } 929 930 // TODO: We could do other simplifications, for example, turning 931 // LIC == Val -> false. 932 } 933 } 934 935 SimplifyCode(Worklist); 936} 937 938/// SimplifyCode - Okay, now that we have simplified some instructions in the 939/// loop, walk over it and constant prop, dce, and fold control flow where 940/// possible. Note that this is effectively a very simple loop-structure-aware 941/// optimizer. During processing of this loop, L could very well be deleted, so 942/// it must not be used. 943/// 944/// FIXME: When the loop optimizer is more mature, separate this out to a new 945/// pass. 946/// 947void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) { 948 while (!Worklist.empty()) { 949 Instruction *I = Worklist.back(); 950 Worklist.pop_back(); 951 952 // Simple constant folding. 953 if (Constant *C = ConstantFoldInstruction(I)) { 954 ReplaceUsesOfWith(I, C, Worklist); 955 continue; 956 } 957 958 // Simple DCE. 959 if (isInstructionTriviallyDead(I)) { 960 DOUT << "Remove dead instruction '" << *I; 961 962 // Add uses to the worklist, which may be dead now. 963 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 964 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 965 Worklist.push_back(Use); 966 I->eraseFromParent(); 967 RemoveFromWorklist(I, Worklist); 968 ++NumSimplify; 969 continue; 970 } 971 972 // Special case hacks that appear commonly in unswitched code. 973 switch (I->getOpcode()) { 974 case Instruction::Select: 975 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 976 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist); 977 continue; 978 } 979 break; 980 case Instruction::And: 981 if (isa<ConstantInt>(I->getOperand(0)) && 982 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 983 cast<BinaryOperator>(I)->swapOperands(); 984 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 985 if (CB->getType() == Type::Int1Ty) { 986 if (CB->isOne()) // X & 1 -> X 987 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 988 else // X & 0 -> 0 989 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 990 continue; 991 } 992 break; 993 case Instruction::Or: 994 if (isa<ConstantInt>(I->getOperand(0)) && 995 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 996 cast<BinaryOperator>(I)->swapOperands(); 997 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 998 if (CB->getType() == Type::Int1Ty) { 999 if (CB->isOne()) // X | 1 -> 1 1000 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1001 else // X | 0 -> X 1002 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1003 continue; 1004 } 1005 break; 1006 case Instruction::Br: { 1007 BranchInst *BI = cast<BranchInst>(I); 1008 if (BI->isUnconditional()) { 1009 // If BI's parent is the only pred of the successor, fold the two blocks 1010 // together. 1011 BasicBlock *Pred = BI->getParent(); 1012 BasicBlock *Succ = BI->getSuccessor(0); 1013 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1014 if (!SinglePred) continue; // Nothing to do. 1015 assert(SinglePred == Pred && "CFG broken"); 1016 1017 DOUT << "Merging blocks: " << Pred->getName() << " <- " 1018 << Succ->getName() << "\n"; 1019 1020 // Resolve any single entry PHI nodes in Succ. 1021 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1022 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); 1023 1024 // Move all of the successor contents from Succ to Pred. 1025 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1026 Succ->end()); 1027 BI->eraseFromParent(); 1028 RemoveFromWorklist(BI, Worklist); 1029 1030 // If Succ has any successors with PHI nodes, update them to have 1031 // entries coming from Pred instead of Succ. 1032 Succ->replaceAllUsesWith(Pred); 1033 1034 // Remove Succ from the loop tree. 1035 LI->removeBlock(Succ); 1036 Succ->eraseFromParent(); 1037 ++NumSimplify; 1038 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1039 // Conditional branch. Turn it into an unconditional branch, then 1040 // remove dead blocks. 1041 break; // FIXME: Enable. 1042 1043 DOUT << "Folded branch: " << *BI; 1044 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1045 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1046 DeadSucc->removePredecessor(BI->getParent(), true); 1047 Worklist.push_back(new BranchInst(LiveSucc, BI)); 1048 BI->eraseFromParent(); 1049 RemoveFromWorklist(BI, Worklist); 1050 ++NumSimplify; 1051 1052 RemoveBlockIfDead(DeadSucc, Worklist); 1053 } 1054 break; 1055 } 1056 } 1057 } 1058} 1059