LoopUnswitch.cpp revision c08b56f6a085f822e21bf3aa18a0554ed9ecd479
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 <set> 44using namespace llvm; 45 46STATISTIC(NumBranches, "Number of branches unswitched"); 47STATISTIC(NumSwitches, "Number of switches unswitched"); 48STATISTIC(NumSelects , "Number of selects unswitched"); 49STATISTIC(NumTrivial , "Number of unswitches that are trivial"); 50STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); 51 52namespace { 53 cl::opt<unsigned> 54 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), 55 cl::init(10), cl::Hidden); 56 57 class LoopUnswitch : public FunctionPass { 58 LoopInfo *LI; // Loop information 59 60 // LoopProcessWorklist - List of loops we need to process. 61 std::vector<Loop*> LoopProcessWorklist; 62 public: 63 virtual bool runOnFunction(Function &F); 64 bool visitLoop(Loop *L); 65 66 /// This transformation requires natural loop information & requires that 67 /// loop preheaders be inserted into the CFG... 68 /// 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.addRequiredID(LoopSimplifyID); 71 AU.addPreservedID(LoopSimplifyID); 72 AU.addRequired<LoopInfo>(); 73 AU.addPreserved<LoopInfo>(); 74 AU.addRequiredID(LCSSAID); 75 AU.addPreservedID(LCSSAID); 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 RegisterPass<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 assert(L->isLCSSAForm()); 158 159 bool Changed = false; 160 161 // Loop over all of the basic blocks in the loop. If we find an interior 162 // block that is branching on a loop-invariant condition, we can unswitch this 163 // loop. 164 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 165 I != E; ++I) { 166 TerminatorInst *TI = (*I)->getTerminator(); 167 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 168 // If this isn't branching on an invariant condition, we can't unswitch 169 // it. 170 if (BI->isConditional()) { 171 // See if this, or some part of it, is loop invariant. If so, we can 172 // unswitch on it if we desire. 173 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); 174 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 175 L)) { 176 ++NumBranches; 177 return true; 178 } 179 } 180 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 181 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 182 if (LoopCond && SI->getNumCases() > 1) { 183 // Find a value to unswitch on: 184 // FIXME: this should chose the most expensive case! 185 Constant *UnswitchVal = SI->getCaseValue(1); 186 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { 187 ++NumSwitches; 188 return true; 189 } 190 } 191 } 192 193 // Scan the instructions to check for unswitchable values. 194 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); 195 BBI != E; ++BBI) 196 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) { 197 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); 198 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), 199 L)) { 200 ++NumSelects; 201 return true; 202 } 203 } 204 } 205 206 assert(L->isLCSSAForm()); 207 208 return Changed; 209} 210 211/// isTrivialLoopExitBlock - Check to see if all paths from BB either: 212/// 1. Exit the loop with no side effects. 213/// 2. Branch to the latch block with no side-effects. 214/// 215/// If these conditions are true, we return true and set ExitBB to the block we 216/// exit through. 217/// 218static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, 219 BasicBlock *&ExitBB, 220 std::set<BasicBlock*> &Visited) { 221 if (!Visited.insert(BB).second) { 222 // Already visited and Ok, end of recursion. 223 return true; 224 } else if (!L->contains(BB)) { 225 // Otherwise, this is a loop exit, this is fine so long as this is the 226 // first exit. 227 if (ExitBB != 0) return false; 228 ExitBB = BB; 229 return true; 230 } 231 232 // Otherwise, this is an unvisited intra-loop node. Check all successors. 233 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { 234 // Check to see if the successor is a trivial loop exit. 235 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited)) 236 return false; 237 } 238 239 // Okay, everything after this looks good, check to make sure that this block 240 // doesn't include any side effects. 241 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 242 if (I->mayWriteToMemory()) 243 return false; 244 245 return true; 246} 247 248/// isTrivialLoopExitBlock - Return true if the specified block unconditionally 249/// leads to an exit from the specified loop, and has no side-effects in the 250/// process. If so, return the block that is exited to, otherwise return null. 251static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { 252 std::set<BasicBlock*> Visited; 253 Visited.insert(L->getHeader()); // Branches to header are ok. 254 BasicBlock *ExitBB = 0; 255 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) 256 return ExitBB; 257 return 0; 258} 259 260/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is 261/// trivial: that is, that the condition controls whether or not the loop does 262/// anything at all. If this is a trivial condition, unswitching produces no 263/// code duplications (equivalently, it produces a simpler loop and a new empty 264/// loop, which gets deleted). 265/// 266/// If this is a trivial condition, return true, otherwise return false. When 267/// returning true, this sets Cond and Val to the condition that controls the 268/// trivial condition: when Cond dynamically equals Val, the loop is known to 269/// exit. Finally, this sets LoopExit to the BB that the loop exits to when 270/// Cond == Val. 271/// 272static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, 273 BasicBlock **LoopExit = 0) { 274 BasicBlock *Header = L->getHeader(); 275 TerminatorInst *HeaderTerm = Header->getTerminator(); 276 277 BasicBlock *LoopExitBB = 0; 278 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) { 279 // If the header block doesn't end with a conditional branch on Cond, we 280 // can't handle it. 281 if (!BI->isConditional() || BI->getCondition() != Cond) 282 return false; 283 284 // Check to see if a successor of the branch is guaranteed to go to the 285 // latch block or exit through a one exit block without having any 286 // side-effects. If so, determine the value of Cond that causes it to do 287 // this. 288 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { 289 if (Val) *Val = ConstantInt::getTrue(); 290 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { 291 if (Val) *Val = ConstantInt::getFalse(); 292 } 293 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) { 294 // If this isn't a switch on Cond, we can't handle it. 295 if (SI->getCondition() != Cond) return false; 296 297 // Check to see if a successor of the switch is guaranteed to go to the 298 // latch block or exit through a one exit block without having any 299 // side-effects. If so, determine the value of Cond that causes it to do 300 // this. Note that we can't trivially unswitch on the default case. 301 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) 302 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) { 303 // Okay, we found a trivial case, remember the value that is trivial. 304 if (Val) *Val = SI->getCaseValue(i); 305 break; 306 } 307 } 308 309 // If we didn't find a single unique LoopExit block, or if the loop exit block 310 // contains phi nodes, this isn't trivial. 311 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin())) 312 return false; // Can't handle this. 313 314 if (LoopExit) *LoopExit = LoopExitBB; 315 316 // We already know that nothing uses any scalar values defined inside of this 317 // loop. As such, we just have to check to see if this loop will execute any 318 // side-effecting instructions (e.g. stores, calls, volatile loads) in the 319 // part of the loop that the code *would* execute. We already checked the 320 // tail, check the header now. 321 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) 322 if (I->mayWriteToMemory()) 323 return false; 324 return true; 325} 326 327/// getLoopUnswitchCost - Return the cost (code size growth) that will happen if 328/// we choose to unswitch the specified loop on the specified value. 329/// 330unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { 331 // If the condition is trivial, always unswitch. There is no code growth for 332 // this case. 333 if (IsTrivialUnswitchCondition(L, LIC)) 334 return 0; 335 336 // FIXME: This is really overly conservative. However, more liberal 337 // estimations have thus far resulted in excessive unswitching, which is bad 338 // both in compile time and in code size. This should be replaced once 339 // someone figures out how a good estimation. 340 return L->getBlocks().size(); 341 342 unsigned Cost = 0; 343 // FIXME: this is brain dead. It should take into consideration code 344 // shrinkage. 345 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 346 I != E; ++I) { 347 BasicBlock *BB = *I; 348 // Do not include empty blocks in the cost calculation. This happen due to 349 // loop canonicalization and will be removed. 350 if (BB->begin() == BasicBlock::iterator(BB->getTerminator())) 351 continue; 352 353 // Count basic blocks. 354 ++Cost; 355 } 356 357 return Cost; 358} 359 360/// UnswitchIfProfitable - We have found that we can unswitch L when 361/// LoopCond == Val to simplify the loop. If we decide that this is profitable, 362/// unswitch the loop, reprocess the pieces, then return true. 363bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ 364 // Check to see if it would be profitable to unswitch this loop. 365 unsigned Cost = getLoopUnswitchCost(L, LoopCond); 366 if (Cost > Threshold) { 367 // FIXME: this should estimate growth by the amount of code shared by the 368 // resultant unswitched loops. 369 // 370 DOUT << "NOT unswitching loop %" 371 << L->getHeader()->getName() << ", cost too high: " 372 << L->getBlocks().size() << "\n"; 373 return false; 374 } 375 376 // If this is a trivial condition to unswitch (which results in no code 377 // duplication), do it now. 378 Constant *CondVal; 379 BasicBlock *ExitBlock; 380 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) { 381 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock); 382 } else { 383 UnswitchNontrivialCondition(LoopCond, Val, L); 384 } 385 386 return true; 387} 388 389/// SplitBlock - Split the specified block at the specified instruction - every 390/// thing before SplitPt stays in Old and everything starting with SplitPt moves 391/// to a new block. The two blocks are joined by an unconditional branch and 392/// the loop info is updated. 393/// 394BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) { 395 BasicBlock::iterator SplitIt = SplitPt; 396 while (isa<PHINode>(SplitIt)) 397 ++SplitIt; 398 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split"); 399 400 // The new block lives in whichever loop the old one did. 401 if (Loop *L = LI->getLoopFor(Old)) 402 L->addBasicBlockToLoop(New, *LI); 403 404 return New; 405} 406 407 408BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) { 409 TerminatorInst *LatchTerm = BB->getTerminator(); 410 unsigned SuccNum = 0; 411 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) { 412 assert(i != e && "Didn't find edge?"); 413 if (LatchTerm->getSuccessor(i) == Succ) { 414 SuccNum = i; 415 break; 416 } 417 } 418 419 // If this is a critical edge, let SplitCriticalEdge do it. 420 if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this)) 421 return LatchTerm->getSuccessor(SuccNum); 422 423 // If the edge isn't critical, then BB has a single successor or Succ has a 424 // single pred. Split the block. 425 BasicBlock::iterator SplitPoint; 426 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 427 // If the successor only has a single pred, split the top of the successor 428 // block. 429 assert(SP == BB && "CFG broken"); 430 return SplitBlock(Succ, Succ->begin()); 431 } else { 432 // Otherwise, if BB has a single successor, split it at the bottom of the 433 // block. 434 assert(BB->getTerminator()->getNumSuccessors() == 1 && 435 "Should have a single succ!"); 436 return SplitBlock(BB, BB->getTerminator()); 437 } 438} 439 440 441 442// RemapInstruction - Convert the instruction operands from referencing the 443// current values into those specified by ValueMap. 444// 445static inline void RemapInstruction(Instruction *I, 446 std::map<const Value *, Value*> &ValueMap) { 447 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { 448 Value *Op = I->getOperand(op); 449 std::map<const Value *, Value*>::iterator It = ValueMap.find(Op); 450 if (It != ValueMap.end()) Op = It->second; 451 I->setOperand(op, Op); 452 } 453} 454 455/// CloneLoop - Recursively clone the specified loop and all of its children, 456/// mapping the blocks with the specified map. 457static Loop *CloneLoop(Loop *L, Loop *PL, std::map<const Value*, Value*> &VM, 458 LoopInfo *LI) { 459 Loop *New = new Loop(); 460 461 if (PL) 462 PL->addChildLoop(New); 463 else 464 LI->addTopLevelLoop(New); 465 466 // Add all of the blocks in L to the new loop. 467 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 468 I != E; ++I) 469 if (LI->getLoopFor(*I) == L) 470 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI); 471 472 // Add all of the subloops to the new loop. 473 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 474 CloneLoop(*I, New, VM, LI); 475 476 return New; 477} 478 479/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values 480/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the 481/// code immediately before InsertPt. 482static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, 483 BasicBlock *TrueDest, 484 BasicBlock *FalseDest, 485 Instruction *InsertPt) { 486 // Insert a conditional branch on LIC to the two preheaders. The original 487 // code is the true version and the new code is the false version. 488 Value *BranchVal = LIC; 489 if (Val->getType() != Type::Int1Ty || !isa<ConstantInt>(Val)) 490 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); 491 else if (Val != ConstantInt::getTrue()) 492 // We want to enter the new loop when the condition is true. 493 std::swap(TrueDest, FalseDest); 494 495 // Insert the new branch. 496 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); 497} 498 499 500/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable 501/// condition in it (a cond branch from its header block to its latch block, 502/// where the path through the loop that doesn't execute its body has no 503/// side-effects), unswitch it. This doesn't involve any code duplication, just 504/// moving the conditional branch outside of the loop and updating loop info. 505void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, 506 Constant *Val, 507 BasicBlock *ExitBlock) { 508 DOUT << "loop-unswitch: Trivial-Unswitch loop %" 509 << L->getHeader()->getName() << " [" << L->getBlocks().size() 510 << " blocks] in Function " << L->getHeader()->getParent()->getName() 511 << " on cond: " << *Val << " == " << *Cond << "\n"; 512 513 // First step, split the preheader, so that we know that there is a safe place 514 // to insert the conditional branch. We will change 'OrigPH' to have a 515 // conditional branch on Cond. 516 BasicBlock *OrigPH = L->getLoopPreheader(); 517 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); 518 519 // Now that we have a place to insert the conditional branch, create a place 520 // to branch to: this is the exit block out of the loop that we should 521 // short-circuit to. 522 523 // Split this block now, so that the loop maintains its exit block, and so 524 // that the jump from the preheader can execute the contents of the exit block 525 // without actually branching to it (the exit block should be dominated by the 526 // loop header, not the preheader). 527 assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); 528 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); 529 530 // Okay, now we have a position to branch from and a position to branch to, 531 // insert the new conditional branch. 532 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, 533 OrigPH->getTerminator()); 534 OrigPH->getTerminator()->eraseFromParent(); 535 536 // We need to reprocess this loop, it could be unswitched again. 537 LoopProcessWorklist.push_back(L); 538 539 // Now that we know that the loop is never entered when this condition is a 540 // particular value, rewrite the loop with this info. We know that this will 541 // at least eliminate the old branch. 542 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false); 543 ++NumTrivial; 544} 545 546 547/// VersionLoop - We determined that the loop is profitable to unswitch when LIC 548/// equal Val. Split it into loop versions and test the condition outside of 549/// either loop. Return the loops created as Out1/Out2. 550void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 551 Loop *L) { 552 Function *F = L->getHeader()->getParent(); 553 DOUT << "loop-unswitch: Unswitching loop %" 554 << L->getHeader()->getName() << " [" << L->getBlocks().size() 555 << " blocks] in Function " << F->getName() 556 << " when '" << *Val << "' == " << *LIC << "\n"; 557 558 // LoopBlocks contains all of the basic blocks of the loop, including the 559 // preheader of the loop, the body of the loop, and the exit blocks of the 560 // loop, in that order. 561 std::vector<BasicBlock*> LoopBlocks; 562 563 // First step, split the preheader and exit blocks, and add these blocks to 564 // the LoopBlocks list. 565 BasicBlock *OrigPreheader = L->getLoopPreheader(); 566 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); 567 568 // We want the loop to come after the preheader, but before the exit blocks. 569 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); 570 571 std::vector<BasicBlock*> ExitBlocks; 572 L->getUniqueExitBlocks(ExitBlocks); 573 574 // Split all of the edges from inside the loop to their exit blocks. Update 575 // the appropriate Phi nodes as we do so. 576 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 577 BasicBlock *ExitBlock = ExitBlocks[i]; 578 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); 579 580 for (unsigned j = 0, e = Preds.size(); j != e; ++j) { 581 assert(L->contains(Preds[j]) && 582 "All preds of loop exit blocks must be the same loop!"); 583 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock); 584 BasicBlock* StartBlock = Preds[j]; 585 BasicBlock* EndBlock; 586 if (MiddleBlock->getSinglePredecessor() == ExitBlock) { 587 EndBlock = MiddleBlock; 588 MiddleBlock = EndBlock->getSinglePredecessor();; 589 } else { 590 EndBlock = ExitBlock; 591 } 592 593 std::set<PHINode*> InsertedPHIs; 594 PHINode* OldLCSSA = 0; 595 for (BasicBlock::iterator I = EndBlock->begin(); 596 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) { 597 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock); 598 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(), 599 OldLCSSA->getName() + ".us-lcssa", 600 MiddleBlock->getTerminator()); 601 NewLCSSA->addIncoming(OldValue, StartBlock); 602 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock), 603 NewLCSSA); 604 InsertedPHIs.insert(NewLCSSA); 605 } 606 607 BasicBlock::iterator InsertPt = EndBlock->begin(); 608 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt; 609 for (BasicBlock::iterator I = MiddleBlock->begin(); 610 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0; 611 ++I) { 612 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(), 613 OldLCSSA->getName() + ".us-lcssa", 614 InsertPt); 615 OldLCSSA->replaceAllUsesWith(NewLCSSA); 616 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); 617 } 618 } 619 } 620 621 // The exit blocks may have been changed due to edge splitting, recompute. 622 ExitBlocks.clear(); 623 L->getUniqueExitBlocks(ExitBlocks); 624 625 // Add exit blocks to the loop blocks. 626 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end()); 627 628 // Next step, clone all of the basic blocks that make up the loop (including 629 // the loop preheader and exit blocks), keeping track of the mapping between 630 // the instructions and blocks. 631 std::vector<BasicBlock*> NewBlocks; 632 NewBlocks.reserve(LoopBlocks.size()); 633 std::map<const Value*, Value*> ValueMap; 634 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { 635 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); 636 NewBlocks.push_back(New); 637 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. 638 } 639 640 // Splice the newly inserted blocks into the function right before the 641 // original preheader. 642 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(), 643 NewBlocks[0], F->end()); 644 645 // Now we create the new Loop object for the versioned loop. 646 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI); 647 Loop *ParentLoop = L->getParentLoop(); 648 if (ParentLoop) { 649 // Make sure to add the cloned preheader and exit blocks to the parent loop 650 // as well. 651 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); 652 } 653 654 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 655 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]); 656 // The new exit block should be in the same loop as the old one. 657 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) 658 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); 659 660 assert(NewExit->getTerminator()->getNumSuccessors() == 1 && 661 "Exit block should have been split to have one successor!"); 662 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); 663 664 // If the successor of the exit block had PHI nodes, add an entry for 665 // NewExit. 666 PHINode *PN; 667 for (BasicBlock::iterator I = ExitSucc->begin(); 668 (PN = dyn_cast<PHINode>(I)); ++I) { 669 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); 670 std::map<const Value *, Value*>::iterator It = ValueMap.find(V); 671 if (It != ValueMap.end()) V = It->second; 672 PN->addIncoming(V, NewExit); 673 } 674 } 675 676 // Rewrite the code to refer to itself. 677 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) 678 for (BasicBlock::iterator I = NewBlocks[i]->begin(), 679 E = NewBlocks[i]->end(); I != E; ++I) 680 RemapInstruction(I, ValueMap); 681 682 // Rewrite the original preheader to select between versions of the loop. 683 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator()); 684 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && 685 "Preheader splitting did not work correctly!"); 686 687 // Emit the new branch that selects between the two versions of this loop. 688 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); 689 OldBR->eraseFromParent(); 690 691 LoopProcessWorklist.push_back(L); 692 LoopProcessWorklist.push_back(NewLoop); 693 694 // Now we rewrite the original code to know that the condition is true and the 695 // new code to know that the condition is false. 696 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false); 697 698 // It's possible that simplifying one loop could cause the other to be 699 // deleted. If so, don't simplify it. 700 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop) 701 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true); 702} 703 704/// RemoveFromWorklist - Remove all instances of I from the worklist vector 705/// specified. 706static void RemoveFromWorklist(Instruction *I, 707 std::vector<Instruction*> &Worklist) { 708 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(), 709 Worklist.end(), I); 710 while (WI != Worklist.end()) { 711 unsigned Offset = WI-Worklist.begin(); 712 Worklist.erase(WI); 713 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); 714 } 715} 716 717/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the 718/// program, replacing all uses with V and update the worklist. 719static void ReplaceUsesOfWith(Instruction *I, Value *V, 720 std::vector<Instruction*> &Worklist) { 721 DOUT << "Replace with '" << *V << "': " << *I; 722 723 // Add uses to the worklist, which may be dead now. 724 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 725 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 726 Worklist.push_back(Use); 727 728 // Add users to the worklist which may be simplified now. 729 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 730 UI != E; ++UI) 731 Worklist.push_back(cast<Instruction>(*UI)); 732 I->replaceAllUsesWith(V); 733 I->eraseFromParent(); 734 RemoveFromWorklist(I, Worklist); 735 ++NumSimplify; 736} 737 738/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop 739/// information, and remove any dead successors it has. 740/// 741void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, 742 std::vector<Instruction*> &Worklist) { 743 if (pred_begin(BB) != pred_end(BB)) { 744 // This block isn't dead, since an edge to BB was just removed, see if there 745 // are any easy simplifications we can do now. 746 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 747 // If it has one pred, fold phi nodes in BB. 748 while (isa<PHINode>(BB->begin())) 749 ReplaceUsesOfWith(BB->begin(), 750 cast<PHINode>(BB->begin())->getIncomingValue(0), 751 Worklist); 752 753 // If this is the header of a loop and the only pred is the latch, we now 754 // have an unreachable loop. 755 if (Loop *L = LI->getLoopFor(BB)) 756 if (L->getHeader() == BB && L->contains(Pred)) { 757 // Remove the branch from the latch to the header block, this makes 758 // the header dead, which will make the latch dead (because the header 759 // dominates the latch). 760 Pred->getTerminator()->eraseFromParent(); 761 new UnreachableInst(Pred); 762 763 // The loop is now broken, remove it from LI. 764 RemoveLoopFromHierarchy(L); 765 766 // Reprocess the header, which now IS dead. 767 RemoveBlockIfDead(BB, Worklist); 768 return; 769 } 770 771 // If pred ends in a uncond branch, add uncond branch to worklist so that 772 // the two blocks will get merged. 773 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 774 if (BI->isUnconditional()) 775 Worklist.push_back(BI); 776 } 777 return; 778 } 779 780 DOUT << "Nuking dead block: " << *BB; 781 782 // Remove the instructions in the basic block from the worklist. 783 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 784 RemoveFromWorklist(I, Worklist); 785 786 // Anything that uses the instructions in this basic block should have their 787 // uses replaced with undefs. 788 if (!I->use_empty()) 789 I->replaceAllUsesWith(UndefValue::get(I->getType())); 790 } 791 792 // If this is the edge to the header block for a loop, remove the loop and 793 // promote all subloops. 794 if (Loop *BBLoop = LI->getLoopFor(BB)) { 795 if (BBLoop->getLoopLatch() == BB) 796 RemoveLoopFromHierarchy(BBLoop); 797 } 798 799 // Remove the block from the loop info, which removes it from any loops it 800 // was in. 801 LI->removeBlock(BB); 802 803 804 // Remove phi node entries in successors for this block. 805 TerminatorInst *TI = BB->getTerminator(); 806 std::vector<BasicBlock*> Succs; 807 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { 808 Succs.push_back(TI->getSuccessor(i)); 809 TI->getSuccessor(i)->removePredecessor(BB); 810 } 811 812 // Unique the successors, remove anything with multiple uses. 813 std::sort(Succs.begin(), Succs.end()); 814 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); 815 816 // Remove the basic block, including all of the instructions contained in it. 817 BB->eraseFromParent(); 818 819 // Remove successor blocks here that are not dead, so that we know we only 820 // have dead blocks in this list. Nondead blocks have a way of becoming dead, 821 // then getting removed before we revisit them, which is badness. 822 // 823 for (unsigned i = 0; i != Succs.size(); ++i) 824 if (pred_begin(Succs[i]) != pred_end(Succs[i])) { 825 // One exception is loop headers. If this block was the preheader for a 826 // loop, then we DO want to visit the loop so the loop gets deleted. 827 // We know that if the successor is a loop header, that this loop had to 828 // be the preheader: the case where this was the latch block was handled 829 // above and headers can only have two predecessors. 830 if (!LI->isLoopHeader(Succs[i])) { 831 Succs.erase(Succs.begin()+i); 832 --i; 833 } 834 } 835 836 for (unsigned i = 0, e = Succs.size(); i != e; ++i) 837 RemoveBlockIfDead(Succs[i], Worklist); 838} 839 840/// RemoveLoopFromHierarchy - We have discovered that the specified loop has 841/// become unwrapped, either because the backedge was deleted, or because the 842/// edge into the header was removed. If the edge into the header from the 843/// latch block was removed, the loop is unwrapped but subloops are still alive, 844/// so they just reparent loops. If the loops are actually dead, they will be 845/// removed later. 846void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { 847 if (Loop *ParentLoop = L->getParentLoop()) { // Not a top-level loop. 848 // Reparent all of the blocks in this loop. Since BBLoop had a parent, 849 // they are now all in it. 850 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 851 I != E; ++I) 852 if (LI->getLoopFor(*I) == L) // Don't change blocks in subloops. 853 LI->changeLoopFor(*I, ParentLoop); 854 855 // Remove the loop from its parent loop. 856 for (Loop::iterator I = ParentLoop->begin(), E = ParentLoop->end();; 857 ++I) { 858 assert(I != E && "Couldn't find loop"); 859 if (*I == L) { 860 ParentLoop->removeChildLoop(I); 861 break; 862 } 863 } 864 865 // Move all subloops into the parent loop. 866 while (L->begin() != L->end()) 867 ParentLoop->addChildLoop(L->removeChildLoop(L->end()-1)); 868 } else { 869 // Reparent all of the blocks in this loop. Since BBLoop had no parent, 870 // they no longer in a loop at all. 871 872 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 873 // Don't change blocks in subloops. 874 if (LI->getLoopFor(L->getBlocks()[i]) == L) { 875 LI->removeBlock(L->getBlocks()[i]); 876 --i; 877 } 878 } 879 880 // Remove the loop from the top-level LoopInfo object. 881 for (LoopInfo::iterator I = LI->begin(), E = LI->end();; ++I) { 882 assert(I != E && "Couldn't find loop"); 883 if (*I == L) { 884 LI->removeLoop(I); 885 break; 886 } 887 } 888 889 // Move all of the subloops to the top-level. 890 while (L->begin() != L->end()) 891 LI->addTopLevelLoop(L->removeChildLoop(L->end()-1)); 892 } 893 894 delete L; 895 RemoveLoopFromWorklist(L); 896} 897 898 899 900// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has 901// the value specified by Val in the specified loop, or we know it does NOT have 902// that value. Rewrite any uses of LIC or of properties correlated to it. 903void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, 904 Constant *Val, 905 bool IsEqual) { 906 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?"); 907 908 // FIXME: Support correlated properties, like: 909 // for (...) 910 // if (li1 < li2) 911 // ... 912 // if (li1 > li2) 913 // ... 914 915 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, 916 // selects, switches. 917 std::vector<User*> Users(LIC->use_begin(), LIC->use_end()); 918 std::vector<Instruction*> Worklist; 919 920 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC 921 // in the loop with the appropriate one directly. 922 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) { 923 Value *Replacement; 924 if (IsEqual) 925 Replacement = Val; 926 else 927 Replacement = ConstantInt::get(Type::Int1Ty, 928 !cast<ConstantInt>(Val)->getZExtValue()); 929 930 for (unsigned i = 0, e = Users.size(); i != e; ++i) 931 if (Instruction *U = cast<Instruction>(Users[i])) { 932 if (!L->contains(U->getParent())) 933 continue; 934 U->replaceUsesOfWith(LIC, Replacement); 935 Worklist.push_back(U); 936 } 937 } else { 938 // Otherwise, we don't know the precise value of LIC, but we do know that it 939 // is certainly NOT "Val". As such, simplify any uses in the loop that we 940 // can. This case occurs when we unswitch switch statements. 941 for (unsigned i = 0, e = Users.size(); i != e; ++i) 942 if (Instruction *U = cast<Instruction>(Users[i])) { 943 if (!L->contains(U->getParent())) 944 continue; 945 946 Worklist.push_back(U); 947 948 // If we know that LIC is not Val, use this info to simplify code. 949 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) { 950 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) { 951 if (SI->getCaseValue(i) == Val) { 952 // Found a dead case value. Don't remove PHI nodes in the 953 // successor if they become single-entry, those PHI nodes may 954 // be in the Users list. 955 956 // FIXME: This is a hack. We need to keep the successor around 957 // and hooked up so as to preserve the loop structure, because 958 // trying to update it is complicated. So instead we preserve the 959 // loop structure and put the block on an dead code path. 960 961 BasicBlock* Old = SI->getParent(); 962 BasicBlock* Split = SplitBlock(Old, SI); 963 964 Instruction* OldTerm = Old->getTerminator(); 965 new BranchInst(Split, SI->getSuccessor(i), 966 ConstantInt::getTrue(), OldTerm); 967 968 Old->getTerminator()->eraseFromParent(); 969 970 971 PHINode *PN; 972 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); 973 (PN = dyn_cast<PHINode>(II)); ++II) { 974 Value *InVal = PN->removeIncomingValue(Split, false); 975 PN->addIncoming(InVal, Old); 976 } 977 978 SI->removeCase(i); 979 break; 980 } 981 } 982 } 983 984 // TODO: We could do other simplifications, for example, turning 985 // LIC == Val -> false. 986 } 987 } 988 989 SimplifyCode(Worklist); 990} 991 992/// SimplifyCode - Okay, now that we have simplified some instructions in the 993/// loop, walk over it and constant prop, dce, and fold control flow where 994/// possible. Note that this is effectively a very simple loop-structure-aware 995/// optimizer. During processing of this loop, L could very well be deleted, so 996/// it must not be used. 997/// 998/// FIXME: When the loop optimizer is more mature, separate this out to a new 999/// pass. 1000/// 1001void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) { 1002 while (!Worklist.empty()) { 1003 Instruction *I = Worklist.back(); 1004 Worklist.pop_back(); 1005 1006 // Simple constant folding. 1007 if (Constant *C = ConstantFoldInstruction(I)) { 1008 ReplaceUsesOfWith(I, C, Worklist); 1009 continue; 1010 } 1011 1012 // Simple DCE. 1013 if (isInstructionTriviallyDead(I)) { 1014 DOUT << "Remove dead instruction '" << *I; 1015 1016 // Add uses to the worklist, which may be dead now. 1017 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 1018 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i))) 1019 Worklist.push_back(Use); 1020 I->eraseFromParent(); 1021 RemoveFromWorklist(I, Worklist); 1022 ++NumSimplify; 1023 continue; 1024 } 1025 1026 // Special case hacks that appear commonly in unswitched code. 1027 switch (I->getOpcode()) { 1028 case Instruction::Select: 1029 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) { 1030 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist); 1031 continue; 1032 } 1033 break; 1034 case Instruction::And: 1035 if (isa<ConstantInt>(I->getOperand(0)) && 1036 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1037 cast<BinaryOperator>(I)->swapOperands(); 1038 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1039 if (CB->getType() == Type::Int1Ty) { 1040 if (CB->getZExtValue()) // X & 1 -> X 1041 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1042 else // X & 0 -> 0 1043 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1044 continue; 1045 } 1046 break; 1047 case Instruction::Or: 1048 if (isa<ConstantInt>(I->getOperand(0)) && 1049 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS 1050 cast<BinaryOperator>(I)->swapOperands(); 1051 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1))) 1052 if (CB->getType() == Type::Int1Ty) { 1053 if (CB->getZExtValue()) // X | 1 -> 1 1054 ReplaceUsesOfWith(I, I->getOperand(1), Worklist); 1055 else // X | 0 -> X 1056 ReplaceUsesOfWith(I, I->getOperand(0), Worklist); 1057 continue; 1058 } 1059 break; 1060 case Instruction::Br: { 1061 BranchInst *BI = cast<BranchInst>(I); 1062 if (BI->isUnconditional()) { 1063 // If BI's parent is the only pred of the successor, fold the two blocks 1064 // together. 1065 BasicBlock *Pred = BI->getParent(); 1066 BasicBlock *Succ = BI->getSuccessor(0); 1067 BasicBlock *SinglePred = Succ->getSinglePredecessor(); 1068 if (!SinglePred) continue; // Nothing to do. 1069 assert(SinglePred == Pred && "CFG broken"); 1070 1071 DOUT << "Merging blocks: " << Pred->getName() << " <- " 1072 << Succ->getName() << "\n"; 1073 1074 // Resolve any single entry PHI nodes in Succ. 1075 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin())) 1076 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); 1077 1078 // Move all of the successor contents from Succ to Pred. 1079 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), 1080 Succ->end()); 1081 BI->eraseFromParent(); 1082 RemoveFromWorklist(BI, Worklist); 1083 1084 // If Succ has any successors with PHI nodes, update them to have 1085 // entries coming from Pred instead of Succ. 1086 Succ->replaceAllUsesWith(Pred); 1087 1088 // Remove Succ from the loop tree. 1089 LI->removeBlock(Succ); 1090 Succ->eraseFromParent(); 1091 ++NumSimplify; 1092 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){ 1093 // Conditional branch. Turn it into an unconditional branch, then 1094 // remove dead blocks. 1095 break; // FIXME: Enable. 1096 1097 DOUT << "Folded branch: " << *BI; 1098 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); 1099 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); 1100 DeadSucc->removePredecessor(BI->getParent(), true); 1101 Worklist.push_back(new BranchInst(LiveSucc, BI)); 1102 BI->eraseFromParent(); 1103 RemoveFromWorklist(BI, Worklist); 1104 ++NumSimplify; 1105 1106 RemoveBlockIfDead(DeadSucc, Worklist); 1107 } 1108 break; 1109 } 1110 } 1111 } 1112} 1113