LoopUnswitch.cpp revision fd98dc9a76b0b00b3f076fb9c09f698b1b592ab1
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 Loop *OrigDestBBL = LI->getLoopFor(BB->getTerminator()->getSuccessor(SuccNum)); 417 if (SplitCriticalEdge(BB->getTerminator(), SuccNum)) { 418 BasicBlock *NewBB = LatchTerm->getSuccessor(SuccNum); 419 420 Loop *BBL = LI->getLoopFor(BB); 421 if (!BBL || !OrigDestBBL) 422 return NewBB; 423 424 // If edge is inside a loop then NewBB is part of same loop. 425 if (BBL == OrigDestBBL) 426 BBL->addBasicBlockToLoop(NewBB, *LI); 427 // If edge is entering loop then NewBB is part of outer loop. 428 else if (BBL->contains(OrigDestBBL->getHeader())) 429 BBL->addBasicBlockToLoop(NewBB, *LI); 430 // If edge is from an inner loop to outer loop then NewBB is part 431 // of outer loop. 432 else if (OrigDestBBL->contains(BBL->getHeader())) 433 OrigDestBBL->addBasicBlockToLoop(NewBB, *LI); 434 // Else edge is connecting two loops and NewBB is part of their parent loop 435 else if (Loop *PL = OrigDestBBL->getParentLoop()) 436 PL->addBasicBlockToLoop(NewBB, *LI); 437 438 return NewBB; 439 } 440 441 // If the edge isn't critical, then BB has a single successor or Succ has a 442 // single pred. Split the block. 443 BasicBlock::iterator SplitPoint; 444 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 445 // If the successor only has a single pred, split the top of the successor 446 // block. 447 assert(SP == BB && "CFG broken"); 448 return SplitBlock(Succ, Succ->begin()); 449 } else { 450 // Otherwise, if BB has a single successor, split it at the bottom of the 451 // block. 452 assert(BB->getTerminator()->getNumSuccessors() == 1 && 453 "Should have a single succ!"); 454 return SplitBlock(BB, BB->getTerminator()); 455 } 456} 457 458 459 460// RemapInstruction - Convert the instruction operands from referencing the 461// current values into those specified by ValueMap. 462// 463static inline void RemapInstruction(Instruction *I, 464 DenseMap<const Value *, Value*> &ValueMap) { 465 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 466 Value *Op = I->getOperand(op); 467 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op); 468 if (It != ValueMap.end()) Op = It->second; 469 I->setOperand(op, Op); 470 } 471} 472 473/// CloneLoop - Recursively clone the specified loop and all of its children, 474/// mapping the blocks with the specified map. 475static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM, 476 LoopInfo *LI, LPPassManager *LPM) { 477 Loop *New = new Loop(); 478 479 LPM->insertLoop(New, PL); 480 481 // Add all of the blocks in L to the new loop. 482 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 483 I != E; ++I) 484 if (LI->getLoopFor(*I) == L) 485 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 486 487 // Add all of the subloops to the new loop. 488 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 489 CloneLoop(*I, New, VM, LI, LPM); 490 491 return New; 492} 493 494/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 495/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 496/// code immediately before InsertPt. 497static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 498 BasicBlock *TrueDest, 499 BasicBlock *FalseDest, 500 Instruction *InsertPt) { 501 // Insert a conditional branch on LIC to the two preheaders. The original 502 // code is the true version and the new code is the false version. 503 Value *BranchVal = LIC; 504 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty) 505 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); 506 else if (Val != ConstantInt::getTrue()) 507 // We want to enter the new loop when the condition is true. 508 std::swap(TrueDest, FalseDest); 509 510 // Insert the new branch. 511 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); 512} 513 514 515/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 516/// condition in it (a cond branch from its header block to its latch block, 517/// where the path through the loop that doesn't execute its body has no 518/// side-effects), unswitch it. This doesn't involve any code duplication, just 519/// moving the conditional branch outside of the loop and updating loop info. 520void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 521 Constant *Val, 522 BasicBlock *ExitBlock) { 523 DOUT << "loop-unswitch: Trivial-Unswitch loop %" 524 << L->getHeader()->getName() << " [" << L->getBlocks().size() 525 << " blocks] in Function " << L->getHeader()->getParent()->getName() 526 << " on cond: " << *Val << " == " << *Cond << "\n"; 527 528 // First step, split the preheader, so that we know that there is a safe place 529 // to insert the conditional branch. We will change 'OrigPH' to have a 530 // conditional branch on Cond. 531 BasicBlock *OrigPH = L->getLoopPreheader(); 532 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); 533 534 // Now that we have a place to insert the conditional branch, create a place 535 // to branch to: this is the exit block out of the loop that we should 536 // short-circuit to. 537 538 // Split this block now, so that the loop maintains its exit block, and so 539 // that the jump from the preheader can execute the contents of the exit block 540 // without actually branching to it (the exit block should be dominated by the 541 // loop header, not the preheader). 542 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 543 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); 544 545 // Okay, now we have a position to branch from and a position to branch to, 546 // insert the new conditional branch. 547 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 548 OrigPH->getTerminator()); 549 OrigPH->getTerminator()->eraseFromParent(); 550 551 // We need to reprocess this loop, it could be unswitched again. 552 LPM->redoLoop(L); 553 554 // Now that we know that the loop is never entered when this condition is a 555 // particular value, rewrite the loop with this info. We know that this will 556 // at least eliminate the old branch. 557 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 558 ++NumTrivial; 559} 560 561 562/// VersionLoop - We determined that the loop is profitable to unswitch when LIC 563/// equal Val. Split it into loop versions and test the condition outside of 564/// either loop. Return the loops created as Out1/Out2. 565void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 566 Loop *L) { 567 Function *F = L->getHeader()->getParent(); 568 DOUT << "loop-unswitch: Unswitching loop %" 569 << L->getHeader()->getName() << " [" << L->getBlocks().size() 570 << " blocks] in Function " << F->getName() 571 << " when '" << *Val << "' == " << *LIC << "\n"; 572 573 // LoopBlocks contains all of the basic blocks of the loop, including the 574 // preheader of the loop, the body of the loop, and the exit blocks of the 575 // loop, in that order. 576 std::vector<BasicBlock*> LoopBlocks; 577 578 // First step, split the preheader and exit blocks, and add these blocks to 579 // the LoopBlocks list. 580 BasicBlock *OrigPreheader = L->getLoopPreheader(); 581 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); 582 583 // We want the loop to come after the preheader, but before the exit blocks. 584 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 585 586 std::vector<BasicBlock*> ExitBlocks; 587 L->getUniqueExitBlocks(ExitBlocks); 588 589 // Split all of the edges from inside the loop to their exit blocks. Update 590 // the appropriate Phi nodes as we do so. 591 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 592 BasicBlock *ExitBlock = ExitBlocks[i]; 593 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 594 595 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 596 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock); 597 BasicBlock* StartBlock = Preds[j]; 598 BasicBlock* EndBlock; 599 if (MiddleBlock->getSinglePredecessor() == ExitBlock) { 600 EndBlock = MiddleBlock; 601 MiddleBlock = EndBlock->getSinglePredecessor();; 602 } else { 603 EndBlock = ExitBlock; 604 } 605 606 std::set<PHINode*> InsertedPHIs; 607 PHINode* OldLCSSA = 0; 608 for (BasicBlock::iterator I = EndBlock->begin(); 609 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) { 610 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock); 611 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(), 612 OldLCSSA->getName() + ".us-lcssa", 613 MiddleBlock->getTerminator()); 614 NewLCSSA->addIncoming(OldValue, StartBlock); 615 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock), 616 NewLCSSA); 617 InsertedPHIs.insert(NewLCSSA); 618 } 619 620 BasicBlock::iterator InsertPt = EndBlock->begin(); 621 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt; 622 for (BasicBlock::iterator I = MiddleBlock->begin(); 623 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0; 624 ++I) { 625 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(), 626 OldLCSSA->getName() + ".us-lcssa", 627 InsertPt); 628 OldLCSSA->replaceAllUsesWith(NewLCSSA); 629 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); 630 } 631 } 632 } 633 634 // The exit blocks may have been changed due to edge splitting, recompute. 635 ExitBlocks.clear(); 636 L->getUniqueExitBlocks(ExitBlocks); 637 638 // Add exit blocks to the loop blocks. 639 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 640 641 // Next step, clone all of the basic blocks that make up the loop (including 642 // the loop preheader and exit blocks), keeping track of the mapping between 643 // the instructions and blocks. 644 std::vector<BasicBlock*> NewBlocks; 645 NewBlocks.reserve(LoopBlocks.size()); 646 DenseMap<const Value*, Value*> ValueMap; 647 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 648 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 649 NewBlocks.push_back(New); 650 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 651 } 652 653 // Splice the newly inserted blocks into the function right before the 654 // original preheader. 655 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 656 NewBlocks[0], F->end()); 657 658 // Now we create the new Loop object for the versioned loop. 659 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); 660 Loop *ParentLoop = L->getParentLoop(); 661 if (ParentLoop) { 662 // Make sure to add the cloned preheader and exit blocks to the parent loop 663 // as well. 664 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 665 } 666 667 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 668 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 669 // The new exit block should be in the same loop as the old one. 670 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 671 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 672 673 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 674 "Exit block should have been split to have one successor!"); 675 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 676 677 // If the successor of the exit block had PHI nodes, add an entry for 678 // NewExit. 679 PHINode *PN; 680 for (BasicBlock::iterator I = ExitSucc->begin(); 681 (PN = dyn_cast<PHINode>(I)); ++I) { 682 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 683 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V); 684 if (It != ValueMap.end()) V = It->second; 685 PN->addIncoming(V, NewExit); 686 } 687 } 688 689 // Rewrite the code to refer to itself. 690 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 691 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 692 E = NewBlocks[i]->end(); I != E; ++I) 693 RemapInstruction(I, ValueMap); 694 695 // Rewrite the original preheader to select between versions of the loop. 696 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 697 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 698 "Preheader splitting did not work correctly!"); 699 700 // Emit the new branch that selects between the two versions of this loop. 701 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 702 OldBR->eraseFromParent(); 703 704 LoopProcessWorklist.push_back(NewLoop); 705 LPM->redoLoop(L); 706 707 // Now we rewrite the original code to know that the condition is true and the 708 // new code to know that the condition is false. 709 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 710 711 // It's possible that simplifying one loop could cause the other to be 712 // deleted. If so, don't simplify it. 713 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 714 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 715} 716 717/// RemoveFromWorklist - Remove all instances of I from the worklist vector 718/// specified. 719static void RemoveFromWorklist(Instruction *I, 720 std::vector<Instruction*> &Worklist) { 721 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 722 Worklist.end(), I); 723 while (WI != Worklist.end()) { 724 unsigned Offset = WI-Worklist.begin(); 725 Worklist.erase(WI); 726 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 727 } 728} 729 730/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 731/// program, replacing all uses with V and update the worklist. 732static void ReplaceUsesOfWith(Instruction *I, Value *V, 733 std::vector<Instruction*> &Worklist) { 734 DOUT << "Replace with '" << *V << "': " << *I; 735 736 // Add uses to the worklist, which may be dead now. 737 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 738 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 739 Worklist.push_back(Use); 740 741 // Add users to the worklist which may be simplified now. 742 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 743 UI != E; ++UI) 744 Worklist.push_back(cast<Instruction>(*UI)); 745 I->replaceAllUsesWith(V); 746 I->eraseFromParent(); 747 RemoveFromWorklist(I, Worklist); 748 ++NumSimplify; 749} 750 751/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 752/// information, and remove any dead successors it has. 753/// 754void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 755 std::vector<Instruction*> &Worklist) { 756 if (pred_begin(BB) != pred_end(BB)) { 757 // This block isn't dead, since an edge to BB was just removed, see if there 758 // are any easy simplifications we can do now. 759 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 760 // If it has one pred, fold phi nodes in BB. 761 while (isa<PHINode>(BB->begin())) 762 ReplaceUsesOfWith(BB->begin(), 763 cast<PHINode>(BB->begin())->getIncomingValue(0), 764 Worklist); 765 766 // If this is the header of a loop and the only pred is the latch, we now 767 // have an unreachable loop. 768 if (Loop *L = LI->getLoopFor(BB)) 769 if (L->getHeader() == BB && L->contains(Pred)) { 770 // Remove the branch from the latch to the header block, this makes 771 // the header dead, which will make the latch dead (because the header 772 // dominates the latch). 773 Pred->getTerminator()->eraseFromParent(); 774 new UnreachableInst(Pred); 775 776 // The loop is now broken, remove it from LI. 777 RemoveLoopFromHierarchy(L); 778 779 // Reprocess the header, which now IS dead. 780 RemoveBlockIfDead(BB, Worklist); 781 return; 782 } 783 784 // If pred ends in a uncond branch, add uncond branch to worklist so that 785 // the two blocks will get merged. 786 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 787 if (BI->isUnconditional()) 788 Worklist.push_back(BI); 789 } 790 return; 791 } 792 793 DOUT << "Nuking dead block: " << *BB; 794 795 // Remove the instructions in the basic block from the worklist. 796 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 797 RemoveFromWorklist(I, Worklist); 798 799 // Anything that uses the instructions in this basic block should have their 800 // uses replaced with undefs. 801 if (!I->use_empty()) 802 I->replaceAllUsesWith(UndefValue::get(I->getType())); 803 } 804 805 // If this is the edge to the header block for a loop, remove the loop and 806 // promote all subloops. 807 if (Loop *BBLoop = LI->getLoopFor(BB)) { 808 if (BBLoop->getLoopLatch() == BB) 809 RemoveLoopFromHierarchy(BBLoop); 810 } 811 812 // Remove the block from the loop info, which removes it from any loops it 813 // was in. 814 LI->removeBlock(BB); 815 816 817 // Remove phi node entries in successors for this block. 818 TerminatorInst *TI = BB->getTerminator(); 819 std::vector<BasicBlock*> Succs; 820 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 821 Succs.push_back(TI->getSuccessor(i)); 822 TI->getSuccessor(i)->removePredecessor(BB); 823 } 824 825 // Unique the successors, remove anything with multiple uses. 826 std::sort(Succs.begin(), Succs.end()); 827 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 828 829 // Remove the basic block, including all of the instructions contained in it. 830 BB->eraseFromParent(); 831 832 // Remove successor blocks here that are not dead, so that we know we only 833 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 834 // then getting removed before we revisit them, which is badness. 835 // 836 for (unsigned i = 0; i != Succs.size(); ++i) 837 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 838 // One exception is loop headers. If this block was the preheader for a 839 // loop, then we DO want to visit the loop so the loop gets deleted. 840 // We know that if the successor is a loop header, that this loop had to 841 // be the preheader: the case where this was the latch block was handled 842 // above and headers can only have two predecessors. 843 if (!LI->isLoopHeader(Succs[i])) { 844 Succs.erase(Succs.begin()+i); 845 --i; 846 } 847 } 848 849 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 850 RemoveBlockIfDead(Succs[i], Worklist); 851} 852 853/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 854/// become unwrapped, either because the backedge was deleted, or because the 855/// edge into the header was removed. If the edge into the header from the 856/// latch block was removed, the loop is unwrapped but subloops are still alive, 857/// so they just reparent loops. If the loops are actually dead, they will be 858/// removed later. 859void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 860 LPM->deleteLoopFromQueue(L); 861 RemoveLoopFromWorklist(L); 862} 863 864 865 866// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 867// the value specified by Val in the specified loop, or we know it does NOT have 868// that value. Rewrite any uses of LIC or of properties correlated to it. 869void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 870 Constant *Val, 871 bool IsEqual) { 872 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 873 874 // FIXME: Support correlated properties, like: 875 // for (...) 876 // if (li1 < li2) 877 // ... 878 // if (li1 > li2) 879 // ... 880 881 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 882 // selects, switches. 883 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 884 std::vector<Instruction*> Worklist; 885 886 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 887 // in the loop with the appropriate one directly. 888 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) { 889 Value *Replacement; 890 if (IsEqual) 891 Replacement = Val; 892 else 893 Replacement = ConstantInt::get(Type::Int1Ty, 894 !cast<ConstantInt>(Val)->getZExtValue()); 895 896 for (unsigned i = 0, e = Users.size(); i != e; ++i) 897 if (Instruction *U = cast<Instruction>(Users[i])) { 898 if (!L->contains(U->getParent())) 899 continue; 900 U->replaceUsesOfWith(LIC, Replacement); 901 Worklist.push_back(U); 902 } 903 } else { 904 // Otherwise, we don't know the precise value of LIC, but we do know that it 905 // is certainly NOT "Val". As such, simplify any uses in the loop that we 906 // can. This case occurs when we unswitch switch statements. 907 for (unsigned i = 0, e = Users.size(); i != e; ++i) 908 if (Instruction *U = cast<Instruction>(Users[i])) { 909 if (!L->contains(U->getParent())) 910 continue; 911 912 Worklist.push_back(U); 913 914 // If we know that LIC is not Val, use this info to simplify code. 915 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 916 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 917 if (SI->getCaseValue(i) == Val) { 918 // Found a dead case value. Don't remove PHI nodes in the 919 // successor if they become single-entry, those PHI nodes may 920 // be in the Users list. 921 922 // FIXME: This is a hack. We need to keep the successor around 923 // and hooked up so as to preserve the loop structure, because 924 // trying to update it is complicated. So instead we preserve the 925 // loop structure and put the block on an dead code path. 926 927 BasicBlock* Old = SI->getParent(); 928 BasicBlock* Split = SplitBlock(Old, SI); 929 930 Instruction* OldTerm = Old->getTerminator(); 931 new BranchInst(Split, SI->getSuccessor(i), 932 ConstantInt::getTrue(), OldTerm); 933 934 Old->getTerminator()->eraseFromParent(); 935 936 937 PHINode *PN; 938 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); 939 (PN = dyn_cast<PHINode>(II)); ++II) { 940 Value *InVal = PN->removeIncomingValue(Split, false); 941 PN->addIncoming(InVal, Old); 942 } 943 944 SI->removeCase(i); 945 break; 946 } 947 } 948 } 949 950 // TODO: We could do other simplifications, for example, turning 951 // LIC == Val -> false. 952 } 953 } 954 955 SimplifyCode(Worklist); 956} 957 958/// SimplifyCode - Okay, now that we have simplified some instructions in the 959/// loop, walk over it and constant prop, dce, and fold control flow where 960/// possible. Note that this is effectively a very simple loop-structure-aware 961/// optimizer. During processing of this loop, L could very well be deleted, so 962/// it must not be used. 963/// 964/// FIXME: When the loop optimizer is more mature, separate this out to a new 965/// pass. 966/// 967void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) { 968 while (!Worklist.empty()) { 969 Instruction *I = Worklist.back(); 970 Worklist.pop_back(); 971 972 // Simple constant folding. 973 if (Constant *C = ConstantFoldInstruction(I)) { 974 ReplaceUsesOfWith(I, C, Worklist); 975 continue; 976 } 977 978 // Simple DCE. 979 if (isInstructionTriviallyDead(I)) { 980 DOUT << "Remove dead instruction '" << *I; 981 982 // Add uses to the worklist, which may be dead now. 983 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 984 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 985 Worklist.push_back(Use); 986 I->eraseFromParent(); 987 RemoveFromWorklist(I, Worklist); 988 ++NumSimplify; 989 continue; 990 } 991 992 // Special case hacks that appear commonly in unswitched code. 993 switch (I->getOpcode()) { 994 case Instruction::Select: 995 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 996 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist); 997 continue; 998 } 999 break; 1000 case Instruction::And: 1001 if (isa<ConstantInt>(I->getOperand(0)) && 1002 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1003 cast<BinaryOperator>(I)->swapOperands(); 1004 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1005 if (CB->getType() == Type::Int1Ty) { 1006 if (CB->isOne()) // X & 1 -> X 1007 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1008 else // X & 0 -> 0 1009 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1010 continue; 1011 } 1012 break; 1013 case Instruction::Or: 1014 if (isa<ConstantInt>(I->getOperand(0)) && 1015 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1016 cast<BinaryOperator>(I)->swapOperands(); 1017 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1018 if (CB->getType() == Type::Int1Ty) { 1019 if (CB->isOne()) // X | 1 -> 1 1020 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1021 else // X | 0 -> X 1022 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1023 continue; 1024 } 1025 break; 1026 case Instruction::Br: { 1027 BranchInst *BI = cast<BranchInst>(I); 1028 if (BI->isUnconditional()) { 1029 // If BI's parent is the only pred of the successor, fold the two blocks 1030 // together. 1031 BasicBlock *Pred = BI->getParent(); 1032 BasicBlock *Succ = BI->getSuccessor(0); 1033 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1034 if (!SinglePred) continue; // Nothing to do. 1035 assert(SinglePred == Pred && "CFG broken"); 1036 1037 DOUT << "Merging blocks: " << Pred->getName() << " <- " 1038 << Succ->getName() << "\n"; 1039 1040 // Resolve any single entry PHI nodes in Succ. 1041 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1042 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); 1043 1044 // Move all of the successor contents from Succ to Pred. 1045 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1046 Succ->end()); 1047 BI->eraseFromParent(); 1048 RemoveFromWorklist(BI, Worklist); 1049 1050 // If Succ has any successors with PHI nodes, update them to have 1051 // entries coming from Pred instead of Succ. 1052 Succ->replaceAllUsesWith(Pred); 1053 1054 // Remove Succ from the loop tree. 1055 LI->removeBlock(Succ); 1056 Succ->eraseFromParent(); 1057 ++NumSimplify; 1058 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1059 // Conditional branch. Turn it into an unconditional branch, then 1060 // remove dead blocks. 1061 break; // FIXME: Enable. 1062 1063 DOUT << "Folded branch: " << *BI; 1064 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1065 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1066 DeadSucc->removePredecessor(BI->getParent(), true); 1067 Worklist.push_back(new BranchInst(LiveSucc, BI)); 1068 BI->eraseFromParent(); 1069 RemoveFromWorklist(BI, Worklist); 1070 ++NumSimplify; 1071 1072 RemoveBlockIfDead(DeadSucc, Worklist); 1073 } 1074 break; 1075 } 1076 } 1077 } 1078} 1079