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/SetOperations.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/AliasSetTracker.h"
38 #include "llvm/Analysis/BasicAliasAnalysis.h"
39 #include "llvm/Analysis/CaptureTracking.h"
40 #include "llvm/Analysis/ConstantFolding.h"
41 #include "llvm/Analysis/GlobalsModRef.h"
42 #include "llvm/Analysis/GuardUtils.h"
43 #include "llvm/Analysis/Loads.h"
44 #include "llvm/Analysis/LoopInfo.h"
45 #include "llvm/Analysis/LoopIterator.h"
46 #include "llvm/Analysis/LoopPass.h"
47 #include "llvm/Analysis/MemoryBuiltins.h"
48 #include "llvm/Analysis/MemorySSA.h"
49 #include "llvm/Analysis/MemorySSAUpdater.h"
50 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
51 #include "llvm/Analysis/ScalarEvolution.h"
52 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
53 #include "llvm/Analysis/TargetLibraryInfo.h"
54 #include "llvm/Analysis/ValueTracking.h"
55 #include "llvm/IR/CFG.h"
56 #include "llvm/IR/Constants.h"
57 #include "llvm/IR/DataLayout.h"
58 #include "llvm/IR/DerivedTypes.h"
59 #include "llvm/IR/Dominators.h"
60 #include "llvm/IR/Instructions.h"
61 #include "llvm/IR/IntrinsicInst.h"
62 #include "llvm/IR/LLVMContext.h"
63 #include "llvm/IR/Metadata.h"
64 #include "llvm/IR/PatternMatch.h"
65 #include "llvm/IR/PredIteratorCache.h"
66 #include "llvm/Support/CommandLine.h"
67 #include "llvm/Support/Debug.h"
68 #include "llvm/Support/raw_ostream.h"
69 #include "llvm/Transforms/Scalar.h"
70 #include "llvm/Transforms/Scalar/LoopPassManager.h"
71 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
72 #include "llvm/Transforms/Utils/Local.h"
73 #include "llvm/Transforms/Utils/LoopUtils.h"
74 #include "llvm/Transforms/Utils/SSAUpdater.h"
79 #define DEBUG_TYPE "licm"
81 STATISTIC(NumCreatedBlocks, "Number of blocks created");
82 STATISTIC(NumClonedBranches, "Number of branches cloned");
83 STATISTIC(NumSunk, "Number of instructions sunk out of loop");
84 STATISTIC(NumHoisted, "Number of instructions hoisted out of loop");
85 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
86 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
87 STATISTIC(NumPromoted, "Number of memory locations promoted to registers");
89 /// Memory promotion is enabled by default.
91 DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false),
92 cl::desc("Disable memory promotion in LICM pass"));
94 static cl::opt<bool> ControlFlowHoisting(
95 "licm-control-flow-hoisting", cl::Hidden, cl::init(false),
96 cl::desc("Enable control flow (and PHI) hoisting in LICM"));
98 static cl::opt<uint32_t> MaxNumUsesTraversed(
99 "licm-max-num-uses-traversed", cl::Hidden, cl::init(8),
100 cl::desc("Max num uses visited for identifying load "
101 "invariance in loop using invariant start (default = 8)"));
103 // Default value of zero implies we use the regular alias set tracker mechanism
104 // instead of the cross product using AA to identify aliasing of the memory
105 // location we are interested in.
107 LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0),
108 cl::desc("How many instruction to cross product using AA"));
110 // Experimental option to allow imprecision in LICM (use MemorySSA cap) in
111 // pathological cases, in exchange for faster compile. This is to be removed
112 // if MemorySSA starts to address the same issue. This flag applies only when
113 // LICM uses MemorySSA instead on AliasSetTracker. When the flag is disabled
114 // (default), LICM calls MemorySSAWalker's getClobberingMemoryAccess, which
115 // gets perfect accuracy. When flag is enabled, LICM will call into MemorySSA's
116 // getDefiningAccess, which may not be precise, since optimizeUses is capped.
117 static cl::opt<bool> EnableLicmCap(
118 "enable-licm-cap", cl::init(false), cl::Hidden,
119 cl::desc("Enable imprecision in LICM (uses MemorySSA cap) in "
120 "pathological cases, in exchange for faster compile"));
122 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
123 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
124 const LoopSafetyInfo *SafetyInfo,
125 TargetTransformInfo *TTI, bool &FreeInLoop);
126 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
127 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
128 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE);
129 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
130 const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo,
131 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE,
133 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
134 const DominatorTree *DT,
136 const LoopSafetyInfo *SafetyInfo,
137 OptimizationRemarkEmitter *ORE,
138 const Instruction *CtxI = nullptr);
139 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
140 AliasSetTracker *CurAST, Loop *CurLoop,
142 static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
144 static Instruction *CloneInstructionInExitBlock(
145 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
146 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU);
148 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
149 AliasSetTracker *AST, MemorySSAUpdater *MSSAU);
151 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
152 ICFLoopSafetyInfo &SafetyInfo);
155 struct LoopInvariantCodeMotion {
156 using ASTrackerMapTy = DenseMap<Loop *, std::unique_ptr<AliasSetTracker>>;
157 bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
158 TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
159 ScalarEvolution *SE, MemorySSA *MSSA,
160 OptimizationRemarkEmitter *ORE, bool DeleteAST);
162 ASTrackerMapTy &getLoopToAliasSetMap() { return LoopToAliasSetMap; }
165 ASTrackerMapTy LoopToAliasSetMap;
167 std::unique_ptr<AliasSetTracker>
168 collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AliasAnalysis *AA);
171 struct LegacyLICMPass : public LoopPass {
172 static char ID; // Pass identification, replacement for typeid
173 LegacyLICMPass() : LoopPass(ID) {
174 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
177 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
179 // If we have run LICM on a previous loop but now we are skipping
180 // (because we've hit the opt-bisect limit), we need to clear the
181 // loop alias information.
182 LICM.getLoopToAliasSetMap().clear();
186 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
187 MemorySSA *MSSA = EnableMSSALoopDependency
188 ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA())
190 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
191 // pass. Function analyses need to be preserved across loop transformations
192 // but ORE cannot be preserved (see comment before the pass definition).
193 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
194 return LICM.runOnLoop(L,
195 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
196 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
197 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
198 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
199 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
200 *L->getHeader()->getParent()),
201 SE ? &SE->getSE() : nullptr, MSSA, &ORE, false);
204 /// This transformation requires natural loop information & requires that
205 /// loop preheaders be inserted into the CFG...
207 void getAnalysisUsage(AnalysisUsage &AU) const override {
208 AU.addPreserved<DominatorTreeWrapperPass>();
209 AU.addPreserved<LoopInfoWrapperPass>();
210 AU.addRequired<TargetLibraryInfoWrapperPass>();
211 if (EnableMSSALoopDependency) {
212 AU.addRequired<MemorySSAWrapperPass>();
213 AU.addPreserved<MemorySSAWrapperPass>();
215 AU.addRequired<TargetTransformInfoWrapperPass>();
216 getLoopAnalysisUsage(AU);
219 using llvm::Pass::doFinalization;
221 bool doFinalization() override {
222 assert(LICM.getLoopToAliasSetMap().empty() &&
223 "Didn't free loop alias sets");
228 LoopInvariantCodeMotion LICM;
230 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
231 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
234 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
236 void deleteAnalysisValue(Value *V, Loop *L) override;
238 /// Simple Analysis hook. Delete loop L from alias set map.
239 void deleteAnalysisLoop(Loop *L) override;
243 PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
244 LoopStandardAnalysisResults &AR, LPMUpdater &) {
246 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
247 Function *F = L.getHeader()->getParent();
249 auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
250 // FIXME: This should probably be optional rather than required.
252 report_fatal_error("LICM: OptimizationRemarkEmitterAnalysis not "
253 "cached at a higher level");
255 LoopInvariantCodeMotion LICM;
256 if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, &AR.TLI, &AR.TTI, &AR.SE,
258 return PreservedAnalyses::all();
260 auto PA = getLoopPassPreservedAnalyses();
262 PA.preserve<DominatorTreeAnalysis>();
263 PA.preserve<LoopAnalysis>();
268 char LegacyLICMPass::ID = 0;
269 INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",
271 INITIALIZE_PASS_DEPENDENCY(LoopPass)
272 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
273 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
274 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
275 INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,
278 Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
280 /// Hoist expressions out of the specified loop. Note, alias info for inner
281 /// loop is not preserved so it is not a good idea to run LICM multiple
282 /// times on one loop.
283 /// We should delete AST for inner loops in the new pass manager to avoid
286 bool LoopInvariantCodeMotion::runOnLoop(
287 Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
288 TargetLibraryInfo *TLI, TargetTransformInfo *TTI, ScalarEvolution *SE,
289 MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, bool DeleteAST) {
290 bool Changed = false;
292 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
294 std::unique_ptr<AliasSetTracker> CurAST;
295 std::unique_ptr<MemorySSAUpdater> MSSAU;
297 LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n");
298 CurAST = collectAliasInfoForLoop(L, LI, AA);
300 LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA. Promotion disabled.\n");
301 MSSAU = make_unique<MemorySSAUpdater>(MSSA);
304 // Get the preheader block to move instructions into...
305 BasicBlock *Preheader = L->getLoopPreheader();
307 // Compute loop safety information.
308 ICFLoopSafetyInfo SafetyInfo(DT);
309 SafetyInfo.computeLoopSafetyInfo(L);
311 // We want to visit all of the instructions in this loop... that are not parts
312 // of our subloops (they have already had their invariants hoisted out of
313 // their loop, into this loop, so there is no need to process the BODIES of
316 // Traverse the body of the loop in depth first order on the dominator tree so
317 // that we are guaranteed to see definitions before we see uses. This allows
318 // us to sink instructions in one pass, without iteration. After sinking
319 // instructions, we perform another pass to hoist them out of the loop.
321 if (L->hasDedicatedExits())
322 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, TTI, L,
323 CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
325 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
326 CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
328 // Now that all loop invariants have been removed from the loop, promote any
329 // memory references to scalars that we can.
330 // Don't sink stores from loops without dedicated block exits. Exits
331 // containing indirect branches are not transformed by loop simplify,
332 // make sure we catch that. An additional load may be generated in the
333 // preheader for SSA updater, so also avoid sinking when no preheader
335 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
336 // Figure out the loop exits and their insertion points
337 SmallVector<BasicBlock *, 8> ExitBlocks;
338 L->getUniqueExitBlocks(ExitBlocks);
340 // We can't insert into a catchswitch.
341 bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
342 return isa<CatchSwitchInst>(Exit->getTerminator());
345 if (!HasCatchSwitch) {
346 SmallVector<Instruction *, 8> InsertPts;
347 InsertPts.reserve(ExitBlocks.size());
348 for (BasicBlock *ExitBlock : ExitBlocks)
349 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
351 PredIteratorCache PIC;
353 bool Promoted = false;
356 // Loop over all of the alias sets in the tracker object.
357 for (AliasSet &AS : *CurAST) {
358 // We can promote this alias set if it has a store, if it is a "Must"
359 // alias set, if the pointer is loop invariant, and if we are not
360 // eliminating any volatile loads or stores.
361 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
362 !L->isLoopInvariant(AS.begin()->getValue()))
367 "Must alias set should have at least one pointer element in it!");
369 SmallSetVector<Value *, 8> PointerMustAliases;
370 for (const auto &ASI : AS)
371 PointerMustAliases.insert(ASI.getValue());
373 Promoted |= promoteLoopAccessesToScalars(
374 PointerMustAliases, ExitBlocks, InsertPts, PIC, LI, DT, TLI, L,
375 CurAST.get(), &SafetyInfo, ORE);
378 // FIXME: Promotion initially disabled when using MemorySSA.
380 // Once we have promoted values across the loop body we have to
381 // recursively reform LCSSA as any nested loop may now have values defined
382 // within the loop used in the outer loop.
383 // FIXME: This is really heavy handed. It would be a bit better to use an
384 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
387 formLCSSARecursively(*L, *DT, LI, SE);
393 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
394 // specifically moving instructions across the loop boundary and so it is
395 // especially in need of sanity checking here.
396 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
397 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
398 "Parent loop not left in LCSSA form after LICM!");
400 // If this loop is nested inside of another one, save the alias information
401 // for when we process the outer loop.
402 if (CurAST.get() && L->getParentLoop() && !DeleteAST)
403 LoopToAliasSetMap[L] = std::move(CurAST);
405 if (MSSAU.get() && VerifyMemorySSA)
406 MSSAU->getMemorySSA()->verifyMemorySSA();
409 SE->forgetLoopDispositions(L);
413 /// Walk the specified region of the CFG (defined by all blocks dominated by
414 /// the specified block, and that are in the current loop) in reverse depth
415 /// first order w.r.t the DominatorTree. This allows us to visit uses before
416 /// definitions, allowing us to sink a loop body in one pass without iteration.
418 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
419 DominatorTree *DT, TargetLibraryInfo *TLI,
420 TargetTransformInfo *TTI, Loop *CurLoop,
421 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
422 ICFLoopSafetyInfo *SafetyInfo,
423 OptimizationRemarkEmitter *ORE) {
426 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
427 CurLoop != nullptr && SafetyInfo != nullptr &&
428 "Unexpected input to sinkRegion.");
429 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
430 "Either AliasSetTracker or MemorySSA should be initialized.");
432 // We want to visit children before parents. We will enque all the parents
433 // before their children in the worklist and process the worklist in reverse
435 SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
437 bool Changed = false;
438 for (DomTreeNode *DTN : reverse(Worklist)) {
439 BasicBlock *BB = DTN->getBlock();
440 // Only need to process the contents of this block if it is not part of a
441 // subloop (which would already have been processed).
442 if (inSubLoop(BB, CurLoop, LI))
445 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
446 Instruction &I = *--II;
448 // If the instruction is dead, we would try to sink it because it isn't
449 // used in the loop, instead, just delete it.
450 if (isInstructionTriviallyDead(&I, TLI)) {
451 LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
454 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
459 // Check to see if we can sink this instruction to the exit blocks
460 // of the loop. We can do this if the all users of the instruction are
461 // outside of the loop. In this case, it doesn't even matter if the
462 // operands of the instruction are loop invariant.
464 bool FreeInLoop = false;
465 if (isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) &&
466 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, ORE) &&
467 !I.mayHaveSideEffects()) {
468 if (sink(I, LI, DT, CurLoop, SafetyInfo, MSSAU, ORE, FreeInLoop)) {
471 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
478 if (MSSAU && VerifyMemorySSA)
479 MSSAU->getMemorySSA()->verifyMemorySSA();
484 // This is a helper class for hoistRegion to make it able to hoist control flow
485 // in order to be able to hoist phis. The way this works is that we initially
486 // start hoisting to the loop preheader, and when we see a loop invariant branch
487 // we make note of this. When we then come to hoist an instruction that's
488 // conditional on such a branch we duplicate the branch and the relevant control
489 // flow, then hoist the instruction into the block corresponding to its original
490 // block in the duplicated control flow.
491 class ControlFlowHoister {
493 // Information about the loop we are hoisting from
497 MemorySSAUpdater *MSSAU;
499 // A map of blocks in the loop to the block their instructions will be hoisted
501 DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap;
503 // The branches that we can hoist, mapped to the block that marks a
504 // convergence point of their control flow.
505 DenseMap<BranchInst *, BasicBlock *> HoistableBranches;
508 ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop,
509 MemorySSAUpdater *MSSAU)
510 : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {}
512 void registerPossiblyHoistableBranch(BranchInst *BI) {
513 // We can only hoist conditional branches with loop invariant operands.
514 if (!ControlFlowHoisting || !BI->isConditional() ||
515 !CurLoop->hasLoopInvariantOperands(BI))
518 // The branch destinations need to be in the loop, and we don't gain
519 // anything by duplicating conditional branches with duplicate successors,
520 // as it's essentially the same as an unconditional branch.
521 BasicBlock *TrueDest = BI->getSuccessor(0);
522 BasicBlock *FalseDest = BI->getSuccessor(1);
523 if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) ||
524 TrueDest == FalseDest)
527 // We can hoist BI if one branch destination is the successor of the other,
528 // or both have common successor which we check by seeing if the
529 // intersection of their successors is non-empty.
530 // TODO: This could be expanded to allowing branches where both ends
531 // eventually converge to a single block.
532 SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc;
533 TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest));
534 FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest));
535 BasicBlock *CommonSucc = nullptr;
536 if (TrueDestSucc.count(FalseDest)) {
537 CommonSucc = FalseDest;
538 } else if (FalseDestSucc.count(TrueDest)) {
539 CommonSucc = TrueDest;
541 set_intersect(TrueDestSucc, FalseDestSucc);
542 // If there's one common successor use that.
543 if (TrueDestSucc.size() == 1)
544 CommonSucc = *TrueDestSucc.begin();
545 // If there's more than one pick whichever appears first in the block list
546 // (we can't use the value returned by TrueDestSucc.begin() as it's
547 // unpredicatable which element gets returned).
548 else if (!TrueDestSucc.empty()) {
549 Function *F = TrueDest->getParent();
550 auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); };
551 auto It = std::find_if(F->begin(), F->end(), IsSucc);
552 assert(It != F->end() && "Could not find successor in function");
556 // The common successor has to be dominated by the branch, as otherwise
557 // there will be some other path to the successor that will not be
558 // controlled by this branch so any phi we hoist would be controlled by the
559 // wrong condition. This also takes care of avoiding hoisting of loop back
561 // TODO: In some cases this could be relaxed if the successor is dominated
562 // by another block that's been hoisted and we can guarantee that the
563 // control flow has been replicated exactly.
564 if (CommonSucc && DT->dominates(BI, CommonSucc))
565 HoistableBranches[BI] = CommonSucc;
568 bool canHoistPHI(PHINode *PN) {
569 // The phi must have loop invariant operands.
570 if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN))
572 // We can hoist phis if the block they are in is the target of hoistable
573 // branches which cover all of the predecessors of the block.
574 SmallPtrSet<BasicBlock *, 8> PredecessorBlocks;
575 BasicBlock *BB = PN->getParent();
576 for (BasicBlock *PredBB : predecessors(BB))
577 PredecessorBlocks.insert(PredBB);
578 // If we have less predecessor blocks than predecessors then the phi will
579 // have more than one incoming value for the same block which we can't
581 // TODO: This could be handled be erasing some of the duplicate incoming
583 if (PredecessorBlocks.size() != pred_size(BB))
585 for (auto &Pair : HoistableBranches) {
586 if (Pair.second == BB) {
587 // Which blocks are predecessors via this branch depends on if the
588 // branch is triangle-like or diamond-like.
589 if (Pair.first->getSuccessor(0) == BB) {
590 PredecessorBlocks.erase(Pair.first->getParent());
591 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
592 } else if (Pair.first->getSuccessor(1) == BB) {
593 PredecessorBlocks.erase(Pair.first->getParent());
594 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
596 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
597 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
601 // PredecessorBlocks will now be empty if for every predecessor of BB we
602 // found a hoistable branch source.
603 return PredecessorBlocks.empty();
606 BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) {
607 if (!ControlFlowHoisting)
608 return CurLoop->getLoopPreheader();
609 // If BB has already been hoisted, return that
610 if (HoistDestinationMap.count(BB))
611 return HoistDestinationMap[BB];
613 // Check if this block is conditional based on a pending branch
614 auto HasBBAsSuccessor =
615 [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) {
616 return BB != Pair.second && (Pair.first->getSuccessor(0) == BB ||
617 Pair.first->getSuccessor(1) == BB);
619 auto It = std::find_if(HoistableBranches.begin(), HoistableBranches.end(),
622 // If not involved in a pending branch, hoist to preheader
623 BasicBlock *InitialPreheader = CurLoop->getLoopPreheader();
624 if (It == HoistableBranches.end()) {
625 LLVM_DEBUG(dbgs() << "LICM using " << InitialPreheader->getName()
626 << " as hoist destination for " << BB->getName()
628 HoistDestinationMap[BB] = InitialPreheader;
629 return InitialPreheader;
631 BranchInst *BI = It->first;
632 assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==
633 HoistableBranches.end() &&
634 "BB is expected to be the target of at most one branch");
636 LLVMContext &C = BB->getContext();
637 BasicBlock *TrueDest = BI->getSuccessor(0);
638 BasicBlock *FalseDest = BI->getSuccessor(1);
639 BasicBlock *CommonSucc = HoistableBranches[BI];
640 BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent());
642 // Create hoisted versions of blocks that currently don't have them
643 auto CreateHoistedBlock = [&](BasicBlock *Orig) {
644 if (HoistDestinationMap.count(Orig))
645 return HoistDestinationMap[Orig];
647 BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent());
648 HoistDestinationMap[Orig] = New;
649 DT->addNewBlock(New, HoistTarget);
650 if (CurLoop->getParentLoop())
651 CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI);
653 LLVM_DEBUG(dbgs() << "LICM created " << New->getName()
654 << " as hoist destination for " << Orig->getName()
658 BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest);
659 BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest);
660 BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc);
662 // Link up these blocks with branches.
663 if (!HoistCommonSucc->getTerminator()) {
664 // The new common successor we've generated will branch to whatever that
665 // hoist target branched to.
666 BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor();
667 assert(TargetSucc && "Expected hoist target to have a single successor");
668 HoistCommonSucc->moveBefore(TargetSucc);
669 BranchInst::Create(TargetSucc, HoistCommonSucc);
671 if (!HoistTrueDest->getTerminator()) {
672 HoistTrueDest->moveBefore(HoistCommonSucc);
673 BranchInst::Create(HoistCommonSucc, HoistTrueDest);
675 if (!HoistFalseDest->getTerminator()) {
676 HoistFalseDest->moveBefore(HoistCommonSucc);
677 BranchInst::Create(HoistCommonSucc, HoistFalseDest);
680 // If BI is being cloned to what was originally the preheader then
681 // HoistCommonSucc will now be the new preheader.
682 if (HoistTarget == InitialPreheader) {
683 // Phis in the loop header now need to use the new preheader.
684 InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc);
686 MSSAU->wireOldPredecessorsToNewImmediatePredecessor(
687 HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget});
688 // The new preheader dominates the loop header.
689 DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc);
690 DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader());
691 DT->changeImmediateDominator(HeaderNode, PreheaderNode);
692 // The preheader hoist destination is now the new preheader, with the
693 // exception of the hoist destination of this branch.
694 for (auto &Pair : HoistDestinationMap)
695 if (Pair.second == InitialPreheader && Pair.first != BI->getParent())
696 Pair.second = HoistCommonSucc;
699 // Now finally clone BI.
701 HoistTarget->getTerminator(),
702 BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition()));
705 assert(CurLoop->getLoopPreheader() &&
706 "Hoisting blocks should not have destroyed preheader");
707 return HoistDestinationMap[BB];
712 /// Walk the specified region of the CFG (defined by all blocks dominated by
713 /// the specified block, and that are in the current loop) in depth first
714 /// order w.r.t the DominatorTree. This allows us to visit definitions before
715 /// uses, allowing us to hoist a loop body in one pass without iteration.
717 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
718 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
719 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
720 ICFLoopSafetyInfo *SafetyInfo,
721 OptimizationRemarkEmitter *ORE) {
723 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
724 CurLoop != nullptr && SafetyInfo != nullptr &&
725 "Unexpected input to hoistRegion.");
726 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
727 "Either AliasSetTracker or MemorySSA should be initialized.");
729 ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
731 // Keep track of instructions that have been hoisted, as they may need to be
732 // re-hoisted if they end up not dominating all of their uses.
733 SmallVector<Instruction *, 16> HoistedInstructions;
735 // For PHI hoisting to work we need to hoist blocks before their successors.
736 // We can do this by iterating through the blocks in the loop in reverse
738 LoopBlocksRPO Worklist(CurLoop);
739 Worklist.perform(LI);
740 bool Changed = false;
741 for (BasicBlock *BB : Worklist) {
742 // Only need to process the contents of this block if it is not part of a
743 // subloop (which would already have been processed).
744 if (inSubLoop(BB, CurLoop, LI))
747 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
748 Instruction &I = *II++;
749 // Try constant folding this instruction. If all the operands are
750 // constants, it is technically hoistable, but it would be better to
752 if (Constant *C = ConstantFoldInstruction(
753 &I, I.getModule()->getDataLayout(), TLI)) {
754 LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C
757 CurAST->copyValue(&I, C);
758 // FIXME MSSA: Such replacements may make accesses unoptimized (D51960).
759 I.replaceAllUsesWith(C);
760 if (isInstructionTriviallyDead(&I, TLI))
761 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
766 // Try hoisting the instruction out to the preheader. We can only do
767 // this if all of the operands of the instruction are loop invariant and
768 // if it is safe to hoist the instruction.
769 // TODO: It may be safe to hoist if we are hoisting to a conditional block
770 // and we have accurately duplicated the control flow from the loop header
772 if (CurLoop->hasLoopInvariantOperands(&I) &&
773 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, ORE) &&
774 isSafeToExecuteUnconditionally(
775 I, DT, CurLoop, SafetyInfo, ORE,
776 CurLoop->getLoopPreheader()->getTerminator())) {
777 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
779 HoistedInstructions.push_back(&I);
784 // Attempt to remove floating point division out of the loop by
785 // converting it to a reciprocal multiplication.
786 if (I.getOpcode() == Instruction::FDiv &&
787 CurLoop->isLoopInvariant(I.getOperand(1)) &&
788 I.hasAllowReciprocal()) {
789 auto Divisor = I.getOperand(1);
790 auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
791 auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
792 ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
793 SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent());
794 ReciprocalDivisor->insertBefore(&I);
797 BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
798 Product->setFastMathFlags(I.getFastMathFlags());
799 SafetyInfo->insertInstructionTo(Product, I.getParent());
800 Product->insertAfter(&I);
801 I.replaceAllUsesWith(Product);
802 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
804 hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB),
805 SafetyInfo, MSSAU, ORE);
806 HoistedInstructions.push_back(ReciprocalDivisor);
811 using namespace PatternMatch;
812 if (((I.use_empty() &&
813 match(&I, m_Intrinsic<Intrinsic::invariant_start>())) ||
815 CurLoop->hasLoopInvariantOperands(&I) &&
816 SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
817 SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop)) {
818 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
820 HoistedInstructions.push_back(&I);
825 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
826 if (CFH.canHoistPHI(PN)) {
827 // Redirect incoming blocks first to ensure that we create hoisted
828 // versions of those blocks before we hoist the phi.
829 for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i)
830 PN->setIncomingBlock(
831 i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i)));
832 hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
834 assert(DT->dominates(PN, BB) && "Conditional PHIs not expected");
840 // Remember possibly hoistable branches so we can actually hoist them
842 if (BranchInst *BI = dyn_cast<BranchInst>(&I))
843 CFH.registerPossiblyHoistableBranch(BI);
847 // If we hoisted instructions to a conditional block they may not dominate
848 // their uses that weren't hoisted (such as phis where some operands are not
849 // loop invariant). If so make them unconditional by moving them to their
850 // immediate dominator. We iterate through the instructions in reverse order
851 // which ensures that when we rehoist an instruction we rehoist its operands,
852 // and also keep track of where in the block we are rehoisting to to make sure
853 // that we rehoist instructions before the instructions that use them.
854 Instruction *HoistPoint = nullptr;
855 if (ControlFlowHoisting) {
856 for (Instruction *I : reverse(HoistedInstructions)) {
857 if (!llvm::all_of(I->uses(),
858 [&](Use &U) { return DT->dominates(I, U); })) {
859 BasicBlock *Dominator =
860 DT->getNode(I->getParent())->getIDom()->getBlock();
861 if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) {
863 assert(DT->dominates(Dominator, HoistPoint->getParent()) &&
864 "New hoist point expected to dominate old hoist point");
865 HoistPoint = Dominator->getTerminator();
867 LLVM_DEBUG(dbgs() << "LICM rehoisting to "
868 << HoistPoint->getParent()->getName()
869 << ": " << *I << "\n");
870 moveInstructionBefore(*I, *HoistPoint, *SafetyInfo);
876 if (MSSAU && VerifyMemorySSA)
877 MSSAU->getMemorySSA()->verifyMemorySSA();
879 // Now that we've finished hoisting make sure that LI and DT are still
883 assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&
884 "Dominator tree verification failed");
892 // Return true if LI is invariant within scope of the loop. LI is invariant if
893 // CurLoop is dominated by an invariant.start representing the same memory
894 // location and size as the memory location LI loads from, and also the
895 // invariant.start has no uses.
896 static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
898 Value *Addr = LI->getOperand(0);
899 const DataLayout &DL = LI->getModule()->getDataLayout();
900 const uint32_t LocSizeInBits = DL.getTypeSizeInBits(
901 cast<PointerType>(Addr->getType())->getElementType());
903 // if the type is i8 addrspace(x)*, we know this is the type of
904 // llvm.invariant.start operand
905 auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()),
906 LI->getPointerAddressSpace());
907 unsigned BitcastsVisited = 0;
908 // Look through bitcasts until we reach the i8* type (this is invariant.start
910 while (Addr->getType() != PtrInt8Ty) {
911 auto *BC = dyn_cast<BitCastInst>(Addr);
912 // Avoid traversing high number of bitcast uses.
913 if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
915 Addr = BC->getOperand(0);
918 unsigned UsesVisited = 0;
919 // Traverse all uses of the load operand value, to see if invariant.start is
920 // one of the uses, and whether it dominates the load instruction.
921 for (auto *U : Addr->users()) {
922 // Avoid traversing for Load operand with high number of users.
923 if (++UsesVisited > MaxNumUsesTraversed)
925 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
926 // If there are escaping uses of invariant.start instruction, the load maybe
928 if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
931 unsigned InvariantSizeInBits =
932 cast<ConstantInt>(II->getArgOperand(0))->getSExtValue() * 8;
933 // Confirm the invariant.start location size contains the load operand size
934 // in bits. Also, the invariant.start should dominate the load, and we
935 // should not hoist the load out of a loop that contains this dominating
937 if (LocSizeInBits <= InvariantSizeInBits &&
938 DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
946 /// Return true if-and-only-if we know how to (mechanically) both hoist and
947 /// sink a given instruction out of a loop. Does not address legality
948 /// concerns such as aliasing or speculation safety.
949 bool isHoistableAndSinkableInst(Instruction &I) {
950 // Only these instructions are hoistable/sinkable.
951 return (isa<LoadInst>(I) || isa<StoreInst>(I) ||
952 isa<CallInst>(I) || isa<FenceInst>(I) ||
953 isa<BinaryOperator>(I) || isa<CastInst>(I) ||
954 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) ||
955 isa<CmpInst>(I) || isa<InsertElementInst>(I) ||
956 isa<ExtractElementInst>(I) || isa<ShuffleVectorInst>(I) ||
957 isa<ExtractValueInst>(I) || isa<InsertValueInst>(I));
959 /// Return true if all of the alias sets within this AST are known not to
960 /// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop.
961 bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU,
964 for (AliasSet &AS : *CurAST) {
965 if (!AS.isForwardingAliasSet() && AS.isMod()) {
971 for (auto *BB : L->getBlocks())
972 if (MSSAU->getMemorySSA()->getBlockDefs(BB))
978 /// Return true if I is the only Instruction with a MemoryAccess in L.
979 bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
980 const MemorySSAUpdater *MSSAU) {
981 for (auto *BB : L->getBlocks())
982 if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) {
984 for (const auto &Acc : *Accs) {
985 if (isa<MemoryPhi>(&Acc))
987 const auto *MUD = cast<MemoryUseOrDef>(&Acc);
988 if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
996 bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
997 Loop *CurLoop, AliasSetTracker *CurAST,
998 MemorySSAUpdater *MSSAU,
999 bool TargetExecutesOncePerLoop,
1000 OptimizationRemarkEmitter *ORE) {
1001 // If we don't understand the instruction, bail early.
1002 if (!isHoistableAndSinkableInst(I))
1005 MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
1007 // Loads have extra constraints we have to verify before we can hoist them.
1008 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
1009 if (!LI->isUnordered())
1010 return false; // Don't sink/hoist volatile or ordered atomic loads!
1012 // Loads from constant memory are always safe to move, even if they end up
1013 // in the same alias set as something that ends up being modified.
1014 if (AA->pointsToConstantMemory(LI->getOperand(0)))
1016 if (LI->getMetadata(LLVMContext::MD_invariant_load))
1019 if (LI->isAtomic() && !TargetExecutesOncePerLoop)
1020 return false; // Don't risk duplicating unordered loads
1022 // This checks for an invariant.start dominating the load.
1023 if (isLoadInvariantInLoop(LI, DT, CurLoop))
1028 Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST,
1031 Invalidated = pointerInvalidatedByLoopWithMSSA(
1032 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop);
1033 // Check loop-invariant address because this may also be a sinkable load
1034 // whose address is not necessarily loop-invariant.
1035 if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1037 return OptimizationRemarkMissed(
1038 DEBUG_TYPE, "LoadWithLoopInvariantAddressInvalidated", LI)
1039 << "failed to move load with loop-invariant address "
1040 "because the loop may invalidate its value";
1043 return !Invalidated;
1044 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1045 // Don't sink or hoist dbg info; it's legal, but not useful.
1046 if (isa<DbgInfoIntrinsic>(I))
1049 // Don't sink calls which can throw.
1053 using namespace PatternMatch;
1054 if (match(CI, m_Intrinsic<Intrinsic::assume>()))
1055 // Assumes don't actually alias anything or throw
1058 // Handle simple cases by querying alias analysis.
1059 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
1060 if (Behavior == FMRB_DoesNotAccessMemory)
1062 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
1063 // A readonly argmemonly function only reads from memory pointed to by
1064 // it's arguments with arbitrary offsets. If we can prove there are no
1065 // writes to this memory in the loop, we can hoist or sink.
1066 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
1067 // TODO: expand to writeable arguments
1068 for (Value *Op : CI->arg_operands())
1069 if (Op->getType()->isPointerTy()) {
1072 Invalidated = pointerInvalidatedByLoop(
1073 MemoryLocation(Op, LocationSize::unknown(), AAMDNodes()),
1074 CurAST, CurLoop, AA);
1076 Invalidated = pointerInvalidatedByLoopWithMSSA(
1077 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop);
1084 // If this call only reads from memory and there are no writes to memory
1085 // in the loop, we can hoist or sink the call as appropriate.
1086 if (isReadOnly(CurAST, MSSAU, CurLoop))
1090 // FIXME: This should use mod/ref information to see if we can hoist or
1094 } else if (auto *FI = dyn_cast<FenceInst>(&I)) {
1095 // Fences alias (most) everything to provide ordering. For the moment,
1096 // just give up if there are any other memory operations in the loop.
1098 auto Begin = CurAST->begin();
1099 assert(Begin != CurAST->end() && "must contain FI");
1100 if (std::next(Begin) != CurAST->end())
1101 // constant memory for instance, TODO: handle better
1103 auto *UniqueI = Begin->getUniqueInstruction();
1105 // other memory op, give up
1107 (void)FI; // suppress unused variable warning
1108 assert(UniqueI == FI && "AS must contain FI");
1111 return isOnlyMemoryAccess(FI, CurLoop, MSSAU);
1112 } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
1113 if (!SI->isUnordered())
1114 return false; // Don't sink/hoist volatile or ordered atomic store!
1116 // We can only hoist a store that we can prove writes a value which is not
1117 // read or overwritten within the loop. For those cases, we fallback to
1118 // load store promotion instead. TODO: We can extend this to cases where
1119 // there is exactly one write to the location and that write dominates an
1120 // arbitrary number of reads in the loop.
1122 auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI));
1124 if (AS.isRef() || !AS.isMustAlias())
1125 // Quick exit test, handled by the full path below as well.
1127 auto *UniqueI = AS.getUniqueInstruction();
1129 // other memory op, give up
1131 assert(UniqueI == SI && "AS must contain SI");
1134 if (isOnlyMemoryAccess(SI, CurLoop, MSSAU))
1136 if (!EnableLicmCap) {
1137 auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI);
1138 if (MSSA->isLiveOnEntryDef(Source) ||
1139 !CurLoop->contains(Source->getBlock()))
1146 assert(!I.mayReadOrWriteMemory() && "unhandled aliasing");
1148 // We've established mechanical ability and aliasing, it's up to the caller
1149 // to check fault safety
1153 /// Returns true if a PHINode is a trivially replaceable with an
1155 /// This is true when all incoming values are that instruction.
1156 /// This pattern occurs most often with LCSSA PHI nodes.
1158 static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
1159 for (const Value *IncValue : PN.incoming_values())
1166 /// Return true if the instruction is free in the loop.
1167 static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop,
1168 const TargetTransformInfo *TTI) {
1170 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1171 if (TTI->getUserCost(GEP) != TargetTransformInfo::TCC_Free)
1173 // For a GEP, we cannot simply use getUserCost because currently it
1174 // optimistically assume that a GEP will fold into addressing mode
1175 // regardless of its users.
1176 const BasicBlock *BB = GEP->getParent();
1177 for (const User *U : GEP->users()) {
1178 const Instruction *UI = cast<Instruction>(U);
1179 if (CurLoop->contains(UI) &&
1180 (BB != UI->getParent() ||
1181 (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
1186 return TTI->getUserCost(&I) == TargetTransformInfo::TCC_Free;
1189 /// Return true if the only users of this instruction are outside of
1190 /// the loop. If this is true, we can sink the instruction to the exit
1191 /// blocks of the loop.
1193 /// We also return true if the instruction could be folded away in lowering.
1194 /// (e.g., a GEP can be folded into a load as an addressing mode in the loop).
1195 static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
1196 const LoopSafetyInfo *SafetyInfo,
1197 TargetTransformInfo *TTI, bool &FreeInLoop) {
1198 const auto &BlockColors = SafetyInfo->getBlockColors();
1199 bool IsFree = isFreeInLoop(I, CurLoop, TTI);
1200 for (const User *U : I.users()) {
1201 const Instruction *UI = cast<Instruction>(U);
1202 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1203 const BasicBlock *BB = PN->getParent();
1204 // We cannot sink uses in catchswitches.
1205 if (isa<CatchSwitchInst>(BB->getTerminator()))
1208 // We need to sink a callsite to a unique funclet. Avoid sinking if the
1209 // phi use is too muddled.
1210 if (isa<CallInst>(I))
1211 if (!BlockColors.empty() &&
1212 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
1216 if (CurLoop->contains(UI)) {
1227 static Instruction *CloneInstructionInExitBlock(
1228 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
1229 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) {
1231 if (auto *CI = dyn_cast<CallInst>(&I)) {
1232 const auto &BlockColors = SafetyInfo->getBlockColors();
1234 // Sinking call-sites need to be handled differently from other
1235 // instructions. The cloned call-site needs a funclet bundle operand
1236 // appropriate for it's location in the CFG.
1237 SmallVector<OperandBundleDef, 1> OpBundles;
1238 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
1239 BundleIdx != BundleEnd; ++BundleIdx) {
1240 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
1241 if (Bundle.getTagID() == LLVMContext::OB_funclet)
1244 OpBundles.emplace_back(Bundle);
1247 if (!BlockColors.empty()) {
1248 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
1249 assert(CV.size() == 1 && "non-unique color for exit block!");
1250 BasicBlock *BBColor = CV.front();
1251 Instruction *EHPad = BBColor->getFirstNonPHI();
1252 if (EHPad->isEHPad())
1253 OpBundles.emplace_back("funclet", EHPad);
1256 New = CallInst::Create(CI, OpBundles);
1261 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
1262 if (!I.getName().empty())
1263 New->setName(I.getName() + ".le");
1265 MemoryAccess *OldMemAcc;
1266 if (MSSAU && (OldMemAcc = MSSAU->getMemorySSA()->getMemoryAccess(&I))) {
1267 // Create a new MemoryAccess and let MemorySSA set its defining access.
1268 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1269 New, nullptr, New->getParent(), MemorySSA::Beginning);
1271 if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
1272 MSSAU->insertDef(MemDef, /*RenameUses=*/true);
1274 auto *MemUse = cast<MemoryUse>(NewMemAcc);
1275 MSSAU->insertUse(MemUse);
1280 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
1281 // particularly cheap because we can rip off the PHI node that we're
1282 // replacing for the number and blocks of the predecessors.
1283 // OPT: If this shows up in a profile, we can instead finish sinking all
1284 // invariant instructions, and then walk their operands to re-establish
1285 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
1286 // sinking bottom-up.
1287 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
1289 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
1290 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
1291 if (!OLoop->contains(&PN)) {
1293 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
1294 OInst->getName() + ".lcssa", &ExitBlock.front());
1295 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1296 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
1302 static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
1303 AliasSetTracker *AST, MemorySSAUpdater *MSSAU) {
1305 AST->deleteValue(&I);
1307 MSSAU->removeMemoryAccess(&I);
1308 SafetyInfo.removeInstruction(&I);
1309 I.eraseFromParent();
1312 static void moveInstructionBefore(Instruction &I, Instruction &Dest,
1313 ICFLoopSafetyInfo &SafetyInfo) {
1314 SafetyInfo.removeInstruction(&I);
1315 SafetyInfo.insertInstructionTo(&I, Dest.getParent());
1316 I.moveBefore(&Dest);
1319 static Instruction *sinkThroughTriviallyReplaceablePHI(
1320 PHINode *TPN, Instruction *I, LoopInfo *LI,
1321 SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
1322 const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
1323 MemorySSAUpdater *MSSAU) {
1324 assert(isTriviallyReplaceablePHI(*TPN, *I) &&
1325 "Expect only trivially replaceable PHI");
1326 BasicBlock *ExitBlock = TPN->getParent();
1328 auto It = SunkCopies.find(ExitBlock);
1329 if (It != SunkCopies.end())
1332 New = SunkCopies[ExitBlock] = CloneInstructionInExitBlock(
1333 *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU);
1337 static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
1338 BasicBlock *BB = PN->getParent();
1339 if (!BB->canSplitPredecessors())
1341 // It's not impossible to split EHPad blocks, but if BlockColors already exist
1342 // it require updating BlockColors for all offspring blocks accordingly. By
1343 // skipping such corner case, we can make updating BlockColors after splitting
1344 // predecessor fairly simple.
1345 if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad())
1347 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1348 BasicBlock *BBPred = *PI;
1349 if (isa<IndirectBrInst>(BBPred->getTerminator()))
1355 static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
1356 LoopInfo *LI, const Loop *CurLoop,
1357 LoopSafetyInfo *SafetyInfo,
1358 MemorySSAUpdater *MSSAU) {
1360 SmallVector<BasicBlock *, 32> ExitBlocks;
1361 CurLoop->getUniqueExitBlocks(ExitBlocks);
1362 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1365 BasicBlock *ExitBB = PN->getParent();
1366 assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.");
1368 // Split predecessors of the loop exit to make instructions in the loop are
1369 // exposed to exit blocks through trivially replaceable PHIs while keeping the
1370 // loop in the canonical form where each predecessor of each exit block should
1371 // be contained within the loop. For example, this will convert the loop below
1381 // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
1387 // br %LE.split, %LB2
1390 // br %LE.split2, %LB1
1392 // %p1 = phi [%v1, %LB1] <-- trivially replaceable
1395 // %p2 = phi [%v2, %LB2] <-- trivially replaceable
1398 // %p = phi [%p1, %LE.split], [%p2, %LE.split2]
1400 const auto &BlockColors = SafetyInfo->getBlockColors();
1401 SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
1402 while (!PredBBs.empty()) {
1403 BasicBlock *PredBB = *PredBBs.begin();
1404 assert(CurLoop->contains(PredBB) &&
1405 "Expect all predecessors are in the loop");
1406 if (PN->getBasicBlockIndex(PredBB) >= 0) {
1407 BasicBlock *NewPred = SplitBlockPredecessors(
1408 ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true);
1409 // Since we do not allow splitting EH-block with BlockColors in
1410 // canSplitPredecessors(), we can simply assign predecessor's color to
1412 if (!BlockColors.empty())
1413 // Grab a reference to the ColorVector to be inserted before getting the
1414 // reference to the vector we are copying because inserting the new
1415 // element in BlockColors might cause the map to be reallocated.
1416 SafetyInfo->copyColors(NewPred, PredBB);
1418 PredBBs.remove(PredBB);
1422 /// When an instruction is found to only be used outside of the loop, this
1423 /// function moves it to the exit blocks and patches up SSA form as needed.
1424 /// This method is guaranteed to remove the original instruction from its
1425 /// position, and may either delete it or move it to outside of the loop.
1427 static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
1428 const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo,
1429 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE,
1431 LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
1433 return OptimizationRemark(DEBUG_TYPE, "InstSunk", &I)
1434 << "sinking " << ore::NV("Inst", &I);
1436 bool Changed = false;
1437 if (isa<LoadInst>(I))
1439 else if (isa<CallInst>(I))
1443 // Iterate over users to be ready for actual sinking. Replace users via
1444 // unrechable blocks with undef and make all user PHIs trivially replcable.
1445 SmallPtrSet<Instruction *, 8> VisitedUsers;
1446 for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) {
1447 auto *User = cast<Instruction>(*UI);
1448 Use &U = UI.getUse();
1451 if (VisitedUsers.count(User) || CurLoop->contains(User))
1454 if (!DT->isReachableFromEntry(User->getParent())) {
1455 U = UndefValue::get(I.getType());
1460 // The user must be a PHI node.
1461 PHINode *PN = cast<PHINode>(User);
1463 // Surprisingly, instructions can be used outside of loops without any
1464 // exits. This can only happen in PHI nodes if the incoming block is
1466 BasicBlock *BB = PN->getIncomingBlock(U);
1467 if (!DT->isReachableFromEntry(BB)) {
1468 U = UndefValue::get(I.getType());
1473 VisitedUsers.insert(PN);
1474 if (isTriviallyReplaceablePHI(*PN, I))
1477 if (!canSplitPredecessors(PN, SafetyInfo))
1480 // Split predecessors of the PHI so that we can make users trivially
1482 splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU);
1484 // Should rebuild the iterators, as they may be invalidated by
1485 // splitPredecessorsOfLoopExit().
1486 UI = I.user_begin();
1490 if (VisitedUsers.empty())
1494 SmallVector<BasicBlock *, 32> ExitBlocks;
1495 CurLoop->getUniqueExitBlocks(ExitBlocks);
1496 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1500 // Clones of this instruction. Don't create more than one per exit block!
1501 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
1503 // If this instruction is only used outside of the loop, then all users are
1504 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
1506 SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
1507 for (auto *UI : Users) {
1508 auto *User = cast<Instruction>(UI);
1510 if (CurLoop->contains(User))
1513 PHINode *PN = cast<PHINode>(User);
1514 assert(ExitBlockSet.count(PN->getParent()) &&
1515 "The LCSSA PHI is not in an exit block!");
1516 // The PHI must be trivially replaceable.
1517 Instruction *New = sinkThroughTriviallyReplaceablePHI(
1518 PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU);
1519 PN->replaceAllUsesWith(New);
1520 eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr);
1526 /// When an instruction is found to only use loop invariant operands that
1527 /// is safe to hoist, this instruction is called to do the dirty work.
1529 static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
1530 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
1531 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE) {
1532 LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getName() << ": " << I
1535 return OptimizationRemark(DEBUG_TYPE, "Hoisted", &I) << "hoisting "
1536 << ore::NV("Inst", &I);
1539 // Metadata can be dependent on conditions we are hoisting above.
1540 // Conservatively strip all metadata on the instruction unless we were
1541 // guaranteed to execute I if we entered the loop, in which case the metadata
1542 // is valid in the loop preheader.
1543 if (I.hasMetadataOtherThanDebugLoc() &&
1544 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
1545 // time in isGuaranteedToExecute if we don't actually have anything to
1546 // drop. It is a compile time optimization, not required for correctness.
1547 !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop))
1548 I.dropUnknownNonDebugMetadata();
1550 if (isa<PHINode>(I))
1551 // Move the new node to the end of the phi list in the destination block.
1552 moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo);
1554 // Move the new node to the destination block, before its terminator.
1555 moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo);
1557 // If moving, I just moved a load or store, so update MemorySSA.
1558 MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
1559 MSSAU->getMemorySSA()->getMemoryAccess(&I));
1561 MSSAU->moveToPlace(OldMemAcc, Dest, MemorySSA::End);
1564 // Do not retain debug locations when we are moving instructions to different
1565 // basic blocks, because we want to avoid jumpy line tables. Calls, however,
1566 // need to retain their debug locs because they may be inlined.
1567 // FIXME: How do we retain source locations without causing poor debugging
1569 if (!isa<CallInst>(I))
1570 I.setDebugLoc(DebugLoc());
1572 if (isa<LoadInst>(I))
1574 else if (isa<CallInst>(I))
1579 /// Only sink or hoist an instruction if it is not a trapping instruction,
1580 /// or if the instruction is known not to trap when moved to the preheader.
1581 /// or if it is a trapping instruction and is guaranteed to execute.
1582 static bool isSafeToExecuteUnconditionally(Instruction &Inst,
1583 const DominatorTree *DT,
1584 const Loop *CurLoop,
1585 const LoopSafetyInfo *SafetyInfo,
1586 OptimizationRemarkEmitter *ORE,
1587 const Instruction *CtxI) {
1588 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
1591 bool GuaranteedToExecute =
1592 SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop);
1594 if (!GuaranteedToExecute) {
1595 auto *LI = dyn_cast<LoadInst>(&Inst);
1596 if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1598 return OptimizationRemarkMissed(
1599 DEBUG_TYPE, "LoadWithLoopInvariantAddressCondExecuted", LI)
1600 << "failed to hoist load with loop-invariant address "
1601 "because load is conditionally executed";
1605 return GuaranteedToExecute;
1609 class LoopPromoter : public LoadAndStorePromoter {
1610 Value *SomePtr; // Designated pointer to store to.
1611 const SmallSetVector<Value *, 8> &PointerMustAliases;
1612 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
1613 SmallVectorImpl<Instruction *> &LoopInsertPts;
1614 PredIteratorCache &PredCache;
1615 AliasSetTracker &AST;
1619 bool UnorderedAtomic;
1621 ICFLoopSafetyInfo &SafetyInfo;
1623 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
1624 if (Instruction *I = dyn_cast<Instruction>(V))
1625 if (Loop *L = LI.getLoopFor(I->getParent()))
1626 if (!L->contains(BB)) {
1627 // We need to create an LCSSA PHI node for the incoming value and
1629 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
1630 I->getName() + ".lcssa", &BB->front());
1631 for (BasicBlock *Pred : PredCache.get(BB))
1632 PN->addIncoming(I, Pred);
1639 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
1640 const SmallSetVector<Value *, 8> &PMA,
1641 SmallVectorImpl<BasicBlock *> &LEB,
1642 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
1643 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
1644 bool UnorderedAtomic, const AAMDNodes &AATags,
1645 ICFLoopSafetyInfo &SafetyInfo)
1646 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
1647 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
1648 LI(li), DL(std::move(dl)), Alignment(alignment),
1649 UnorderedAtomic(UnorderedAtomic), AATags(AATags), SafetyInfo(SafetyInfo)
1652 bool isInstInList(Instruction *I,
1653 const SmallVectorImpl<Instruction *> &) const override {
1655 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1656 Ptr = LI->getOperand(0);
1658 Ptr = cast<StoreInst>(I)->getPointerOperand();
1659 return PointerMustAliases.count(Ptr);
1662 void doExtraRewritesBeforeFinalDeletion() const override {
1663 // Insert stores after in the loop exit blocks. Each exit block gets a
1664 // store of the live-out values that feed them. Since we've already told
1665 // the SSA updater about the defs in the loop and the preheader
1666 // definition, it is all set and we can start using it.
1667 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
1668 BasicBlock *ExitBlock = LoopExitBlocks[i];
1669 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
1670 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
1671 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
1672 Instruction *InsertPos = LoopInsertPts[i];
1673 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
1674 if (UnorderedAtomic)
1675 NewSI->setOrdering(AtomicOrdering::Unordered);
1676 NewSI->setAlignment(Alignment);
1677 NewSI->setDebugLoc(DL);
1679 NewSI->setAAMetadata(AATags);
1683 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
1684 // Update alias analysis.
1685 AST.copyValue(LI, V);
1687 void instructionDeleted(Instruction *I) const override {
1688 SafetyInfo.removeInstruction(I);
1694 /// Return true iff we can prove that a caller of this function can not inspect
1695 /// the contents of the provided object in a well defined program.
1696 bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) {
1697 if (isa<AllocaInst>(Object))
1698 // Since the alloca goes out of scope, we know the caller can't retain a
1699 // reference to it and be well defined. Thus, we don't need to check for
1703 // For all other objects we need to know that the caller can't possibly
1704 // have gotten a reference to the object. There are two components of
1706 // 1) Object can't be escaped by this function. This is what
1707 // PointerMayBeCaptured checks.
1708 // 2) Object can't have been captured at definition site. For this, we
1709 // need to know the return value is noalias. At the moment, we use a
1710 // weaker condition and handle only AllocLikeFunctions (which are
1711 // known to be noalias). TODO
1712 return isAllocLikeFn(Object, TLI) &&
1713 !PointerMayBeCaptured(Object, true, true);
1718 /// Try to promote memory values to scalars by sinking stores out of the
1719 /// loop and moving loads to before the loop. We do this by looping over
1720 /// the stores in the loop, looking for stores to Must pointers which are
1723 bool llvm::promoteLoopAccessesToScalars(
1724 const SmallSetVector<Value *, 8> &PointerMustAliases,
1725 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1726 SmallVectorImpl<Instruction *> &InsertPts, PredIteratorCache &PIC,
1727 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
1728 Loop *CurLoop, AliasSetTracker *CurAST, ICFLoopSafetyInfo *SafetyInfo,
1729 OptimizationRemarkEmitter *ORE) {
1731 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
1732 CurAST != nullptr && SafetyInfo != nullptr &&
1733 "Unexpected Input to promoteLoopAccessesToScalars");
1735 Value *SomePtr = *PointerMustAliases.begin();
1736 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1738 // It is not safe to promote a load/store from the loop if the load/store is
1739 // conditional. For example, turning:
1741 // for () { if (c) *P += 1; }
1745 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
1747 // is not safe, because *P may only be valid to access if 'c' is true.
1749 // The safety property divides into two parts:
1750 // p1) The memory may not be dereferenceable on entry to the loop. In this
1751 // case, we can't insert the required load in the preheader.
1752 // p2) The memory model does not allow us to insert a store along any dynamic
1753 // path which did not originally have one.
1755 // If at least one store is guaranteed to execute, both properties are
1756 // satisfied, and promotion is legal.
1758 // This, however, is not a necessary condition. Even if no store/load is
1759 // guaranteed to execute, we can still establish these properties.
1760 // We can establish (p1) by proving that hoisting the load into the preheader
1761 // is safe (i.e. proving dereferenceability on all paths through the loop). We
1762 // can use any access within the alias set to prove dereferenceability,
1763 // since they're all must alias.
1765 // There are two ways establish (p2):
1766 // a) Prove the location is thread-local. In this case the memory model
1767 // requirement does not apply, and stores are safe to insert.
1768 // b) Prove a store dominates every exit block. In this case, if an exit
1769 // blocks is reached, the original dynamic path would have taken us through
1770 // the store, so inserting a store into the exit block is safe. Note that this
1771 // is different from the store being guaranteed to execute. For instance,
1772 // if an exception is thrown on the first iteration of the loop, the original
1773 // store is never executed, but the exit blocks are not executed either.
1775 bool DereferenceableInPH = false;
1776 bool SafeToInsertStore = false;
1778 SmallVector<Instruction *, 64> LoopUses;
1780 // We start with an alignment of one and try to find instructions that allow
1781 // us to prove better alignment.
1782 unsigned Alignment = 1;
1783 // Keep track of which types of access we see
1784 bool SawUnorderedAtomic = false;
1785 bool SawNotAtomic = false;
1788 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
1790 bool IsKnownThreadLocalObject = false;
1791 if (SafetyInfo->anyBlockMayThrow()) {
1792 // If a loop can throw, we have to insert a store along each unwind edge.
1793 // That said, we can't actually make the unwind edge explicit. Therefore,
1794 // we have to prove that the store is dead along the unwind edge. We do
1795 // this by proving that the caller can't have a reference to the object
1796 // after return and thus can't possibly load from the object.
1797 Value *Object = GetUnderlyingObject(SomePtr, MDL);
1798 if (!isKnownNonEscaping(Object, TLI))
1800 // Subtlety: Alloca's aren't visible to callers, but *are* potentially
1801 // visible to other threads if captured and used during their lifetimes.
1802 IsKnownThreadLocalObject = !isa<AllocaInst>(Object);
1805 // Check that all of the pointers in the alias set have the same type. We
1806 // cannot (yet) promote a memory location that is loaded and stored in
1807 // different sizes. While we are at it, collect alignment and AA info.
1808 for (Value *ASIV : PointerMustAliases) {
1809 // Check that all of the pointers in the alias set have the same type. We
1810 // cannot (yet) promote a memory location that is loaded and stored in
1812 if (SomePtr->getType() != ASIV->getType())
1815 for (User *U : ASIV->users()) {
1816 // Ignore instructions that are outside the loop.
1817 Instruction *UI = dyn_cast<Instruction>(U);
1818 if (!UI || !CurLoop->contains(UI))
1821 // If there is an non-load/store instruction in the loop, we can't promote
1823 if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
1824 if (!Load->isUnordered())
1827 SawUnorderedAtomic |= Load->isAtomic();
1828 SawNotAtomic |= !Load->isAtomic();
1830 if (!DereferenceableInPH)
1831 DereferenceableInPH = isSafeToExecuteUnconditionally(
1832 *Load, DT, CurLoop, SafetyInfo, ORE, Preheader->getTerminator());
1833 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
1834 // Stores *of* the pointer are not interesting, only stores *to* the
1836 if (UI->getOperand(1) != ASIV)
1838 if (!Store->isUnordered())
1841 SawUnorderedAtomic |= Store->isAtomic();
1842 SawNotAtomic |= !Store->isAtomic();
1844 // If the store is guaranteed to execute, both properties are satisfied.
1845 // We may want to check if a store is guaranteed to execute even if we
1846 // already know that promotion is safe, since it may have higher
1847 // alignment than any other guaranteed stores, in which case we can
1848 // raise the alignment on the promoted store.
1849 unsigned InstAlignment = Store->getAlignment();
1852 MDL.getABITypeAlignment(Store->getValueOperand()->getType());
1854 if (!DereferenceableInPH || !SafeToInsertStore ||
1855 (InstAlignment > Alignment)) {
1856 if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) {
1857 DereferenceableInPH = true;
1858 SafeToInsertStore = true;
1859 Alignment = std::max(Alignment, InstAlignment);
1863 // If a store dominates all exit blocks, it is safe to sink.
1864 // As explained above, if an exit block was executed, a dominating
1865 // store must have been executed at least once, so we are not
1866 // introducing stores on paths that did not have them.
1867 // Note that this only looks at explicit exit blocks. If we ever
1868 // start sinking stores into unwind edges (see above), this will break.
1869 if (!SafeToInsertStore)
1870 SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
1871 return DT->dominates(Store->getParent(), Exit);
1874 // If the store is not guaranteed to execute, we may still get
1875 // deref info through it.
1876 if (!DereferenceableInPH) {
1877 DereferenceableInPH = isDereferenceableAndAlignedPointer(
1878 Store->getPointerOperand(), Store->getAlignment(), MDL,
1879 Preheader->getTerminator(), DT);
1882 return false; // Not a load or store.
1884 // Merge the AA tags.
1885 if (LoopUses.empty()) {
1886 // On the first load/store, just take its AA tags.
1887 UI->getAAMetadata(AATags);
1888 } else if (AATags) {
1889 UI->getAAMetadata(AATags, /* Merge = */ true);
1892 LoopUses.push_back(UI);
1896 // If we found both an unordered atomic instruction and a non-atomic memory
1897 // access, bail. We can't blindly promote non-atomic to atomic since we
1898 // might not be able to lower the result. We can't downgrade since that
1899 // would violate memory model. Also, align 0 is an error for atomics.
1900 if (SawUnorderedAtomic && SawNotAtomic)
1903 // If we couldn't prove we can hoist the load, bail.
1904 if (!DereferenceableInPH)
1907 // We know we can hoist the load, but don't have a guaranteed store.
1908 // Check whether the location is thread-local. If it is, then we can insert
1909 // stores along paths which originally didn't have them without violating the
1911 if (!SafeToInsertStore) {
1912 if (IsKnownThreadLocalObject)
1913 SafeToInsertStore = true;
1915 Value *Object = GetUnderlyingObject(SomePtr, MDL);
1917 (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
1918 !PointerMayBeCaptured(Object, true, true);
1922 // If we've still failed to prove we can sink the store, give up.
1923 if (!SafeToInsertStore)
1926 // Otherwise, this is safe to promote, lets do it!
1927 LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtr
1930 return OptimizationRemark(DEBUG_TYPE, "PromoteLoopAccessesToScalar",
1932 << "Moving accesses to memory location out of the loop";
1936 // Grab a debug location for the inserted loads/stores; given that the
1937 // inserted loads/stores have little relation to the original loads/stores,
1938 // this code just arbitrarily picks a location from one, since any debug
1939 // location is better than none.
1940 DebugLoc DL = LoopUses[0]->getDebugLoc();
1942 // We use the SSAUpdater interface to insert phi nodes as required.
1943 SmallVector<PHINode *, 16> NewPHIs;
1944 SSAUpdater SSA(&NewPHIs);
1945 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
1946 InsertPts, PIC, *CurAST, *LI, DL, Alignment,
1947 SawUnorderedAtomic, AATags, *SafetyInfo);
1949 // Set up the preheader to have a definition of the value. It is the live-out
1950 // value from the preheader that uses in the loop will use.
1951 LoadInst *PreheaderLoad = new LoadInst(
1952 SomePtr, SomePtr->getName() + ".promoted", Preheader->getTerminator());
1953 if (SawUnorderedAtomic)
1954 PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
1955 PreheaderLoad->setAlignment(Alignment);
1956 PreheaderLoad->setDebugLoc(DL);
1958 PreheaderLoad->setAAMetadata(AATags);
1959 SSA.AddAvailableValue(Preheader, PreheaderLoad);
1961 // Rewrite all the loads in the loop and remember all the definitions from
1962 // stores in the loop.
1963 Promoter.run(LoopUses);
1965 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
1966 if (PreheaderLoad->use_empty())
1967 eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, nullptr);
1972 /// Returns an owning pointer to an alias set which incorporates aliasing info
1973 /// from L and all subloops of L.
1974 /// FIXME: In new pass manager, there is no helper function to handle loop
1975 /// analysis such as cloneBasicBlockAnalysis, so the AST needs to be recomputed
1976 /// from scratch for every loop. Hook up with the helper functions when
1977 /// available in the new pass manager to avoid redundant computation.
1978 std::unique_ptr<AliasSetTracker>
1979 LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
1980 AliasAnalysis *AA) {
1981 std::unique_ptr<AliasSetTracker> CurAST;
1982 SmallVector<Loop *, 4> RecomputeLoops;
1983 for (Loop *InnerL : L->getSubLoops()) {
1984 auto MapI = LoopToAliasSetMap.find(InnerL);
1985 // If the AST for this inner loop is missing it may have been merged into
1986 // some other loop's AST and then that loop unrolled, and so we need to
1988 if (MapI == LoopToAliasSetMap.end()) {
1989 RecomputeLoops.push_back(InnerL);
1992 std::unique_ptr<AliasSetTracker> InnerAST = std::move(MapI->second);
1995 // What if InnerLoop was modified by other passes ?
1996 // Once we've incorporated the inner loop's AST into ours, we don't need
1997 // the subloop's anymore.
1998 CurAST->add(*InnerAST);
2000 CurAST = std::move(InnerAST);
2002 LoopToAliasSetMap.erase(MapI);
2005 CurAST = make_unique<AliasSetTracker>(*AA);
2007 // Add everything from the sub loops that are no longer directly available.
2008 for (Loop *InnerL : RecomputeLoops)
2009 for (BasicBlock *BB : InnerL->blocks())
2012 // And merge in this loop (without anything from inner loops).
2013 for (BasicBlock *BB : L->blocks())
2014 if (LI->getLoopFor(BB) == L)
2020 /// Simple analysis hook. Clone alias set info.
2022 void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
2024 auto ASTIt = LICM.getLoopToAliasSetMap().find(L);
2025 if (ASTIt == LICM.getLoopToAliasSetMap().end())
2028 ASTIt->second->copyValue(From, To);
2031 /// Simple Analysis hook. Delete value V from alias set
2033 void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) {
2034 auto ASTIt = LICM.getLoopToAliasSetMap().find(L);
2035 if (ASTIt == LICM.getLoopToAliasSetMap().end())
2038 ASTIt->second->deleteValue(V);
2041 /// Simple Analysis hook. Delete value L from alias set map.
2043 void LegacyLICMPass::deleteAnalysisLoop(Loop *L) {
2044 if (!LICM.getLoopToAliasSetMap().count(L))
2047 LICM.getLoopToAliasSetMap().erase(L);
2050 static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
2051 AliasSetTracker *CurAST, Loop *CurLoop,
2052 AliasAnalysis *AA) {
2053 // First check to see if any of the basic blocks in CurLoop invalidate *V.
2054 bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod();
2056 if (!isInvalidatedAccordingToAST || !LICMN2Theshold)
2057 return isInvalidatedAccordingToAST;
2059 // Check with a diagnostic analysis if we can refine the information above.
2060 // This is to identify the limitations of using the AST.
2061 // The alias set mechanism used by LICM has a major weakness in that it
2062 // combines all things which may alias into a single set *before* asking
2063 // modref questions. As a result, a single readonly call within a loop will
2064 // collapse all loads and stores into a single alias set and report
2065 // invalidation if the loop contains any store. For example, readonly calls
2066 // with deopt states have this form and create a general alias set with all
2067 // loads and stores. In order to get any LICM in loops containing possible
2068 // deopt states we need a more precise invalidation of checking the mod ref
2069 // info of each instruction within the loop and LI. This has a complexity of
2070 // O(N^2), so currently, it is used only as a diagnostic tool since the
2071 // default value of LICMN2Threshold is zero.
2073 // Don't look at nested loops.
2074 if (CurLoop->begin() != CurLoop->end())
2078 for (BasicBlock *BB : CurLoop->getBlocks())
2079 for (Instruction &I : *BB) {
2080 if (N >= LICMN2Theshold) {
2081 LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "
2082 << *(MemLoc.Ptr) << "\n");
2086 auto Res = AA->getModRefInfo(&I, MemLoc);
2087 if (isModSet(Res)) {
2088 LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "
2089 << *(MemLoc.Ptr) << "\n");
2093 LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n");
2097 static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
2099 MemoryAccess *Source;
2100 // See declaration of EnableLicmCap for usage details.
2102 Source = MU->getDefiningAccess();
2104 Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU);
2105 return !MSSA->isLiveOnEntryDef(Source) &&
2106 CurLoop->contains(Source->getBlock());
2109 /// Little predicate that returns true if the specified basic block is in
2110 /// a subloop of the current one, not the current one itself.
2112 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
2113 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
2114 return LI->getLoopFor(BB) != CurLoop;