LoopUnswitch.cpp revision f8bf116f808ecc6a9ca09cb761fe2a169e006b57
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/Function.h" 33#include "llvm/Instructions.h" 34#include "llvm/Analysis/LoopInfo.h" 35#include "llvm/Transforms/Utils/Cloning.h" 36#include "llvm/Transforms/Utils/Local.h" 37#include "llvm/Transforms/Utils/BasicBlockUtils.h" 38#include "llvm/ADT/Statistic.h" 39#include "llvm/ADT/PostOrderIterator.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/CommandLine.h" 42#include <algorithm> 43#include <iostream> 44#include <set> 45using namespace llvm; 46 47namespace { 48 Statistic<> NumBranches("loop-unswitch", "Number of branches unswitched"); 49 Statistic<> NumSwitches("loop-unswitch", "Number of switches unswitched"); 50 Statistic<> NumSelects ("loop-unswitch", "Number of selects unswitched"); 51 Statistic<> NumTrivial ("loop-unswitch", 52 "Number of unswitches that are trivial"); 53 Statistic<> NumSimplify("loop-unswitch", 54 "Number of simplifications of unswitched code"); 55 cl::opt<unsigned> 56 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 57 cl::init(10), cl::Hidden); 58 59 class LoopUnswitch : public FunctionPass { 60 LoopInfo *LI; // Loop information 61 62 // LoopProcessWorklist - List of loops we need to process. 63 std::vector<Loop*> LoopProcessWorklist; 64 public: 65 virtual bool runOnFunction(Function &F); 66 bool visitLoop(Loop *L); 67 68 /// This transformation requires natural loop information & requires that 69 /// loop preheaders be inserted into the CFG... 70 /// 71 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 72 AU.addRequiredID(LoopSimplifyID); 73 AU.addPreservedID(LoopSimplifyID); 74 AU.addRequired<LoopInfo>(); 75 AU.addPreserved<LoopInfo>(); 76 } 77 78 private: 79 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, 80 /// remove it. 81 void RemoveLoopFromWorklist(Loop *L) { 82 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(), 83 LoopProcessWorklist.end(), L); 84 if (I != LoopProcessWorklist.end()) 85 LoopProcessWorklist.erase(I); 86 } 87 88 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L); 89 unsigned getLoopUnswitchCost(Loop *L, Value *LIC); 90 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, 91 BasicBlock *ExitBlock); 92 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); 93 BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To); 94 BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt); 95 96 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 97 Constant *Val, bool isEqual); 98 99 void SimplifyCode(std::vector<Instruction*> &Worklist); 100 void RemoveBlockIfDead(BasicBlock *BB, 101 std::vector<Instruction*> &Worklist); 102 void RemoveLoopFromHierarchy(Loop *L); 103 }; 104 RegisterOpt<LoopUnswitch> X("loop-unswitch", "Unswitch loops"); 105} 106 107FunctionPass *llvm::createLoopUnswitchPass() { return new LoopUnswitch(); } 108 109bool LoopUnswitch::runOnFunction(Function &F) { 110 bool Changed = false; 111 LI = &getAnalysis<LoopInfo>(); 112 113 // Populate the worklist of loops to process in post-order. 114 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 115 for (po_iterator<Loop*> LI = po_begin(*I), E = po_end(*I); LI != E; ++LI) 116 LoopProcessWorklist.push_back(*LI); 117 118 // Process the loops in worklist order, this is a post-order visitation of 119 // the loops. We use a worklist of loops so that loops can be removed at any 120 // time if they are deleted (e.g. the backedge of a loop is removed). 121 while (!LoopProcessWorklist.empty()) { 122 Loop *L = LoopProcessWorklist.back(); 123 LoopProcessWorklist.pop_back(); 124 Changed |= visitLoop(L); 125 } 126 127 return Changed; 128} 129 130/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is 131/// invariant in the loop, or has an invariant piece, return the invariant. 132/// Otherwise, return null. 133static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { 134 // Constants should be folded, not unswitched on! 135 if (isa<Constant>(Cond)) return false; 136 137 // TODO: Handle: br (VARIANT|INVARIANT). 138 // TODO: Hoist simple expressions out of loops. 139 if (L->isLoopInvariant(Cond)) return Cond; 140 141 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond)) 142 if (BO->getOpcode() == Instruction::And || 143 BO->getOpcode() == Instruction::Or) { 144 // If either the left or right side is invariant, we can unswitch on this, 145 // which will cause the branch to go away in one loop and the condition to 146 // simplify in the other one. 147 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed)) 148 return LHS; 149 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) 150 return RHS; 151 } 152 153 return 0; 154} 155 156bool LoopUnswitch::visitLoop(Loop *L) { 157 bool Changed = false; 158 159 // Loop over all of the basic blocks in the loop. If we find an interior 160 // block that is branching on a loop-invariant condition, we can unswitch this 161 // loop. 162 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 163 I != E; ++I) { 164 TerminatorInst *TI = (*I)->getTerminator(); 165 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 166 // If this isn't branching on an invariant condition, we can't unswitch 167 // it. 168 if (BI->isConditional()) { 169 // See if this, or some part of it, is loop invariant. If so, we can 170 // unswitch on it if we desire. 171 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); 172 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::True, L)) { 173 ++NumBranches; 174 return true; 175 } 176 } 177 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 178 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 179 if (LoopCond && SI->getNumCases() > 1) { 180 // Find a value to unswitch on: 181 // FIXME: this should chose the most expensive case! 182 Constant *UnswitchVal = SI->getCaseValue(1); 183 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { 184 ++NumSwitches; 185 return true; 186 } 187 } 188 } 189 190 // Scan the instructions to check for unswitchable values. 191 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 192 BBI != E; ++BBI) 193 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 194 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 195 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::True, L)) { 196 ++NumSelects; 197 return true; 198 } 199 } 200 } 201 202 return Changed; 203} 204 205 206/// LoopValuesUsedOutsideLoop - Return true if there are any values defined in 207/// the loop that are used by instructions outside of it. 208static bool LoopValuesUsedOutsideLoop(Loop *L) { 209 // We will be doing lots of "loop contains block" queries. Loop::contains is 210 // linear time, use a set to speed this up. 211 std::set<BasicBlock*> LoopBlocks; 212 213 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 214 BB != E; ++BB) 215 LoopBlocks.insert(*BB); 216 217 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 218 BB != E; ++BB) { 219 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ++I) 220 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; 221 ++UI) { 222 BasicBlock *UserBB = cast<Instruction>(*UI)->getParent(); 223 if (!LoopBlocks.count(UserBB)) 224 return true; 225 } 226 } 227 return false; 228} 229 230/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 231/// 1. Exit the loop with no side effects. 232/// 2. Branch to the latch block with no side-effects. 233/// 234/// If these conditions are true, we return true and set ExitBB to the block we 235/// exit through. 236/// 237static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 238 BasicBlock *&ExitBB, 239 std::set<BasicBlock*> &Visited) { 240 if (!Visited.insert(BB).second) { 241 // Already visited and Ok, end of recursion. 242 return true; 243 } else if (!L->contains(BB)) { 244 // Otherwise, this is a loop exit, this is fine so long as this is the 245 // first exit. 246 if (ExitBB != 0) return false; 247 ExitBB = BB; 248 return true; 249 } 250 251 // Otherwise, this is an unvisited intra-loop node. Check all successors. 252 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 253 // Check to see if the successor is a trivial loop exit. 254 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 255 return false; 256 } 257 258 // Okay, everything after this looks good, check to make sure that this block 259 // doesn't include any side effects. 260 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 261 if (I->mayWriteToMemory()) 262 return false; 263 264 return true; 265} 266 267/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 268/// leads to an exit from the specified loop, and has no side-effects in the 269/// process. If so, return the block that is exited to, otherwise return null. 270static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 271 std::set<BasicBlock*> Visited; 272 Visited.insert(L->getHeader()); // Branches to header are ok. 273 BasicBlock *ExitBB = 0; 274 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 275 return ExitBB; 276 return 0; 277} 278 279/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 280/// trivial: that is, that the condition controls whether or not the loop does 281/// anything at all. If this is a trivial condition, unswitching produces no 282/// code duplications (equivalently, it produces a simpler loop and a new empty 283/// loop, which gets deleted). 284/// 285/// If this is a trivial condition, return true, otherwise return false. When 286/// returning true, this sets Cond and Val to the condition that controls the 287/// trivial condition: when Cond dynamically equals Val, the loop is known to 288/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 289/// Cond == Val. 290/// 291static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, 292 BasicBlock **LoopExit = 0) { 293 BasicBlock *Header = L->getHeader(); 294 TerminatorInst *HeaderTerm = Header->getTerminator(); 295 296 BasicBlock *LoopExitBB = 0; 297 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 298 // If the header block doesn't end with a conditional branch on Cond, we 299 // can't handle it. 300 if (!BI->isConditional() || BI->getCondition() != Cond) 301 return false; 302 303 // Check to see if a successor of the branch is guaranteed to go to the 304 // latch block or exit through a one exit block without having any 305 // side-effects. If so, determine the value of Cond that causes it to do 306 // this. 307 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { 308 if (Val) *Val = ConstantBool::True; 309 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { 310 if (Val) *Val = ConstantBool::False; 311 } 312 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 313 // If this isn't a switch on Cond, we can't handle it. 314 if (SI->getCondition() != Cond) return false; 315 316 // Check to see if a successor of the switch is guaranteed to go to the 317 // latch block or exit through a one exit block without having any 318 // side-effects. If so, determine the value of Cond that causes it to do 319 // this. Note that we can't trivially unswitch on the default case. 320 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 321 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) { 322 // Okay, we found a trivial case, remember the value that is trivial. 323 if (Val) *Val = SI->getCaseValue(i); 324 break; 325 } 326 } 327 328 // If we didn't find a single unique LoopExit block, or if the loop exit block 329 // contains phi nodes, this isn't trivial. 330 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 331 return false; // Can't handle this. 332 333 if (LoopExit) *LoopExit = LoopExitBB; 334 335 // We already know that nothing uses any scalar values defined inside of this 336 // loop. As such, we just have to check to see if this loop will execute any 337 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 338 // part of the loop that the code *would* execute. We already checked the 339 // tail, check the header now. 340 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 341 if (I->mayWriteToMemory()) 342 return false; 343 return true; 344} 345 346/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 347/// we choose to unswitch the specified loop on the specified value. 348/// 349unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { 350 // If the condition is trivial, always unswitch. There is no code growth for 351 // this case. 352 if (IsTrivialUnswitchCondition(L, LIC)) 353 return 0; 354 355 unsigned Cost = 0; 356 // FIXME: this is brain dead. It should take into consideration code 357 // shrinkage. 358 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 359 I != E; ++I) { 360 BasicBlock *BB = *I; 361 // Do not include empty blocks in the cost calculation. This happen due to 362 // loop canonicalization and will be removed. 363 if (BB->begin() == BasicBlock::iterator(BB->getTerminator())) 364 continue; 365 366 // Count basic blocks. 367 ++Cost; 368 } 369 370 return Cost; 371} 372 373/// UnswitchIfProfitable - We have found that we can unswitch L when 374/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 375/// unswitch the loop, reprocess the pieces, then return true. 376bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ 377 // Check to see if it would be profitable to unswitch this loop. 378 if (getLoopUnswitchCost(L, LoopCond) > Threshold) { 379 // FIXME: this should estimate growth by the amount of code shared by the 380 // resultant unswitched loops. 381 // 382 DEBUG(std::cerr << "NOT unswitching loop %" 383 << L->getHeader()->getName() << ", cost too high: " 384 << L->getBlocks().size() << "\n"); 385 return false; 386 } 387 388 // If this loop has live-out values, we can't unswitch it. We need something 389 // like loop-closed SSA form in order to know how to insert PHI nodes for 390 // these values. 391 if (LoopValuesUsedOutsideLoop(L)) { 392 DEBUG(std::cerr << "NOT unswitching loop %" << L->getHeader()->getName() 393 << ", a loop value is used outside loop!\n"); 394 return false; 395 } 396 397 // If this is a trivial condition to unswitch (which results in no code 398 // duplication), do it now. 399 Constant *CondVal; 400 BasicBlock *ExitBlock; 401 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) { 402 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock); 403 } else { 404 UnswitchNontrivialCondition(LoopCond, Val, L); 405 } 406 407 return true; 408} 409 410/// SplitBlock - Split the specified block at the specified instruction - every 411/// thing before SplitPt stays in Old and everything starting with SplitPt moves 412/// to a new block. The two blocks are joined by an unconditional branch and 413/// the loop info is updated. 414/// 415BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) { 416 BasicBlock::iterator SplitIt = SplitPt; 417 while (isa<PHINode>(SplitIt)) 418 ++SplitIt; 419 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split"); 420 421 // The new block lives in whichever loop the old one did. 422 if (Loop *L = LI->getLoopFor(Old)) 423 L->addBasicBlockToLoop(New, *LI); 424 425 return New; 426} 427 428 429BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) { 430 TerminatorInst *LatchTerm = BB->getTerminator(); 431 unsigned SuccNum = 0; 432 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) { 433 assert(i != e && "Didn't find edge?"); 434 if (LatchTerm->getSuccessor(i) == Succ) { 435 SuccNum = i; 436 break; 437 } 438 } 439 440 // If this is a critical edge, let SplitCriticalEdge do it. 441 if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this)) 442 return LatchTerm->getSuccessor(SuccNum); 443 444 // If the edge isn't critical, then BB has a single successor or Succ has a 445 // single pred. Split the block. 446 BasicBlock::iterator SplitPoint; 447 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 448 // If the successor only has a single pred, split the top of the successor 449 // block. 450 assert(SP == BB && "CFG broken"); 451 return SplitBlock(Succ, Succ->begin()); 452 } else { 453 // Otherwise, if BB has a single successor, split it at the bottom of the 454 // block. 455 assert(BB->getTerminator()->getNumSuccessors() == 1 && 456 "Should have a single succ!"); 457 return SplitBlock(BB, BB->getTerminator()); 458 } 459} 460 461 462 463// RemapInstruction - Convert the instruction operands from referencing the 464// current values into those specified by ValueMap. 465// 466static inline void RemapInstruction(Instruction *I, 467 std::map<const Value *, Value*> &ValueMap) { 468 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 469 Value *Op = I->getOperand(op); 470 std::map<const Value *, Value*>::iterator It = ValueMap.find(Op); 471 if (It != ValueMap.end()) Op = It->second; 472 I->setOperand(op, Op); 473 } 474} 475 476/// CloneLoop - Recursively clone the specified loop and all of its children, 477/// mapping the blocks with the specified map. 478static Loop *CloneLoop(Loop *L, Loop *PL, std::map<const Value*, Value*> &VM, 479 LoopInfo *LI) { 480 Loop *New = new Loop(); 481 482 if (PL) 483 PL->addChildLoop(New); 484 else 485 LI->addTopLevelLoop(New); 486 487 // Add all of the blocks in L to the new loop. 488 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 489 I != E; ++I) 490 if (LI->getLoopFor(*I) == L) 491 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 492 493 // Add all of the subloops to the new loop. 494 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 495 CloneLoop(*I, New, VM, LI); 496 497 return New; 498} 499 500/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 501/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 502/// code immediately before InsertPt. 503static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 504 BasicBlock *TrueDest, 505 BasicBlock *FalseDest, 506 Instruction *InsertPt) { 507 // Insert a conditional branch on LIC to the two preheaders. The original 508 // code is the true version and the new code is the false version. 509 Value *BranchVal = LIC; 510 if (!isa<ConstantBool>(Val)) { 511 BranchVal = BinaryOperator::createSetEQ(LIC, Val, "tmp", InsertPt); 512 } else if (Val != ConstantBool::True) { 513 // We want to enter the new loop when the condition is true. 514 std::swap(TrueDest, FalseDest); 515 } 516 517 // Insert the new branch. 518 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); 519} 520 521 522/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 523/// condition in it (a cond branch from its header block to its latch block, 524/// where the path through the loop that doesn't execute its body has no 525/// side-effects), unswitch it. This doesn't involve any code duplication, just 526/// moving the conditional branch outside of the loop and updating loop info. 527void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 528 Constant *Val, 529 BasicBlock *ExitBlock) { 530 DEBUG(std::cerr << "loop-unswitch: Trivial-Unswitch loop %" 531 << L->getHeader()->getName() << " [" << L->getBlocks().size() 532 << " blocks] in Function " << L->getHeader()->getParent()->getName() 533 << " on cond: " << *Val << " == " << *Cond << "\n"); 534 535 // First step, split the preheader, so that we know that there is a safe place 536 // to insert the conditional branch. We will change 'OrigPH' to have a 537 // conditional branch on Cond. 538 BasicBlock *OrigPH = L->getLoopPreheader(); 539 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); 540 541 // Now that we have a place to insert the conditional branch, create a place 542 // to branch to: this is the exit block out of the loop that we should 543 // short-circuit to. 544 545 // Split this block now, so that the loop maintains its exit block, and so 546 // that the jump from the preheader can execute the contents of the exit block 547 // without actually branching to it (the exit block should be dominated by the 548 // loop header, not the preheader). 549 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 550 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); 551 552 // Okay, now we have a position to branch from and a position to branch to, 553 // insert the new conditional branch. 554 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 555 OrigPH->getTerminator()); 556 OrigPH->getTerminator()->eraseFromParent(); 557 558 // We need to reprocess this loop, it could be unswitched again. 559 LoopProcessWorklist.push_back(L); 560 561 // Now that we know that the loop is never entered when this condition is a 562 // particular value, rewrite the loop with this info. We know that this will 563 // at least eliminate the old branch. 564 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 565 ++NumTrivial; 566} 567 568 569/// VersionLoop - We determined that the loop is profitable to unswitch when LIC 570/// equal Val. Split it into loop versions and test the condition outside of 571/// either loop. Return the loops created as Out1/Out2. 572void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 573 Loop *L) { 574 Function *F = L->getHeader()->getParent(); 575 DEBUG(std::cerr << "loop-unswitch: Unswitching loop %" 576 << L->getHeader()->getName() << " [" << L->getBlocks().size() 577 << " blocks] in Function " << F->getName() 578 << " when '" << *Val << "' == " << *LIC << "\n"); 579 580 // LoopBlocks contains all of the basic blocks of the loop, including the 581 // preheader of the loop, the body of the loop, and the exit blocks of the 582 // loop, in that order. 583 std::vector<BasicBlock*> LoopBlocks; 584 585 // First step, split the preheader and exit blocks, and add these blocks to 586 // the LoopBlocks list. 587 BasicBlock *OrigPreheader = L->getLoopPreheader(); 588 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); 589 590 // We want the loop to come after the preheader, but before the exit blocks. 591 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 592 593 std::vector<BasicBlock*> ExitBlocks; 594 L->getExitBlocks(ExitBlocks); 595 std::sort(ExitBlocks.begin(), ExitBlocks.end()); 596 ExitBlocks.erase(std::unique(ExitBlocks.begin(), ExitBlocks.end()), 597 ExitBlocks.end()); 598 599 // Split all of the edges from inside the loop to their exit blocks. This 600 // unswitching trivial: no phi nodes to update. 601 unsigned NumBlocks = L->getBlocks().size(); 602 603 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 604 BasicBlock *ExitBlock = ExitBlocks[i]; 605 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 606 607 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 608 assert(L->contains(Preds[j]) && 609 "All preds of loop exit blocks must be the same loop!"); 610 SplitEdge(Preds[j], ExitBlock); 611 } 612 } 613 614 // The exit blocks may have been changed due to edge splitting, recompute. 615 ExitBlocks.clear(); 616 L->getExitBlocks(ExitBlocks); 617 std::sort(ExitBlocks.begin(), ExitBlocks.end()); 618 ExitBlocks.erase(std::unique(ExitBlocks.begin(), ExitBlocks.end()), 619 ExitBlocks.end()); 620 621 // Add exit blocks to the loop blocks. 622 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 623 624 // Next step, clone all of the basic blocks that make up the loop (including 625 // the loop preheader and exit blocks), keeping track of the mapping between 626 // the instructions and blocks. 627 std::vector<BasicBlock*> NewBlocks; 628 NewBlocks.reserve(LoopBlocks.size()); 629 std::map<const Value*, Value*> ValueMap; 630 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 631 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 632 NewBlocks.push_back(New); 633 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 634 } 635 636 // Splice the newly inserted blocks into the function right before the 637 // original preheader. 638 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 639 NewBlocks[0], F->end()); 640 641 // Now we create the new Loop object for the versioned loop. 642 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI); 643 Loop *ParentLoop = L->getParentLoop(); 644 if (ParentLoop) { 645 // Make sure to add the cloned preheader and exit blocks to the parent loop 646 // as well. 647 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 648 } 649 650 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 651 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 652 // The new exit block should be in the same loop as the old one. 653 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 654 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 655 656 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 657 "Exit block should have been split to have one successor!"); 658 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 659 660 // If the successor of the exit block had PHI nodes, add an entry for 661 // NewExit. 662 PHINode *PN; 663 for (BasicBlock::iterator I = ExitSucc->begin(); 664 (PN = dyn_cast<PHINode>(I)); ++I) { 665 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 666 std::map<const Value *, Value*>::iterator It = ValueMap.find(V); 667 if (It != ValueMap.end()) V = It->second; 668 PN->addIncoming(V, NewExit); 669 } 670 } 671 672 // Rewrite the code to refer to itself. 673 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 674 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 675 E = NewBlocks[i]->end(); I != E; ++I) 676 RemapInstruction(I, ValueMap); 677 678 // Rewrite the original preheader to select between versions of the loop. 679 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 680 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 681 "Preheader splitting did not work correctly!"); 682 683 // Emit the new branch that selects between the two versions of this loop. 684 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 685 OldBR->eraseFromParent(); 686 687 LoopProcessWorklist.push_back(L); 688 LoopProcessWorklist.push_back(NewLoop); 689 690 // Now we rewrite the original code to know that the condition is true and the 691 // new code to know that the condition is false. 692 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 693 694 // It's possible that simplifying one loop could cause the other to be 695 // deleted. If so, don't simplify it. 696 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 697 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 698} 699 700/// RemoveFromWorklist - Remove all instances of I from the worklist vector 701/// specified. 702static void RemoveFromWorklist(Instruction *I, 703 std::vector<Instruction*> &Worklist) { 704 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 705 Worklist.end(), I); 706 while (WI != Worklist.end()) { 707 unsigned Offset = WI-Worklist.begin(); 708 Worklist.erase(WI); 709 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 710 } 711} 712 713/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 714/// program, replacing all uses with V and update the worklist. 715static void ReplaceUsesOfWith(Instruction *I, Value *V, 716 std::vector<Instruction*> &Worklist) { 717 DEBUG(std::cerr << "Replace with '" << *V << "': " << *I); 718 719 // Add uses to the worklist, which may be dead now. 720 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 721 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 722 Worklist.push_back(Use); 723 724 // Add users to the worklist which may be simplified now. 725 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 726 UI != E; ++UI) 727 Worklist.push_back(cast<Instruction>(*UI)); 728 I->replaceAllUsesWith(V); 729 I->eraseFromParent(); 730 RemoveFromWorklist(I, Worklist); 731 ++NumSimplify; 732} 733 734/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 735/// information, and remove any dead successors it has. 736/// 737void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 738 std::vector<Instruction*> &Worklist) { 739 if (pred_begin(BB) != pred_end(BB)) { 740 // This block isn't dead, since an edge to BB was just removed, see if there 741 // are any easy simplifications we can do now. 742 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 743 // If it has one pred, fold phi nodes in BB. 744 while (isa<PHINode>(BB->begin())) 745 ReplaceUsesOfWith(BB->begin(), 746 cast<PHINode>(BB->begin())->getIncomingValue(0), 747 Worklist); 748 749 // If this is the header of a loop and the only pred is the latch, we now 750 // have an unreachable loop. 751 if (Loop *L = LI->getLoopFor(BB)) 752 if (L->getHeader() == BB && L->contains(Pred)) { 753 // Remove the branch from the latch to the header block, this makes 754 // the header dead, which will make the latch dead (because the header 755 // dominates the latch). 756 Pred->getTerminator()->eraseFromParent(); 757 new UnreachableInst(Pred); 758 759 // The loop is now broken, remove it from LI. 760 RemoveLoopFromHierarchy(L); 761 762 // Reprocess the header, which now IS dead. 763 RemoveBlockIfDead(BB, Worklist); 764 return; 765 } 766 767 // If pred ends in a uncond branch, add uncond branch to worklist so that 768 // the two blocks will get merged. 769 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 770 if (BI->isUnconditional()) 771 Worklist.push_back(BI); 772 } 773 return; 774 } 775 776 DEBUG(std::cerr << "Nuking dead block: " << *BB); 777 778 // Remove the instructions in the basic block from the worklist. 779 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 780 RemoveFromWorklist(I, Worklist); 781 782 // Anything that uses the instructions in this basic block should have their 783 // uses replaced with undefs. 784 if (!I->use_empty()) 785 I->replaceAllUsesWith(UndefValue::get(I->getType())); 786 } 787 788 // If this is the edge to the header block for a loop, remove the loop and 789 // promote all subloops. 790 if (Loop *BBLoop = LI->getLoopFor(BB)) { 791 if (BBLoop->getLoopLatch() == BB) 792 RemoveLoopFromHierarchy(BBLoop); 793 } 794 795 // Remove the block from the loop info, which removes it from any loops it 796 // was in. 797 LI->removeBlock(BB); 798 799 800 // Remove phi node entries in successors for this block. 801 TerminatorInst *TI = BB->getTerminator(); 802 std::vector<BasicBlock*> Succs; 803 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 804 Succs.push_back(TI->getSuccessor(i)); 805 TI->getSuccessor(i)->removePredecessor(BB); 806 } 807 808 // Unique the successors, remove anything with multiple uses. 809 std::sort(Succs.begin(), Succs.end()); 810 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 811 812 // Remove the basic block, including all of the instructions contained in it. 813 BB->eraseFromParent(); 814 815 // Remove successor blocks here that are not dead, so that we know we only 816 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 817 // then getting removed before we revisit them, which is badness. 818 // 819 for (unsigned i = 0; i != Succs.size(); ++i) 820 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 821 // One exception is loop headers. If this block was the preheader for a 822 // loop, then we DO want to visit the loop so the loop gets deleted. 823 // We know that if the successor is a loop header, that this loop had to 824 // be the preheader: the case where this was the latch block was handled 825 // above and headers can only have two predecessors. 826 if (!LI->isLoopHeader(Succs[i])) { 827 Succs.erase(Succs.begin()+i); 828 --i; 829 } 830 } 831 832 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 833 RemoveBlockIfDead(Succs[i], Worklist); 834} 835 836/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 837/// become unwrapped, either because the backedge was deleted, or because the 838/// edge into the header was removed. If the edge into the header from the 839/// latch block was removed, the loop is unwrapped but subloops are still alive, 840/// so they just reparent loops. If the loops are actually dead, they will be 841/// removed later. 842void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 843 if (Loop *ParentLoop = L->getParentLoop()) { // Not a top-level loop. 844 // Reparent all of the blocks in this loop. Since BBLoop had a parent, 845 // they are now all in it. 846 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 847 I != E; ++I) 848 if (LI->getLoopFor(*I) == L) // Don't change blocks in subloops. 849 LI->changeLoopFor(*I, ParentLoop); 850 851 // Remove the loop from its parent loop. 852 for (Loop::iterator I = ParentLoop->begin(), E = ParentLoop->end();; 853 ++I) { 854 assert(I != E && "Couldn't find loop"); 855 if (*I == L) { 856 ParentLoop->removeChildLoop(I); 857 break; 858 } 859 } 860 861 // Move all subloops into the parent loop. 862 while (L->begin() != L->end()) 863 ParentLoop->addChildLoop(L->removeChildLoop(L->end()-1)); 864 } else { 865 // Reparent all of the blocks in this loop. Since BBLoop had no parent, 866 // they no longer in a loop at all. 867 868 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 869 // Don't change blocks in subloops. 870 if (LI->getLoopFor(L->getBlocks()[i]) == L) { 871 LI->removeBlock(L->getBlocks()[i]); 872 --i; 873 } 874 } 875 876 // Remove the loop from the top-level LoopInfo object. 877 for (LoopInfo::iterator I = LI->begin(), E = LI->end();; ++I) { 878 assert(I != E && "Couldn't find loop"); 879 if (*I == L) { 880 LI->removeLoop(I); 881 break; 882 } 883 } 884 885 // Move all of the subloops to the top-level. 886 while (L->begin() != L->end()) 887 LI->addTopLevelLoop(L->removeChildLoop(L->end()-1)); 888 } 889 890 delete L; 891 RemoveLoopFromWorklist(L); 892} 893 894 895 896// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 897// the value specified by Val in the specified loop, or we know it does NOT have 898// that value. Rewrite any uses of LIC or of properties correlated to it. 899void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 900 Constant *Val, 901 bool IsEqual) { 902 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 903 904 // FIXME: Support correlated properties, like: 905 // for (...) 906 // if (li1 < li2) 907 // ... 908 // if (li1 > li2) 909 // ... 910 911 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 912 // selects, switches. 913 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 914 std::vector<Instruction*> Worklist; 915 916 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 917 // in the loop with the appropriate one directly. 918 if (IsEqual || isa<ConstantBool>(Val)) { 919 Value *Replacement; 920 if (IsEqual) 921 Replacement = Val; 922 else 923 Replacement = ConstantBool::get(!cast<ConstantBool>(Val)->getValue()); 924 925 for (unsigned i = 0, e = Users.size(); i != e; ++i) 926 if (Instruction *U = cast<Instruction>(Users[i])) { 927 if (!L->contains(U->getParent())) 928 continue; 929 U->replaceUsesOfWith(LIC, Replacement); 930 Worklist.push_back(U); 931 } 932 } else { 933 // Otherwise, we don't know the precise value of LIC, but we do know that it 934 // is certainly NOT "Val". As such, simplify any uses in the loop that we 935 // can. This case occurs when we unswitch switch statements. 936 for (unsigned i = 0, e = Users.size(); i != e; ++i) 937 if (Instruction *U = cast<Instruction>(Users[i])) { 938 if (!L->contains(U->getParent())) 939 continue; 940 941 Worklist.push_back(U); 942 943 // If we know that LIC is not Val, use this info to simplify code. 944 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 945 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 946 if (SI->getCaseValue(i) == Val) { 947 // Found a dead case value. Don't remove PHI nodes in the 948 // successor if they become single-entry, those PHI nodes may 949 // be in the Users list. 950 SI->getSuccessor(i)->removePredecessor(SI->getParent(), true); 951 SI->removeCase(i); 952 break; 953 } 954 } 955 } 956 957 // TODO: We could do other simplifications, for example, turning 958 // LIC == Val -> false. 959 } 960 } 961 962 SimplifyCode(Worklist); 963} 964 965/// SimplifyCode - Okay, now that we have simplified some instructions in the 966/// loop, walk over it and constant prop, dce, and fold control flow where 967/// possible. Note that this is effectively a very simple loop-structure-aware 968/// optimizer. During processing of this loop, L could very well be deleted, so 969/// it must not be used. 970/// 971/// FIXME: When the loop optimizer is more mature, separate this out to a new 972/// pass. 973/// 974void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) { 975 while (!Worklist.empty()) { 976 Instruction *I = Worklist.back(); 977 Worklist.pop_back(); 978 979 // Simple constant folding. 980 if (Constant *C = ConstantFoldInstruction(I)) { 981 ReplaceUsesOfWith(I, C, Worklist); 982 continue; 983 } 984 985 // Simple DCE. 986 if (isInstructionTriviallyDead(I)) { 987 DEBUG(std::cerr << "Remove dead instruction '" << *I); 988 989 // Add uses to the worklist, which may be dead now. 990 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 991 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 992 Worklist.push_back(Use); 993 I->eraseFromParent(); 994 RemoveFromWorklist(I, Worklist); 995 ++NumSimplify; 996 continue; 997 } 998 999 // Special case hacks that appear commonly in unswitched code. 1000 switch (I->getOpcode()) { 1001 case Instruction::Select: 1002 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(0))) { 1003 ReplaceUsesOfWith(I, I->getOperand(!CB->getValue()+1), Worklist); 1004 continue; 1005 } 1006 break; 1007 case Instruction::And: 1008 if (isa<ConstantBool>(I->getOperand(0))) // constant -> RHS 1009 cast<BinaryOperator>(I)->swapOperands(); 1010 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(1))) { 1011 if (CB->getValue()) // X & 1 -> X 1012 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1013 else // X & 0 -> 0 1014 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1015 continue; 1016 } 1017 break; 1018 case Instruction::Or: 1019 if (isa<ConstantBool>(I->getOperand(0))) // constant -> RHS 1020 cast<BinaryOperator>(I)->swapOperands(); 1021 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(1))) { 1022 if (CB->getValue()) // X | 1 -> 1 1023 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1024 else // X | 0 -> X 1025 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1026 continue; 1027 } 1028 break; 1029 case Instruction::Br: { 1030 BranchInst *BI = cast<BranchInst>(I); 1031 if (BI->isUnconditional()) { 1032 // If BI's parent is the only pred of the successor, fold the two blocks 1033 // together. 1034 BasicBlock *Pred = BI->getParent(); 1035 BasicBlock *Succ = BI->getSuccessor(0); 1036 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1037 if (!SinglePred) continue; // Nothing to do. 1038 assert(SinglePred == Pred && "CFG broken"); 1039 1040 DEBUG(std::cerr << "Merging blocks: " << Pred->getName() << " <- " 1041 << Succ->getName() << "\n"); 1042 1043 // Resolve any single entry PHI nodes in Succ. 1044 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1045 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); 1046 1047 // Move all of the successor contents from Succ to Pred. 1048 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1049 Succ->end()); 1050 BI->eraseFromParent(); 1051 RemoveFromWorklist(BI, Worklist); 1052 1053 // If Succ has any successors with PHI nodes, update them to have 1054 // entries coming from Pred instead of Succ. 1055 Succ->replaceAllUsesWith(Pred); 1056 1057 // Remove Succ from the loop tree. 1058 LI->removeBlock(Succ); 1059 Succ->eraseFromParent(); 1060 ++NumSimplify; 1061 } else if (ConstantBool *CB = dyn_cast<ConstantBool>(BI->getCondition())){ 1062 // Conditional branch. Turn it into an unconditional branch, then 1063 // remove dead blocks. 1064 break; // FIXME: Enable. 1065 1066 DEBUG(std::cerr << "Folded branch: " << *BI); 1067 BasicBlock *DeadSucc = BI->getSuccessor(CB->getValue()); 1068 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getValue()); 1069 DeadSucc->removePredecessor(BI->getParent(), true); 1070 Worklist.push_back(new BranchInst(LiveSucc, BI)); 1071 BI->eraseFromParent(); 1072 RemoveFromWorklist(BI, Worklist); 1073 ++NumSimplify; 1074 1075 RemoveBlockIfDead(DeadSucc, Worklist); 1076 } 1077 break; 1078 } 1079 } 1080 } 1081} 1082