1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
16 // This pass uses alias analysis for two purposes:
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.
31 //===----------------------------------------------------------------------===//
33 #include "llvm/Transforms/Scalar/LICM.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/BasicAliasAnalysis.h"
38 #include "llvm/Analysis/CaptureTracking.h"
39 #include "llvm/Analysis/ConstantFolding.h"
40 #include "llvm/Analysis/GlobalsModRef.h"
41 #include "llvm/Analysis/Loads.h"
42 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/Analysis/LoopPass.h"
44 #include "llvm/Analysis/LoopPassManager.h"
45 #include "llvm/Analysis/MemoryBuiltins.h"
46 #include "llvm/Analysis/ScalarEvolution.h"
47 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/ValueTracking.h"
50 #include "llvm/IR/CFG.h"
51 #include "llvm/IR/Constants.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/DerivedTypes.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/LLVMContext.h"
58 #include "llvm/IR/Metadata.h"
59 #include "llvm/IR/PredIteratorCache.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/Local.h"
65 #include "llvm/Transforms/Utils/LoopUtils.h"
66 #include "llvm/Transforms/Utils/SSAUpdater.h"
71 #define DEBUG_TYPE "licm"
73 STATISTIC(NumSunk, "Number of instructions sunk out of loop");
74 STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
75 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
76 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
77 STATISTIC(NumPromoted, "Number of memory locations promoted to registers");
80 DisablePromotion("disable-licm-promotion", cl::Hidden,
81 cl::desc("Disable memory promotion in LICM pass"));
83 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
84 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
85 const LoopSafetyInfo *SafetyInfo);
86 static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
87 const LoopSafetyInfo *SafetyInfo);
88 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
89 const Loop *CurLoop, AliasSetTracker *CurAST,
90 const LoopSafetyInfo *SafetyInfo);
91 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
92 const DominatorTree *DT,
94 const LoopSafetyInfo *SafetyInfo,
95 const Instruction *CtxI = nullptr);
96 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
97 const AAMDNodes &AAInfo,
98 AliasSetTracker *CurAST);
100 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
102 const LoopSafetyInfo *SafetyInfo);
105 struct LoopInvariantCodeMotion {
106 bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
107 TargetLibraryInfo *TLI, ScalarEvolution *SE, bool DeleteAST);
109 DenseMap<Loop *, AliasSetTracker *> &getLoopToAliasSetMap() {
110 return LoopToAliasSetMap;
114 DenseMap<Loop *, AliasSetTracker *> LoopToAliasSetMap;
116 AliasSetTracker *collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
120 struct LegacyLICMPass : public LoopPass {
121 static char ID; // Pass identification, replacement for typeid
122 LegacyLICMPass() : LoopPass(ID) {
123 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
126 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
128 // If we have run LICM on a previous loop but now we are skipping
129 // (because we've hit the opt-bisect limit), we need to clear the
130 // loop alias information.
131 for (auto <AS : LICM.getLoopToAliasSetMap())
133 LICM.getLoopToAliasSetMap().clear();
137 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
138 return LICM.runOnLoop(L,
139 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
140 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
141 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
142 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
143 SE ? &SE->getSE() : nullptr, false);
146 /// This transformation requires natural loop information & requires that
147 /// loop preheaders be inserted into the CFG...
149 void getAnalysisUsage(AnalysisUsage &AU) const override {
150 AU.setPreservesCFG();
151 AU.addRequired<TargetLibraryInfoWrapperPass>();
152 getLoopAnalysisUsage(AU);
155 using llvm::Pass::doFinalization;
157 bool doFinalization() override {
158 assert(LICM.getLoopToAliasSetMap().empty() &&
159 "Didn't free loop alias sets");
164 LoopInvariantCodeMotion LICM;
166 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
167 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
170 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
172 void deleteAnalysisValue(Value *V, Loop *L) override;
174 /// Simple Analysis hook. Delete loop L from alias set map.
175 void deleteAnalysisLoop(Loop *L) override;
179 PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM) {
181 AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager();
182 Function *F = L.getHeader()->getParent();
184 auto *AA = FAM.getCachedResult<AAManager>(*F);
185 auto *LI = FAM.getCachedResult<LoopAnalysis>(*F);
186 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F);
187 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F);
188 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F);
189 assert((AA && LI && DT && TLI && SE) && "Analyses for LICM not available");
191 LoopInvariantCodeMotion LICM;
193 if (!LICM.runOnLoop(&L, AA, LI, DT, TLI, SE, true))
194 return PreservedAnalyses::all();
196 // FIXME: There is no setPreservesCFG in the new PM. When that becomes
197 // available, it should be used here.
198 return getLoopPassPreservedAnalyses();
201 char LegacyLICMPass::ID = 0;
202 INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
204 INITIALIZE_PASS_DEPENDENCY(LoopPass)
205 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
206 INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
209 Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
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 /// We should delete AST for inner loops in the new pass manager to avoid
217 bool LoopInvariantCodeMotion::runOnLoop(Loop *L, AliasAnalysis *AA,
218 LoopInfo *LI, DominatorTree *DT,
219 TargetLibraryInfo *TLI,
220 ScalarEvolution *SE, bool DeleteAST) {
221 bool Changed = false;
223 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
225 AliasSetTracker *CurAST = collectAliasInfoForLoop(L, LI, AA);
227 // Get the preheader block to move instructions into...
228 BasicBlock *Preheader = L->getLoopPreheader();
230 // Compute loop safety information.
231 LoopSafetyInfo SafetyInfo;
232 computeLoopSafetyInfo(&SafetyInfo, L);
234 // We want to visit all of the instructions in this loop... that are not parts
235 // of our subloops (they have already had their invariants hoisted out of
236 // their loop, into this loop, so there is no need to process the BODIES of
239 // Traverse the body of the loop in depth first order on the dominator tree so
240 // that we are guaranteed to see definitions before we see uses. This allows
241 // us to sink instructions in one pass, without iteration. After sinking
242 // instructions, we perform another pass to hoist them out of the loop.
244 if (L->hasDedicatedExits())
245 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
246 CurAST, &SafetyInfo);
248 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
249 CurAST, &SafetyInfo);
251 // Now that all loop invariants have been removed from the loop, promote any
252 // memory references to scalars that we can.
253 // Don't sink stores from loops without dedicated block exits. Exits
254 // containing indirect branches are not transformed by loop simplify,
255 // make sure we catch that. An additional load may be generated in the
256 // preheader for SSA updater, so also avoid sinking when no preheader
258 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
259 // Figure out the loop exits and their insertion points
260 SmallVector<BasicBlock *, 8> ExitBlocks;
261 L->getUniqueExitBlocks(ExitBlocks);
263 // We can't insert into a catchswitch.
264 bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
265 return isa<CatchSwitchInst>(Exit->getTerminator());
268 if (!HasCatchSwitch) {
269 SmallVector<Instruction *, 8> InsertPts;
270 InsertPts.reserve(ExitBlocks.size());
271 for (BasicBlock *ExitBlock : ExitBlocks)
272 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
274 PredIteratorCache PIC;
276 bool Promoted = false;
278 // Loop over all of the alias sets in the tracker object.
279 for (AliasSet &AS : *CurAST)
281 promoteLoopAccessesToScalars(AS, ExitBlocks, InsertPts, PIC, LI, DT,
282 TLI, L, CurAST, &SafetyInfo);
284 // Once we have promoted values across the loop body we have to
285 // recursively reform LCSSA as any nested loop may now have values defined
286 // within the loop used in the outer loop.
287 // FIXME: This is really heavy handed. It would be a bit better to use an
288 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
291 formLCSSARecursively(*L, *DT, LI, SE);
297 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
298 // specifically moving instructions across the loop boundary and so it is
299 // especially in need of sanity checking here.
300 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
301 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
302 "Parent loop not left in LCSSA form after LICM!");
304 // If this loop is nested inside of another one, save the alias information
305 // for when we process the outer loop.
306 if (L->getParentLoop() && !DeleteAST)
307 LoopToAliasSetMap[L] = CurAST;
312 SE->forgetLoopDispositions(L);
316 /// Walk the specified region of the CFG (defined by all blocks dominated by
317 /// the specified block, and that are in the current loop) in reverse depth
318 /// first order w.r.t the DominatorTree. This allows us to visit uses before
319 /// definitions, allowing us to sink a loop body in one pass without iteration.
321 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
322 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
323 AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
326 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
327 CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&
328 "Unexpected input to sinkRegion");
330 BasicBlock *BB = N->getBlock();
331 // If this subregion is not in the top level loop at all, exit.
332 if (!CurLoop->contains(BB))
335 // We are processing blocks in reverse dfo, so process children first.
336 bool Changed = false;
337 const std::vector<DomTreeNode *> &Children = N->getChildren();
338 for (DomTreeNode *Child : Children)
339 Changed |= sinkRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
341 // Only need to process the contents of this block if it is not part of a
342 // subloop (which would already have been processed).
343 if (inSubLoop(BB, CurLoop, LI))
346 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
347 Instruction &I = *--II;
349 // If the instruction is dead, we would try to sink it because it isn't used
350 // in the loop, instead, just delete it.
351 if (isInstructionTriviallyDead(&I, TLI)) {
352 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
354 CurAST->deleteValue(&I);
360 // Check to see if we can sink this instruction to the exit blocks
361 // of the loop. We can do this if the all users of the instruction are
362 // outside of the loop. In this case, it doesn't even matter if the
363 // operands of the instruction are loop invariant.
365 if (isNotUsedInLoop(I, CurLoop, SafetyInfo) &&
366 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo)) {
368 Changed |= sink(I, LI, DT, CurLoop, CurAST, SafetyInfo);
374 /// Walk the specified region of the CFG (defined by all blocks dominated by
375 /// the specified block, and that are in the current loop) in depth first
376 /// order w.r.t the DominatorTree. This allows us to visit definitions before
377 /// uses, allowing us to hoist a loop body in one pass without iteration.
379 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
380 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
381 AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
383 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
384 CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr &&
385 "Unexpected input to hoistRegion");
387 BasicBlock *BB = N->getBlock();
389 // If this subregion is not in the top level loop at all, exit.
390 if (!CurLoop->contains(BB))
393 // Only need to process the contents of this block if it is not part of a
394 // subloop (which would already have been processed).
395 bool Changed = false;
396 if (!inSubLoop(BB, CurLoop, LI))
397 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
398 Instruction &I = *II++;
399 // Try constant folding this instruction. If all the operands are
400 // constants, it is technically hoistable, but it would be better to just
402 if (Constant *C = ConstantFoldInstruction(
403 &I, I.getModule()->getDataLayout(), TLI)) {
404 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
405 CurAST->copyValue(&I, C);
406 I.replaceAllUsesWith(C);
407 if (isInstructionTriviallyDead(&I, TLI)) {
408 CurAST->deleteValue(&I);
415 // Try hoisting the instruction out to the preheader. We can only do this
416 // if all of the operands of the instruction are loop invariant and if it
417 // is safe to hoist the instruction.
419 if (CurLoop->hasLoopInvariantOperands(&I) &&
420 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo) &&
421 isSafeToExecuteUnconditionally(
422 I, DT, CurLoop, SafetyInfo,
423 CurLoop->getLoopPreheader()->getTerminator()))
424 Changed |= hoist(I, DT, CurLoop, SafetyInfo);
427 const std::vector<DomTreeNode *> &Children = N->getChildren();
428 for (DomTreeNode *Child : Children)
429 Changed |= hoistRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
433 /// Computes loop safety information, checks loop body & header
434 /// for the possibility of may throw exception.
436 void llvm::computeLoopSafetyInfo(LoopSafetyInfo *SafetyInfo, Loop *CurLoop) {
437 assert(CurLoop != nullptr && "CurLoop cant be null");
438 BasicBlock *Header = CurLoop->getHeader();
439 // Setting default safety values.
440 SafetyInfo->MayThrow = false;
441 SafetyInfo->HeaderMayThrow = false;
442 // Iterate over header and compute safety info.
443 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
444 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
445 SafetyInfo->HeaderMayThrow |=
446 !isGuaranteedToTransferExecutionToSuccessor(&*I);
448 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
449 // Iterate over loop instructions and compute safety info.
450 for (Loop::block_iterator BB = CurLoop->block_begin(),
451 BBE = CurLoop->block_end();
452 (BB != BBE) && !SafetyInfo->MayThrow; ++BB)
453 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
454 (I != E) && !SafetyInfo->MayThrow; ++I)
455 SafetyInfo->MayThrow |= !isGuaranteedToTransferExecutionToSuccessor(&*I);
457 // Compute funclet colors if we might sink/hoist in a function with a funclet
458 // personality routine.
459 Function *Fn = CurLoop->getHeader()->getParent();
460 if (Fn->hasPersonalityFn())
461 if (Constant *PersonalityFn = Fn->getPersonalityFn())
462 if (isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)))
463 SafetyInfo->BlockColors = colorEHFunclets(*Fn);
466 bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
467 Loop *CurLoop, AliasSetTracker *CurAST,
468 LoopSafetyInfo *SafetyInfo) {
469 // Loads have extra constraints we have to verify before we can hoist them.
470 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
471 if (!LI->isUnordered())
472 return false; // Don't hoist volatile/atomic loads!
474 // Loads from constant memory are always safe to move, even if they end up
475 // in the same alias set as something that ends up being modified.
476 if (AA->pointsToConstantMemory(LI->getOperand(0)))
478 if (LI->getMetadata(LLVMContext::MD_invariant_load))
481 // Don't hoist loads which have may-aliased stores in loop.
483 if (LI->getType()->isSized())
484 Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType());
487 LI->getAAMetadata(AAInfo);
489 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
490 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
491 // Don't sink or hoist dbg info; it's legal, but not useful.
492 if (isa<DbgInfoIntrinsic>(I))
495 // Don't sink calls which can throw.
499 // Handle simple cases by querying alias analysis.
500 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
501 if (Behavior == FMRB_DoesNotAccessMemory)
503 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
504 // A readonly argmemonly function only reads from memory pointed to by
505 // it's arguments with arbitrary offsets. If we can prove there are no
506 // writes to this memory in the loop, we can hoist or sink.
507 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
508 for (Value *Op : CI->arg_operands())
509 if (Op->getType()->isPointerTy() &&
510 pointerInvalidatedByLoop(Op, MemoryLocation::UnknownSize,
511 AAMDNodes(), CurAST))
515 // If this call only reads from memory and there are no writes to memory
516 // in the loop, we can hoist or sink the call as appropriate.
517 bool FoundMod = false;
518 for (AliasSet &AS : *CurAST) {
519 if (!AS.isForwardingAliasSet() && AS.isMod()) {
528 // FIXME: This should use mod/ref information to see if we can hoist or
534 // Only these instructions are hoistable/sinkable.
535 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
536 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
537 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
538 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
539 !isa<InsertValueInst>(I))
542 // SafetyInfo is nullptr if we are checking for sinking from preheader to
543 // loop body. It will be always safe as there is no speculative execution.
547 // TODO: Plumb the context instruction through to make hoisting and sinking
548 // more powerful. Hoisting of loads already works due to the special casing
550 return isSafeToExecuteUnconditionally(I, DT, CurLoop, SafetyInfo, nullptr);
553 /// Returns true if a PHINode is a trivially replaceable with an
555 /// This is true when all incoming values are that instruction.
556 /// This pattern occurs most often with LCSSA PHI nodes.
558 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
559 for (const Value *IncValue : PN.incoming_values())
566 /// Return true if the only users of this instruction are outside of
567 /// the loop. If this is true, we can sink the instruction to the exit
568 /// blocks of the loop.
570 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
571 const LoopSafetyInfo *SafetyInfo) {
572 const auto &BlockColors = SafetyInfo->BlockColors;
573 for (const User *U : I.users()) {
574 const Instruction *UI = cast<Instruction>(U);
575 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
576 const BasicBlock *BB = PN->getParent();
577 // We cannot sink uses in catchswitches.
578 if (isa<CatchSwitchInst>(BB->getTerminator()))
581 // We need to sink a callsite to a unique funclet. Avoid sinking if the
582 // phi use is too muddled.
583 if (isa<CallInst>(I))
584 if (!BlockColors.empty() &&
585 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
588 // A PHI node where all of the incoming values are this instruction are
589 // special -- they can just be RAUW'ed with the instruction and thus
590 // don't require a use in the predecessor. This is a particular important
591 // special case because it is the pattern found in LCSSA form.
592 if (isTriviallyReplacablePHI(*PN, I)) {
593 if (CurLoop->contains(PN))
599 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
600 // values. Check for such a use being inside the loop.
601 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
602 if (PN->getIncomingValue(i) == &I)
603 if (CurLoop->contains(PN->getIncomingBlock(i)))
609 if (CurLoop->contains(UI))
616 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
618 const LoopSafetyInfo *SafetyInfo) {
620 if (auto *CI = dyn_cast<CallInst>(&I)) {
621 const auto &BlockColors = SafetyInfo->BlockColors;
623 // Sinking call-sites need to be handled differently from other
624 // instructions. The cloned call-site needs a funclet bundle operand
625 // appropriate for it's location in the CFG.
626 SmallVector<OperandBundleDef, 1> OpBundles;
627 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
628 BundleIdx != BundleEnd; ++BundleIdx) {
629 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
630 if (Bundle.getTagID() == LLVMContext::OB_funclet)
633 OpBundles.emplace_back(Bundle);
636 if (!BlockColors.empty()) {
637 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
638 assert(CV.size() == 1 && "non-unique color for exit block!");
639 BasicBlock *BBColor = CV.front();
640 Instruction *EHPad = BBColor->getFirstNonPHI();
641 if (EHPad->isEHPad())
642 OpBundles.emplace_back("funclet", EHPad);
645 New = CallInst::Create(CI, OpBundles);
650 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
651 if (!I.getName().empty())
652 New->setName(I.getName() + ".le");
654 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
655 // particularly cheap because we can rip off the PHI node that we're
656 // replacing for the number and blocks of the predecessors.
657 // OPT: If this shows up in a profile, we can instead finish sinking all
658 // invariant instructions, and then walk their operands to re-establish
659 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
660 // sinking bottom-up.
661 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
663 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
664 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
665 if (!OLoop->contains(&PN)) {
667 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
668 OInst->getName() + ".lcssa", &ExitBlock.front());
669 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
670 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
676 /// When an instruction is found to only be used outside of the loop, this
677 /// function moves it to the exit blocks and patches up SSA form as needed.
678 /// This method is guaranteed to remove the original instruction from its
679 /// position, and may either delete it or move it to outside of the loop.
681 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
682 const Loop *CurLoop, AliasSetTracker *CurAST,
683 const LoopSafetyInfo *SafetyInfo) {
684 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
685 bool Changed = false;
686 if (isa<LoadInst>(I))
688 else if (isa<CallInst>(I))
694 SmallVector<BasicBlock *, 32> ExitBlocks;
695 CurLoop->getUniqueExitBlocks(ExitBlocks);
696 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
700 // Clones of this instruction. Don't create more than one per exit block!
701 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
703 // If this instruction is only used outside of the loop, then all users are
704 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
706 while (!I.use_empty()) {
707 Value::user_iterator UI = I.user_begin();
708 auto *User = cast<Instruction>(*UI);
709 if (!DT->isReachableFromEntry(User->getParent())) {
710 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
713 // The user must be a PHI node.
714 PHINode *PN = cast<PHINode>(User);
716 // Surprisingly, instructions can be used outside of loops without any
717 // exits. This can only happen in PHI nodes if the incoming block is
719 Use &U = UI.getUse();
720 BasicBlock *BB = PN->getIncomingBlock(U);
721 if (!DT->isReachableFromEntry(BB)) {
722 U = UndefValue::get(I.getType());
726 BasicBlock *ExitBlock = PN->getParent();
727 assert(ExitBlockSet.count(ExitBlock) &&
728 "The LCSSA PHI is not in an exit block!");
731 auto It = SunkCopies.find(ExitBlock);
732 if (It != SunkCopies.end())
735 New = SunkCopies[ExitBlock] =
736 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI, SafetyInfo);
738 PN->replaceAllUsesWith(New);
739 PN->eraseFromParent();
742 CurAST->deleteValue(&I);
747 /// When an instruction is found to only use loop invariant operands that
748 /// is safe to hoist, this instruction is called to do the dirty work.
750 static bool hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
751 const LoopSafetyInfo *SafetyInfo) {
752 auto *Preheader = CurLoop->getLoopPreheader();
753 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": " << I
756 // Metadata can be dependent on conditions we are hoisting above.
757 // Conservatively strip all metadata on the instruction unless we were
758 // guaranteed to execute I if we entered the loop, in which case the metadata
759 // is valid in the loop preheader.
760 if (I.hasMetadataOtherThanDebugLoc() &&
761 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
762 // time in isGuaranteedToExecute if we don't actually have anything to
763 // drop. It is a compile time optimization, not required for correctness.
764 !isGuaranteedToExecute(I, DT, CurLoop, SafetyInfo))
765 I.dropUnknownNonDebugMetadata();
767 // Move the new node to the Preheader, before its terminator.
768 I.moveBefore(Preheader->getTerminator());
770 // Do not retain debug locations when we are moving instructions to different
771 // basic blocks, because we want to avoid jumpy line tables. Calls, however,
772 // need to retain their debug locs because they may be inlined.
773 // FIXME: How do we retain source locations without causing poor debugging
775 if (!isa<CallInst>(I))
776 I.setDebugLoc(DebugLoc());
778 if (isa<LoadInst>(I))
780 else if (isa<CallInst>(I))
786 /// Only sink or hoist an instruction if it is not a trapping instruction,
787 /// or if the instruction is known not to trap when moved to the preheader.
788 /// or if it is a trapping instruction and is guaranteed to execute.
789 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
790 const DominatorTree *DT,
792 const LoopSafetyInfo *SafetyInfo,
793 const Instruction *CtxI) {
794 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
797 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
801 class LoopPromoter : public LoadAndStorePromoter {
802 Value *SomePtr; // Designated pointer to store to.
803 SmallPtrSetImpl<Value *> &PointerMustAliases;
804 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
805 SmallVectorImpl<Instruction *> &LoopInsertPts;
806 PredIteratorCache &PredCache;
807 AliasSetTracker &AST;
813 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
814 if (Instruction *I = dyn_cast<Instruction>(V))
815 if (Loop *L = LI.getLoopFor(I->getParent()))
816 if (!L->contains(BB)) {
817 // We need to create an LCSSA PHI node for the incoming value and
819 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
820 I->getName() + ".lcssa", &BB->front());
821 for (BasicBlock *Pred : PredCache.get(BB))
822 PN->addIncoming(I, Pred);
829 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
830 SmallPtrSetImpl<Value *> &PMA,
831 SmallVectorImpl<BasicBlock *> &LEB,
832 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
833 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
834 const AAMDNodes &AATags)
835 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
836 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
837 LI(li), DL(std::move(dl)), Alignment(alignment), AATags(AATags) {}
839 bool isInstInList(Instruction *I,
840 const SmallVectorImpl<Instruction *> &) const override {
842 if (LoadInst *LI = dyn_cast<LoadInst>(I))
843 Ptr = LI->getOperand(0);
845 Ptr = cast<StoreInst>(I)->getPointerOperand();
846 return PointerMustAliases.count(Ptr);
849 void doExtraRewritesBeforeFinalDeletion() const override {
850 // Insert stores after in the loop exit blocks. Each exit block gets a
851 // store of the live-out values that feed them. Since we've already told
852 // the SSA updater about the defs in the loop and the preheader
853 // definition, it is all set and we can start using it.
854 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
855 BasicBlock *ExitBlock = LoopExitBlocks[i];
856 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
857 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
858 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
859 Instruction *InsertPos = LoopInsertPts[i];
860 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
861 NewSI->setAlignment(Alignment);
862 NewSI->setDebugLoc(DL);
864 NewSI->setAAMetadata(AATags);
868 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
869 // Update alias analysis.
870 AST.copyValue(LI, V);
872 void instructionDeleted(Instruction *I) const override { AST.deleteValue(I); }
874 } // end anon namespace
876 /// Try to promote memory values to scalars by sinking stores out of the
877 /// loop and moving loads to before the loop. We do this by looping over
878 /// the stores in the loop, looking for stores to Must pointers which are
881 bool llvm::promoteLoopAccessesToScalars(
882 AliasSet &AS, SmallVectorImpl<BasicBlock *> &ExitBlocks,
883 SmallVectorImpl<Instruction *> &InsertPts, PredIteratorCache &PIC,
884 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
885 Loop *CurLoop, AliasSetTracker *CurAST, LoopSafetyInfo *SafetyInfo) {
887 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
888 CurAST != nullptr && SafetyInfo != nullptr &&
889 "Unexpected Input to promoteLoopAccessesToScalars");
891 // We can promote this alias set if it has a store, if it is a "Must" alias
892 // set, if the pointer is loop invariant, and if we are not eliminating any
893 // volatile loads or stores.
894 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
895 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
898 assert(!AS.empty() &&
899 "Must alias set should have at least one pointer element in it!");
901 Value *SomePtr = AS.begin()->getValue();
902 BasicBlock *Preheader = CurLoop->getLoopPreheader();
904 // It isn't safe to promote a load/store from the loop if the load/store is
905 // conditional. For example, turning:
907 // for () { if (c) *P += 1; }
911 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
913 // is not safe, because *P may only be valid to access if 'c' is true.
915 // The safety property divides into two parts:
916 // p1) The memory may not be dereferenceable on entry to the loop. In this
917 // case, we can't insert the required load in the preheader.
918 // p2) The memory model does not allow us to insert a store along any dynamic
919 // path which did not originally have one.
921 // If at least one store is guaranteed to execute, both properties are
922 // satisfied, and promotion is legal.
924 // This, however, is not a necessary condition. Even if no store/load is
925 // guaranteed to execute, we can still establish these properties.
926 // We can establish (p1) by proving that hoisting the load into the preheader
927 // is safe (i.e. proving dereferenceability on all paths through the loop). We
928 // can use any access within the alias set to prove dereferenceability,
929 // since they're all must alias.
931 // There are two ways establish (p2):
932 // a) Prove the location is thread-local. In this case the memory model
933 // requirement does not apply, and stores are safe to insert.
934 // b) Prove a store dominates every exit block. In this case, if an exit
935 // blocks is reached, the original dynamic path would have taken us through
936 // the store, so inserting a store into the exit block is safe. Note that this
937 // is different from the store being guaranteed to execute. For instance,
938 // if an exception is thrown on the first iteration of the loop, the original
939 // store is never executed, but the exit blocks are not executed either.
941 bool DereferenceableInPH = false;
942 bool SafeToInsertStore = false;
944 SmallVector<Instruction *, 64> LoopUses;
945 SmallPtrSet<Value *, 4> PointerMustAliases;
947 // We start with an alignment of one and try to find instructions that allow
948 // us to prove better alignment.
949 unsigned Alignment = 1;
952 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
954 if (SafetyInfo->MayThrow) {
955 // If a loop can throw, we have to insert a store along each unwind edge.
956 // That said, we can't actually make the unwind edge explicit. Therefore,
957 // we have to prove that the store is dead along the unwind edge.
959 // Currently, this code just special-cases alloca instructions.
960 if (!isa<AllocaInst>(GetUnderlyingObject(SomePtr, MDL)))
964 // Check that all of the pointers in the alias set have the same type. We
965 // cannot (yet) promote a memory location that is loaded and stored in
966 // different sizes. While we are at it, collect alignment and AA info.
967 for (const auto &ASI : AS) {
968 Value *ASIV = ASI.getValue();
969 PointerMustAliases.insert(ASIV);
971 // Check that all of the pointers in the alias set have the same type. We
972 // cannot (yet) promote a memory location that is loaded and stored in
974 if (SomePtr->getType() != ASIV->getType())
977 for (User *U : ASIV->users()) {
978 // Ignore instructions that are outside the loop.
979 Instruction *UI = dyn_cast<Instruction>(U);
980 if (!UI || !CurLoop->contains(UI))
983 // If there is an non-load/store instruction in the loop, we can't promote
985 if (const LoadInst *Load = dyn_cast<LoadInst>(UI)) {
986 assert(!Load->isVolatile() && "AST broken");
987 if (!Load->isSimple())
990 if (!DereferenceableInPH)
991 DereferenceableInPH = isSafeToExecuteUnconditionally(
992 *Load, DT, CurLoop, SafetyInfo, Preheader->getTerminator());
993 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
994 // Stores *of* the pointer are not interesting, only stores *to* the
996 if (UI->getOperand(1) != ASIV)
998 assert(!Store->isVolatile() && "AST broken");
999 if (!Store->isSimple())
1002 // If the store is guaranteed to execute, both properties are satisfied.
1003 // We may want to check if a store is guaranteed to execute even if we
1004 // already know that promotion is safe, since it may have higher
1005 // alignment than any other guaranteed stores, in which case we can
1006 // raise the alignment on the promoted store.
1007 unsigned InstAlignment = Store->getAlignment();
1010 MDL.getABITypeAlignment(Store->getValueOperand()->getType());
1012 if (!DereferenceableInPH || !SafeToInsertStore ||
1013 (InstAlignment > Alignment)) {
1014 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
1015 DereferenceableInPH = true;
1016 SafeToInsertStore = true;
1017 Alignment = std::max(Alignment, InstAlignment);
1021 // If a store dominates all exit blocks, it is safe to sink.
1022 // As explained above, if an exit block was executed, a dominating
1023 // store must have been been executed at least once, so we are not
1024 // introducing stores on paths that did not have them.
1025 // Note that this only looks at explicit exit blocks. If we ever
1026 // start sinking stores into unwind edges (see above), this will break.
1027 if (!SafeToInsertStore)
1028 SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
1029 return DT->dominates(Store->getParent(), Exit);
1032 // If the store is not guaranteed to execute, we may still get
1033 // deref info through it.
1034 if (!DereferenceableInPH) {
1035 DereferenceableInPH = isDereferenceableAndAlignedPointer(
1036 Store->getPointerOperand(), Store->getAlignment(), MDL,
1037 Preheader->getTerminator(), DT);
1040 return false; // Not a load or store.
1042 // Merge the AA tags.
1043 if (LoopUses.empty()) {
1044 // On the first load/store, just take its AA tags.
1045 UI->getAAMetadata(AATags);
1046 } else if (AATags) {
1047 UI->getAAMetadata(AATags, /* Merge = */ true);
1050 LoopUses.push_back(UI);
1055 // If we couldn't prove we can hoist the load, bail.
1056 if (!DereferenceableInPH)
1059 // We know we can hoist the load, but don't have a guaranteed store.
1060 // Check whether the location is thread-local. If it is, then we can insert
1061 // stores along paths which originally didn't have them without violating the
1063 if (!SafeToInsertStore) {
1064 Value *Object = GetUnderlyingObject(SomePtr, MDL);
1066 (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
1067 !PointerMayBeCaptured(Object, true, true);
1070 // If we've still failed to prove we can sink the store, give up.
1071 if (!SafeToInsertStore)
1074 // Otherwise, this is safe to promote, lets do it!
1075 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr
1079 // Grab a debug location for the inserted loads/stores; given that the
1080 // inserted loads/stores have little relation to the original loads/stores,
1081 // this code just arbitrarily picks a location from one, since any debug
1082 // location is better than none.
1083 DebugLoc DL = LoopUses[0]->getDebugLoc();
1085 // We use the SSAUpdater interface to insert phi nodes as required.
1086 SmallVector<PHINode *, 16> NewPHIs;
1087 SSAUpdater SSA(&NewPHIs);
1088 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
1089 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
1091 // Set up the preheader to have a definition of the value. It is the live-out
1092 // value from the preheader that uses in the loop will use.
1093 LoadInst *PreheaderLoad = new LoadInst(
1094 SomePtr, SomePtr->getName() + ".promoted", Preheader->getTerminator());
1095 PreheaderLoad->setAlignment(Alignment);
1096 PreheaderLoad->setDebugLoc(DL);
1098 PreheaderLoad->setAAMetadata(AATags);
1099 SSA.AddAvailableValue(Preheader, PreheaderLoad);
1101 // Rewrite all the loads in the loop and remember all the definitions from
1102 // stores in the loop.
1103 Promoter.run(LoopUses);
1105 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
1106 if (PreheaderLoad->use_empty())
1107 PreheaderLoad->eraseFromParent();
1112 /// Returns an owning pointer to an alias set which incorporates aliasing info
1113 /// from L and all subloops of L.
1114 /// FIXME: In new pass manager, there is no helper function to handle loop
1115 /// analysis such as cloneBasicBlockAnalysis, so the AST needs to be recomputed
1116 /// from scratch for every loop. Hook up with the helper functions when
1117 /// available in the new pass manager to avoid redundant computation.
1119 LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
1120 AliasAnalysis *AA) {
1121 AliasSetTracker *CurAST = nullptr;
1122 SmallVector<Loop *, 4> RecomputeLoops;
1123 for (Loop *InnerL : L->getSubLoops()) {
1124 auto MapI = LoopToAliasSetMap.find(InnerL);
1125 // If the AST for this inner loop is missing it may have been merged into
1126 // some other loop's AST and then that loop unrolled, and so we need to
1128 if (MapI == LoopToAliasSetMap.end()) {
1129 RecomputeLoops.push_back(InnerL);
1132 AliasSetTracker *InnerAST = MapI->second;
1134 if (CurAST != nullptr) {
1135 // What if InnerLoop was modified by other passes ?
1136 CurAST->add(*InnerAST);
1138 // Once we've incorporated the inner loop's AST into ours, we don't need
1139 // the subloop's anymore.
1144 LoopToAliasSetMap.erase(MapI);
1146 if (CurAST == nullptr)
1147 CurAST = new AliasSetTracker(*AA);
1149 auto mergeLoop = [&](Loop *L) {
1150 // Loop over the body of this loop, looking for calls, invokes, and stores.
1151 // Because subloops have already been incorporated into AST, we skip blocks
1153 for (BasicBlock *BB : L->blocks())
1154 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
1155 CurAST->add(*BB); // Incorporate the specified basic block
1158 // Add everything from the sub loops that are no longer directly available.
1159 for (Loop *InnerL : RecomputeLoops)
1162 // And merge in this loop.
1168 /// Simple analysis hook. Clone alias set info.
1170 void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
1172 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1176 AST->copyValue(From, To);
1179 /// Simple Analysis hook. Delete value V from alias set
1181 void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) {
1182 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1186 AST->deleteValue(V);
1189 /// Simple Analysis hook. Delete value L from alias set map.
1191 void LegacyLICMPass::deleteAnalysisLoop(Loop *L) {
1192 AliasSetTracker *AST = LICM.getLoopToAliasSetMap().lookup(L);
1197 LICM.getLoopToAliasSetMap().erase(L);
1200 /// Return true if the body of this loop may store into the memory
1201 /// location pointed to by V.
1203 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
1204 const AAMDNodes &AAInfo,
1205 AliasSetTracker *CurAST) {
1206 // Check to see if any of the basic blocks in CurLoop invalidate *V.
1207 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
1210 /// Little predicate that returns true if the specified basic block is in
1211 /// a subloop of the current one, not the current one itself.
1213 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1214 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1215 return LI->getLoopFor(BB) != CurLoop;