LICM.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs loop invariant code motion, attempting to remove as much 11// code from the body of a loop as possible. It does this by either hoisting 12// code into the preheader block, or by sinking code to the exit blocks if it is 13// safe. This pass also promotes must-aliased memory locations in the loop to 14// live in registers, thus hoisting and sinking "invariant" loads and stores. 15// 16// This pass uses alias analysis for two purposes: 17// 18// 1. Moving loop invariant loads and calls out of loops. If we can determine 19// that a load or call inside of a loop never aliases anything stored to, 20// we can hoist it or sink it like any other instruction. 21// 2. Scalar Promotion of Memory - If there is a store instruction inside of 22// the loop, we try to move the store to happen AFTER the loop instead of 23// inside of the loop. This can only happen if a few conditions are true: 24// A. The pointer stored through is loop invariant 25// B. There are no stores or loads in the loop which _may_ alias the 26// pointer. There are no calls in the loop which mod/ref the pointer. 27// If these conditions are true, we can promote the loads and stores in the 28// loop of the pointer to use a temporary alloca'd variable. We then use 29// the SSAUpdater to construct the appropriate SSA form for the value. 30// 31//===----------------------------------------------------------------------===// 32 33#include "llvm/Transforms/Scalar.h" 34#include "llvm/ADT/Statistic.h" 35#include "llvm/Analysis/AliasAnalysis.h" 36#include "llvm/Analysis/AliasSetTracker.h" 37#include "llvm/Analysis/ConstantFolding.h" 38#include "llvm/Analysis/LoopInfo.h" 39#include "llvm/Analysis/LoopPass.h" 40#include "llvm/Analysis/ScalarEvolution.h" 41#include "llvm/Analysis/ValueTracking.h" 42#include "llvm/IR/CFG.h" 43#include "llvm/IR/Constants.h" 44#include "llvm/IR/DataLayout.h" 45#include "llvm/IR/DerivedTypes.h" 46#include "llvm/IR/Dominators.h" 47#include "llvm/IR/Instructions.h" 48#include "llvm/IR/IntrinsicInst.h" 49#include "llvm/IR/LLVMContext.h" 50#include "llvm/IR/Metadata.h" 51#include "llvm/IR/PredIteratorCache.h" 52#include "llvm/Support/CommandLine.h" 53#include "llvm/Support/Debug.h" 54#include "llvm/Support/raw_ostream.h" 55#include "llvm/Target/TargetLibraryInfo.h" 56#include "llvm/Transforms/Utils/Local.h" 57#include "llvm/Transforms/Utils/LoopUtils.h" 58#include "llvm/Transforms/Utils/SSAUpdater.h" 59#include <algorithm> 60using namespace llvm; 61 62#define DEBUG_TYPE "licm" 63 64STATISTIC(NumSunk , "Number of instructions sunk out of loop"); 65STATISTIC(NumHoisted , "Number of instructions hoisted out of loop"); 66STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk"); 67STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk"); 68STATISTIC(NumPromoted , "Number of memory locations promoted to registers"); 69 70static cl::opt<bool> 71DisablePromotion("disable-licm-promotion", cl::Hidden, 72 cl::desc("Disable memory promotion in LICM pass")); 73 74namespace { 75 struct LICM : public LoopPass { 76 static char ID; // Pass identification, replacement for typeid 77 LICM() : LoopPass(ID) { 78 initializeLICMPass(*PassRegistry::getPassRegistry()); 79 } 80 81 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 82 83 /// This transformation requires natural loop information & requires that 84 /// loop preheaders be inserted into the CFG... 85 /// 86 void getAnalysisUsage(AnalysisUsage &AU) const override { 87 AU.setPreservesCFG(); 88 AU.addRequired<DominatorTreeWrapperPass>(); 89 AU.addRequired<LoopInfo>(); 90 AU.addRequiredID(LoopSimplifyID); 91 AU.addPreservedID(LoopSimplifyID); 92 AU.addRequiredID(LCSSAID); 93 AU.addPreservedID(LCSSAID); 94 AU.addRequired<AliasAnalysis>(); 95 AU.addPreserved<AliasAnalysis>(); 96 AU.addPreserved<ScalarEvolution>(); 97 AU.addRequired<TargetLibraryInfo>(); 98 } 99 100 using llvm::Pass::doFinalization; 101 102 bool doFinalization() override { 103 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets"); 104 return false; 105 } 106 107 private: 108 AliasAnalysis *AA; // Current AliasAnalysis information 109 LoopInfo *LI; // Current LoopInfo 110 DominatorTree *DT; // Dominator Tree for the current Loop. 111 112 const DataLayout *DL; // DataLayout for constant folding. 113 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding. 114 115 // State that is updated as we process loops. 116 bool Changed; // Set to true when we change anything. 117 BasicBlock *Preheader; // The preheader block of the current loop... 118 Loop *CurLoop; // The current loop we are working on... 119 AliasSetTracker *CurAST; // AliasSet information for the current loop... 120 bool MayThrow; // The current loop contains an instruction which 121 // may throw, thus preventing code motion of 122 // instructions with side effects. 123 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap; 124 125 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 126 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, 127 Loop *L) override; 128 129 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 130 /// set. 131 void deleteAnalysisValue(Value *V, Loop *L) override; 132 133 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks 134 /// dominated by the specified block, and that are in the current loop) in 135 /// reverse depth first order w.r.t the DominatorTree. This allows us to 136 /// visit uses before definitions, allowing us to sink a loop body in one 137 /// pass without iteration. 138 /// 139 void SinkRegion(DomTreeNode *N); 140 141 /// HoistRegion - Walk the specified region of the CFG (defined by all 142 /// blocks dominated by the specified block, and that are in the current 143 /// loop) in depth first order w.r.t the DominatorTree. This allows us to 144 /// visit definitions before uses, allowing us to hoist a loop body in one 145 /// pass without iteration. 146 /// 147 void HoistRegion(DomTreeNode *N); 148 149 /// inSubLoop - Little predicate that returns true if the specified basic 150 /// block is in a subloop of the current one, not the current one itself. 151 /// 152 bool inSubLoop(BasicBlock *BB) { 153 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop"); 154 return LI->getLoopFor(BB) != CurLoop; 155 } 156 157 /// sink - When an instruction is found to only be used outside of the loop, 158 /// this function moves it to the exit blocks and patches up SSA form as 159 /// needed. 160 /// 161 void sink(Instruction &I); 162 163 /// hoist - When an instruction is found to only use loop invariant operands 164 /// that is safe to hoist, this instruction is called to do the dirty work. 165 /// 166 void hoist(Instruction &I); 167 168 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it 169 /// is not a trapping instruction or if it is a trapping instruction and is 170 /// guaranteed to execute. 171 /// 172 bool isSafeToExecuteUnconditionally(Instruction &I); 173 174 /// isGuaranteedToExecute - Check that the instruction is guaranteed to 175 /// execute. 176 /// 177 bool isGuaranteedToExecute(Instruction &I); 178 179 /// pointerInvalidatedByLoop - Return true if the body of this loop may 180 /// store into the memory location pointed to by V. 181 /// 182 bool pointerInvalidatedByLoop(Value *V, uint64_t Size, 183 const MDNode *TBAAInfo) { 184 // Check to see if any of the basic blocks in CurLoop invalidate *V. 185 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod(); 186 } 187 188 bool canSinkOrHoistInst(Instruction &I); 189 bool isNotUsedInLoop(Instruction &I); 190 191 void PromoteAliasSet(AliasSet &AS, 192 SmallVectorImpl<BasicBlock*> &ExitBlocks, 193 SmallVectorImpl<Instruction*> &InsertPts, 194 PredIteratorCache &PIC); 195 }; 196} 197 198char LICM::ID = 0; 199INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false) 200INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 201INITIALIZE_PASS_DEPENDENCY(LoopInfo) 202INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 203INITIALIZE_PASS_DEPENDENCY(LCSSA) 204INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 205INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 206INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 207INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false) 208 209Pass *llvm::createLICMPass() { return new LICM(); } 210 211/// Hoist expressions out of the specified loop. Note, alias info for inner 212/// loop is not preserved so it is not a good idea to run LICM multiple 213/// times on one loop. 214/// 215bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) { 216 if (skipOptnoneFunction(L)) 217 return false; 218 219 Changed = false; 220 221 // Get our Loop and Alias Analysis information... 222 LI = &getAnalysis<LoopInfo>(); 223 AA = &getAnalysis<AliasAnalysis>(); 224 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 225 226 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 227 DL = DLP ? &DLP->getDataLayout() : nullptr; 228 TLI = &getAnalysis<TargetLibraryInfo>(); 229 230 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form."); 231 232 CurAST = new AliasSetTracker(*AA); 233 // Collect Alias info from subloops. 234 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end(); 235 LoopItr != LoopItrE; ++LoopItr) { 236 Loop *InnerL = *LoopItr; 237 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL]; 238 assert(InnerAST && "Where is my AST?"); 239 240 // What if InnerLoop was modified by other passes ? 241 CurAST->add(*InnerAST); 242 243 // Once we've incorporated the inner loop's AST into ours, we don't need the 244 // subloop's anymore. 245 delete InnerAST; 246 LoopToAliasSetMap.erase(InnerL); 247 } 248 249 CurLoop = L; 250 251 // Get the preheader block to move instructions into... 252 Preheader = L->getLoopPreheader(); 253 254 // Loop over the body of this loop, looking for calls, invokes, and stores. 255 // Because subloops have already been incorporated into AST, we skip blocks in 256 // subloops. 257 // 258 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 259 I != E; ++I) { 260 BasicBlock *BB = *I; 261 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops. 262 CurAST->add(*BB); // Incorporate the specified basic block 263 } 264 265 MayThrow = false; 266 // TODO: We've already searched for instructions which may throw in subloops. 267 // We may want to reuse this information. 268 for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end(); 269 (BB != BBE) && !MayThrow ; ++BB) 270 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); 271 (I != E) && !MayThrow; ++I) 272 MayThrow |= I->mayThrow(); 273 274 // We want to visit all of the instructions in this loop... that are not parts 275 // of our subloops (they have already had their invariants hoisted out of 276 // their loop, into this loop, so there is no need to process the BODIES of 277 // the subloops). 278 // 279 // Traverse the body of the loop in depth first order on the dominator tree so 280 // that we are guaranteed to see definitions before we see uses. This allows 281 // us to sink instructions in one pass, without iteration. After sinking 282 // instructions, we perform another pass to hoist them out of the loop. 283 // 284 if (L->hasDedicatedExits()) 285 SinkRegion(DT->getNode(L->getHeader())); 286 if (Preheader) 287 HoistRegion(DT->getNode(L->getHeader())); 288 289 // Now that all loop invariants have been removed from the loop, promote any 290 // memory references to scalars that we can. 291 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) { 292 SmallVector<BasicBlock *, 8> ExitBlocks; 293 SmallVector<Instruction *, 8> InsertPts; 294 PredIteratorCache PIC; 295 296 // Loop over all of the alias sets in the tracker object. 297 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 298 I != E; ++I) 299 PromoteAliasSet(*I, ExitBlocks, InsertPts, PIC); 300 301 // Once we have promoted values across the loop body we have to recursively 302 // reform LCSSA as any nested loop may now have values defined within the 303 // loop used in the outer loop. 304 // FIXME: This is really heavy handed. It would be a bit better to use an 305 // SSAUpdater strategy during promotion that was LCSSA aware and reformed 306 // it as it went. 307 if (Changed) 308 formLCSSARecursively(*L, *DT, getAnalysisIfAvailable<ScalarEvolution>()); 309 } 310 311 // Check that neither this loop nor its parent have had LCSSA broken. LICM is 312 // specifically moving instructions across the loop boundary and so it is 313 // especially in need of sanity checking here. 314 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!"); 315 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && 316 "Parent loop not left in LCSSA form after LICM!"); 317 318 // Clear out loops state information for the next iteration 319 CurLoop = nullptr; 320 Preheader = nullptr; 321 322 // If this loop is nested inside of another one, save the alias information 323 // for when we process the outer loop. 324 if (L->getParentLoop()) 325 LoopToAliasSetMap[L] = CurAST; 326 else 327 delete CurAST; 328 return Changed; 329} 330 331/// SinkRegion - Walk the specified region of the CFG (defined by all blocks 332/// dominated by the specified block, and that are in the current loop) in 333/// reverse depth first order w.r.t the DominatorTree. This allows us to visit 334/// uses before definitions, allowing us to sink a loop body in one pass without 335/// iteration. 336/// 337void LICM::SinkRegion(DomTreeNode *N) { 338 assert(N != nullptr && "Null dominator tree node?"); 339 BasicBlock *BB = N->getBlock(); 340 341 // If this subregion is not in the top level loop at all, exit. 342 if (!CurLoop->contains(BB)) return; 343 344 // We are processing blocks in reverse dfo, so process children first. 345 const std::vector<DomTreeNode*> &Children = N->getChildren(); 346 for (unsigned i = 0, e = Children.size(); i != e; ++i) 347 SinkRegion(Children[i]); 348 349 // Only need to process the contents of this block if it is not part of a 350 // subloop (which would already have been processed). 351 if (inSubLoop(BB)) return; 352 353 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) { 354 Instruction &I = *--II; 355 356 // If the instruction is dead, we would try to sink it because it isn't used 357 // in the loop, instead, just delete it. 358 if (isInstructionTriviallyDead(&I, TLI)) { 359 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n'); 360 ++II; 361 CurAST->deleteValue(&I); 362 I.eraseFromParent(); 363 Changed = true; 364 continue; 365 } 366 367 // Check to see if we can sink this instruction to the exit blocks 368 // of the loop. We can do this if the all users of the instruction are 369 // outside of the loop. In this case, it doesn't even matter if the 370 // operands of the instruction are loop invariant. 371 // 372 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) { 373 ++II; 374 sink(I); 375 } 376 } 377} 378 379/// HoistRegion - Walk the specified region of the CFG (defined by all blocks 380/// dominated by the specified block, and that are in the current loop) in depth 381/// first order w.r.t the DominatorTree. This allows us to visit definitions 382/// before uses, allowing us to hoist a loop body in one pass without iteration. 383/// 384void LICM::HoistRegion(DomTreeNode *N) { 385 assert(N != nullptr && "Null dominator tree node?"); 386 BasicBlock *BB = N->getBlock(); 387 388 // If this subregion is not in the top level loop at all, exit. 389 if (!CurLoop->contains(BB)) return; 390 391 // Only need to process the contents of this block if it is not part of a 392 // subloop (which would already have been processed). 393 if (!inSubLoop(BB)) 394 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) { 395 Instruction &I = *II++; 396 397 // Try constant folding this instruction. If all the operands are 398 // constants, it is technically hoistable, but it would be better to just 399 // fold it. 400 if (Constant *C = ConstantFoldInstruction(&I, DL, TLI)) { 401 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n'); 402 CurAST->copyValue(&I, C); 403 CurAST->deleteValue(&I); 404 I.replaceAllUsesWith(C); 405 I.eraseFromParent(); 406 continue; 407 } 408 409 // Try hoisting the instruction out to the preheader. We can only do this 410 // if all of the operands of the instruction are loop invariant and if it 411 // is safe to hoist the instruction. 412 // 413 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) && 414 isSafeToExecuteUnconditionally(I)) 415 hoist(I); 416 } 417 418 const std::vector<DomTreeNode*> &Children = N->getChildren(); 419 for (unsigned i = 0, e = Children.size(); i != e; ++i) 420 HoistRegion(Children[i]); 421} 422 423/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this 424/// instruction. 425/// 426bool LICM::canSinkOrHoistInst(Instruction &I) { 427 // Loads have extra constraints we have to verify before we can hoist them. 428 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) { 429 if (!LI->isUnordered()) 430 return false; // Don't hoist volatile/atomic loads! 431 432 // Loads from constant memory are always safe to move, even if they end up 433 // in the same alias set as something that ends up being modified. 434 if (AA->pointsToConstantMemory(LI->getOperand(0))) 435 return true; 436 if (LI->getMetadata("invariant.load")) 437 return true; 438 439 // Don't hoist loads which have may-aliased stores in loop. 440 uint64_t Size = 0; 441 if (LI->getType()->isSized()) 442 Size = AA->getTypeStoreSize(LI->getType()); 443 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, 444 LI->getMetadata(LLVMContext::MD_tbaa)); 445 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) { 446 // Don't sink or hoist dbg info; it's legal, but not useful. 447 if (isa<DbgInfoIntrinsic>(I)) 448 return false; 449 450 // Handle simple cases by querying alias analysis. 451 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI); 452 if (Behavior == AliasAnalysis::DoesNotAccessMemory) 453 return true; 454 if (AliasAnalysis::onlyReadsMemory(Behavior)) { 455 // If this call only reads from memory and there are no writes to memory 456 // in the loop, we can hoist or sink the call as appropriate. 457 bool FoundMod = false; 458 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end(); 459 I != E; ++I) { 460 AliasSet &AS = *I; 461 if (!AS.isForwardingAliasSet() && AS.isMod()) { 462 FoundMod = true; 463 break; 464 } 465 } 466 if (!FoundMod) return true; 467 } 468 469 // FIXME: This should use mod/ref information to see if we can hoist or 470 // sink the call. 471 472 return false; 473 } 474 475 // Only these instructions are hoistable/sinkable. 476 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) && 477 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) && 478 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) && 479 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) && 480 !isa<InsertValueInst>(I)) 481 return false; 482 483 return isSafeToExecuteUnconditionally(I); 484} 485 486/// \brief Returns true if a PHINode is a trivially replaceable with an 487/// Instruction. 488/// 489/// This is true when all incoming values are that instruction. This pattern 490/// occurs most often with LCSSA PHI nodes. 491static bool isTriviallyReplacablePHI(PHINode &PN, Instruction &I) { 492 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) 493 if (PN.getIncomingValue(i) != &I) 494 return false; 495 496 return true; 497} 498 499/// isNotUsedInLoop - Return true if the only users of this instruction are 500/// outside of the loop. If this is true, we can sink the instruction to the 501/// exit blocks of the loop. 502/// 503bool LICM::isNotUsedInLoop(Instruction &I) { 504 for (User *U : I.users()) { 505 Instruction *UI = cast<Instruction>(U); 506 if (PHINode *PN = dyn_cast<PHINode>(UI)) { 507 // A PHI node where all of the incoming values are this instruction are 508 // special -- they can just be RAUW'ed with the instruction and thus 509 // don't require a use in the predecessor. This is a particular important 510 // special case because it is the pattern found in LCSSA form. 511 if (isTriviallyReplacablePHI(*PN, I)) { 512 if (CurLoop->contains(PN)) 513 return false; 514 else 515 continue; 516 } 517 518 // Otherwise, PHI node uses occur in predecessor blocks if the incoming 519 // values. Check for such a use being inside the loop. 520 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 521 if (PN->getIncomingValue(i) == &I) 522 if (CurLoop->contains(PN->getIncomingBlock(i))) 523 return false; 524 525 continue; 526 } 527 528 if (CurLoop->contains(UI)) 529 return false; 530 } 531 return true; 532} 533 534/// sink - When an instruction is found to only be used outside of the loop, 535/// this function moves it to the exit blocks and patches up SSA form as needed. 536/// This method is guaranteed to remove the original instruction from its 537/// position, and may either delete it or move it to outside of the loop. 538/// 539void LICM::sink(Instruction &I) { 540 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n"); 541 542 if (isa<LoadInst>(I)) ++NumMovedLoads; 543 else if (isa<CallInst>(I)) ++NumMovedCalls; 544 ++NumSunk; 545 Changed = true; 546 547#ifndef NDEBUG 548 SmallVector<BasicBlock *, 32> ExitBlocks; 549 CurLoop->getUniqueExitBlocks(ExitBlocks); 550 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end()); 551#endif 552 553 // If this instruction is only used outside of the loop, then all users are 554 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of 555 // the instruction. 556 while (!I.use_empty()) { 557 // The user must be a PHI node. 558 PHINode *PN = cast<PHINode>(I.user_back()); 559 560 BasicBlock *ExitBlock = PN->getParent(); 561 assert(ExitBlockSet.count(ExitBlock) && 562 "The LCSSA PHI is not in an exit block!"); 563 564 Instruction *New = I.clone(); 565 ExitBlock->getInstList().insert(ExitBlock->getFirstInsertionPt(), New); 566 if (!I.getName().empty()) 567 New->setName(I.getName() + ".le"); 568 569 // Build LCSSA PHI nodes for any in-loop operands. Note that this is 570 // particularly cheap because we can rip off the PHI node that we're 571 // replacing for the number and blocks of the predecessors. 572 // OPT: If this shows up in a profile, we can instead finish sinking all 573 // invariant instructions, and then walk their operands to re-establish 574 // LCSSA. That will eliminate creating PHI nodes just to nuke them when 575 // sinking bottom-up. 576 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE; 577 ++OI) 578 if (Instruction *OInst = dyn_cast<Instruction>(*OI)) 579 if (Loop *OLoop = LI->getLoopFor(OInst->getParent())) 580 if (!OLoop->contains(PN)) { 581 PHINode *OpPN = PHINode::Create( 582 OInst->getType(), PN->getNumIncomingValues(), 583 OInst->getName() + ".lcssa", ExitBlock->begin()); 584 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 585 OpPN->addIncoming(OInst, PN->getIncomingBlock(i)); 586 *OI = OpPN; 587 } 588 589 PN->replaceAllUsesWith(New); 590 PN->eraseFromParent(); 591 } 592 593 CurAST->deleteValue(&I); 594 I.eraseFromParent(); 595} 596 597/// hoist - When an instruction is found to only use loop invariant operands 598/// that is safe to hoist, this instruction is called to do the dirty work. 599/// 600void LICM::hoist(Instruction &I) { 601 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " 602 << I << "\n"); 603 604 // Move the new node to the Preheader, before its terminator. 605 I.moveBefore(Preheader->getTerminator()); 606 607 if (isa<LoadInst>(I)) ++NumMovedLoads; 608 else if (isa<CallInst>(I)) ++NumMovedCalls; 609 ++NumHoisted; 610 Changed = true; 611} 612 613/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is 614/// not a trapping instruction or if it is a trapping instruction and is 615/// guaranteed to execute. 616/// 617bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) { 618 // If it is not a trapping instruction, it is always safe to hoist. 619 if (isSafeToSpeculativelyExecute(&Inst)) 620 return true; 621 622 return isGuaranteedToExecute(Inst); 623} 624 625bool LICM::isGuaranteedToExecute(Instruction &Inst) { 626 627 // Somewhere in this loop there is an instruction which may throw and make us 628 // exit the loop. 629 if (MayThrow) 630 return false; 631 632 // Otherwise we have to check to make sure that the instruction dominates all 633 // of the exit blocks. If it doesn't, then there is a path out of the loop 634 // which does not execute this instruction, so we can't hoist it. 635 636 // If the instruction is in the header block for the loop (which is very 637 // common), it is always guaranteed to dominate the exit blocks. Since this 638 // is a common case, and can save some work, check it now. 639 if (Inst.getParent() == CurLoop->getHeader()) 640 return true; 641 642 // Get the exit blocks for the current loop. 643 SmallVector<BasicBlock*, 8> ExitBlocks; 644 CurLoop->getExitBlocks(ExitBlocks); 645 646 // Verify that the block dominates each of the exit blocks of the loop. 647 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 648 if (!DT->dominates(Inst.getParent(), ExitBlocks[i])) 649 return false; 650 651 // As a degenerate case, if the loop is statically infinite then we haven't 652 // proven anything since there are no exit blocks. 653 if (ExitBlocks.empty()) 654 return false; 655 656 return true; 657} 658 659namespace { 660 class LoopPromoter : public LoadAndStorePromoter { 661 Value *SomePtr; // Designated pointer to store to. 662 SmallPtrSet<Value*, 4> &PointerMustAliases; 663 SmallVectorImpl<BasicBlock*> &LoopExitBlocks; 664 SmallVectorImpl<Instruction*> &LoopInsertPts; 665 PredIteratorCache &PredCache; 666 AliasSetTracker &AST; 667 LoopInfo &LI; 668 DebugLoc DL; 669 int Alignment; 670 MDNode *TBAATag; 671 672 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const { 673 if (Instruction *I = dyn_cast<Instruction>(V)) 674 if (Loop *L = LI.getLoopFor(I->getParent())) 675 if (!L->contains(BB)) { 676 // We need to create an LCSSA PHI node for the incoming value and 677 // store that. 678 PHINode *PN = PHINode::Create( 679 I->getType(), PredCache.GetNumPreds(BB), 680 I->getName() + ".lcssa", BB->begin()); 681 for (BasicBlock **PI = PredCache.GetPreds(BB); *PI; ++PI) 682 PN->addIncoming(I, *PI); 683 return PN; 684 } 685 return V; 686 } 687 688 public: 689 LoopPromoter(Value *SP, const SmallVectorImpl<Instruction *> &Insts, 690 SSAUpdater &S, SmallPtrSet<Value *, 4> &PMA, 691 SmallVectorImpl<BasicBlock *> &LEB, 692 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC, 693 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment, 694 MDNode *TBAATag) 695 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA), 696 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast), 697 LI(li), DL(dl), Alignment(alignment), TBAATag(TBAATag) {} 698 699 bool isInstInList(Instruction *I, 700 const SmallVectorImpl<Instruction*> &) const override { 701 Value *Ptr; 702 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 703 Ptr = LI->getOperand(0); 704 else 705 Ptr = cast<StoreInst>(I)->getPointerOperand(); 706 return PointerMustAliases.count(Ptr); 707 } 708 709 void doExtraRewritesBeforeFinalDeletion() const override { 710 // Insert stores after in the loop exit blocks. Each exit block gets a 711 // store of the live-out values that feed them. Since we've already told 712 // the SSA updater about the defs in the loop and the preheader 713 // definition, it is all set and we can start using it. 714 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) { 715 BasicBlock *ExitBlock = LoopExitBlocks[i]; 716 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock); 717 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock); 718 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock); 719 Instruction *InsertPos = LoopInsertPts[i]; 720 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos); 721 NewSI->setAlignment(Alignment); 722 NewSI->setDebugLoc(DL); 723 if (TBAATag) NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag); 724 } 725 } 726 727 void replaceLoadWithValue(LoadInst *LI, Value *V) const override { 728 // Update alias analysis. 729 AST.copyValue(LI, V); 730 } 731 void instructionDeleted(Instruction *I) const override { 732 AST.deleteValue(I); 733 } 734 }; 735} // end anon namespace 736 737/// PromoteAliasSet - Try to promote memory values to scalars by sinking 738/// stores out of the loop and moving loads to before the loop. We do this by 739/// looping over the stores in the loop, looking for stores to Must pointers 740/// which are loop invariant. 741/// 742void LICM::PromoteAliasSet(AliasSet &AS, 743 SmallVectorImpl<BasicBlock*> &ExitBlocks, 744 SmallVectorImpl<Instruction*> &InsertPts, 745 PredIteratorCache &PIC) { 746 // We can promote this alias set if it has a store, if it is a "Must" alias 747 // set, if the pointer is loop invariant, and if we are not eliminating any 748 // volatile loads or stores. 749 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() || 750 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue())) 751 return; 752 753 assert(!AS.empty() && 754 "Must alias set should have at least one pointer element in it!"); 755 Value *SomePtr = AS.begin()->getValue(); 756 757 // It isn't safe to promote a load/store from the loop if the load/store is 758 // conditional. For example, turning: 759 // 760 // for () { if (c) *P += 1; } 761 // 762 // into: 763 // 764 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp; 765 // 766 // is not safe, because *P may only be valid to access if 'c' is true. 767 // 768 // It is safe to promote P if all uses are direct load/stores and if at 769 // least one is guaranteed to be executed. 770 bool GuaranteedToExecute = false; 771 772 SmallVector<Instruction*, 64> LoopUses; 773 SmallPtrSet<Value*, 4> PointerMustAliases; 774 775 // We start with an alignment of one and try to find instructions that allow 776 // us to prove better alignment. 777 unsigned Alignment = 1; 778 MDNode *TBAATag = nullptr; 779 780 // Check that all of the pointers in the alias set have the same type. We 781 // cannot (yet) promote a memory location that is loaded and stored in 782 // different sizes. While we are at it, collect alignment and TBAA info. 783 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) { 784 Value *ASIV = ASI->getValue(); 785 PointerMustAliases.insert(ASIV); 786 787 // Check that all of the pointers in the alias set have the same type. We 788 // cannot (yet) promote a memory location that is loaded and stored in 789 // different sizes. 790 if (SomePtr->getType() != ASIV->getType()) 791 return; 792 793 for (User *U : ASIV->users()) { 794 // Ignore instructions that are outside the loop. 795 Instruction *UI = dyn_cast<Instruction>(U); 796 if (!UI || !CurLoop->contains(UI)) 797 continue; 798 799 // If there is an non-load/store instruction in the loop, we can't promote 800 // it. 801 if (LoadInst *load = dyn_cast<LoadInst>(UI)) { 802 assert(!load->isVolatile() && "AST broken"); 803 if (!load->isSimple()) 804 return; 805 } else if (StoreInst *store = dyn_cast<StoreInst>(UI)) { 806 // Stores *of* the pointer are not interesting, only stores *to* the 807 // pointer. 808 if (UI->getOperand(1) != ASIV) 809 continue; 810 assert(!store->isVolatile() && "AST broken"); 811 if (!store->isSimple()) 812 return; 813 814 // Note that we only check GuaranteedToExecute inside the store case 815 // so that we do not introduce stores where they did not exist before 816 // (which would break the LLVM concurrency model). 817 818 // If the alignment of this instruction allows us to specify a more 819 // restrictive (and performant) alignment and if we are sure this 820 // instruction will be executed, update the alignment. 821 // Larger is better, with the exception of 0 being the best alignment. 822 unsigned InstAlignment = store->getAlignment(); 823 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0) 824 if (isGuaranteedToExecute(*UI)) { 825 GuaranteedToExecute = true; 826 Alignment = InstAlignment; 827 } 828 829 if (!GuaranteedToExecute) 830 GuaranteedToExecute = isGuaranteedToExecute(*UI); 831 832 } else 833 return; // Not a load or store. 834 835 // Merge the TBAA tags. 836 if (LoopUses.empty()) { 837 // On the first load/store, just take its TBAA tag. 838 TBAATag = UI->getMetadata(LLVMContext::MD_tbaa); 839 } else if (TBAATag) { 840 TBAATag = MDNode::getMostGenericTBAA(TBAATag, 841 UI->getMetadata(LLVMContext::MD_tbaa)); 842 } 843 844 LoopUses.push_back(UI); 845 } 846 } 847 848 // If there isn't a guaranteed-to-execute instruction, we can't promote. 849 if (!GuaranteedToExecute) 850 return; 851 852 // Otherwise, this is safe to promote, lets do it! 853 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n'); 854 Changed = true; 855 ++NumPromoted; 856 857 // Grab a debug location for the inserted loads/stores; given that the 858 // inserted loads/stores have little relation to the original loads/stores, 859 // this code just arbitrarily picks a location from one, since any debug 860 // location is better than none. 861 DebugLoc DL = LoopUses[0]->getDebugLoc(); 862 863 // Figure out the loop exits and their insertion points, if this is the 864 // first promotion. 865 if (ExitBlocks.empty()) { 866 CurLoop->getUniqueExitBlocks(ExitBlocks); 867 InsertPts.resize(ExitBlocks.size()); 868 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) 869 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt(); 870 } 871 872 // We use the SSAUpdater interface to insert phi nodes as required. 873 SmallVector<PHINode*, 16> NewPHIs; 874 SSAUpdater SSA(&NewPHIs); 875 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks, 876 InsertPts, PIC, *CurAST, *LI, DL, Alignment, TBAATag); 877 878 // Set up the preheader to have a definition of the value. It is the live-out 879 // value from the preheader that uses in the loop will use. 880 LoadInst *PreheaderLoad = 881 new LoadInst(SomePtr, SomePtr->getName()+".promoted", 882 Preheader->getTerminator()); 883 PreheaderLoad->setAlignment(Alignment); 884 PreheaderLoad->setDebugLoc(DL); 885 if (TBAATag) PreheaderLoad->setMetadata(LLVMContext::MD_tbaa, TBAATag); 886 SSA.AddAvailableValue(Preheader, PreheaderLoad); 887 888 // Rewrite all the loads in the loop and remember all the definitions from 889 // stores in the loop. 890 Promoter.run(LoopUses); 891 892 // If the SSAUpdater didn't use the load in the preheader, just zap it now. 893 if (PreheaderLoad->use_empty()) 894 PreheaderLoad->eraseFromParent(); 895} 896 897 898/// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info. 899void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) { 900 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 901 if (!AST) 902 return; 903 904 AST->copyValue(From, To); 905} 906 907/// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias 908/// set. 909void LICM::deleteAnalysisValue(Value *V, Loop *L) { 910 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L); 911 if (!AST) 912 return; 913 914 AST->deleteValue(V); 915} 916