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.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/ConstantFolding.h"
39 #include "llvm/Analysis/GlobalsModRef.h"
40 #include "llvm/Analysis/LoopInfo.h"
41 #include "llvm/Analysis/LoopPass.h"
42 #include "llvm/Analysis/ScalarEvolution.h"
43 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
44 #include "llvm/Analysis/TargetLibraryInfo.h"
45 #include "llvm/Analysis/ValueTracking.h"
46 #include "llvm/IR/CFG.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DataLayout.h"
49 #include "llvm/IR/DerivedTypes.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Instructions.h"
52 #include "llvm/IR/IntrinsicInst.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/IR/Metadata.h"
55 #include "llvm/IR/PredIteratorCache.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include "llvm/Transforms/Utils/LoopUtils.h"
61 #include "llvm/Transforms/Utils/SSAUpdater.h"
65 #define DEBUG_TYPE "licm"
67 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
68 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
69 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
70 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
71 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
74 DisablePromotion("disable-licm-promotion", cl::Hidden,
75 cl::desc("Disable memory promotion in LICM pass"));
77 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
78 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
79 const LICMSafetyInfo *SafetyInfo);
80 static bool hoist(Instruction &I, BasicBlock *Preheader);
81 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
82 const Loop *CurLoop, AliasSetTracker *CurAST,
83 const LICMSafetyInfo *SafetyInfo);
84 static bool isGuaranteedToExecute(const Instruction &Inst,
85 const DominatorTree *DT,
87 const LICMSafetyInfo *SafetyInfo);
88 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
89 const DominatorTree *DT,
90 const TargetLibraryInfo *TLI,
92 const LICMSafetyInfo *SafetyInfo,
93 const Instruction *CtxI = nullptr);
94 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
95 const AAMDNodes &AAInfo,
96 AliasSetTracker *CurAST);
98 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
100 const LICMSafetyInfo *SafetyInfo);
101 static bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA,
102 DominatorTree *DT, TargetLibraryInfo *TLI,
103 Loop *CurLoop, AliasSetTracker *CurAST,
104 LICMSafetyInfo *SafetyInfo);
107 struct LICM : public LoopPass {
108 static char ID; // Pass identification, replacement for typeid
109 LICM() : LoopPass(ID) {
110 initializeLICMPass(*PassRegistry::getPassRegistry());
113 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
115 /// This transformation requires natural loop information & requires that
116 /// loop preheaders be inserted into the CFG...
118 void getAnalysisUsage(AnalysisUsage &AU) const override {
119 AU.setPreservesCFG();
120 AU.addRequired<DominatorTreeWrapperPass>();
121 AU.addRequired<LoopInfoWrapperPass>();
122 AU.addRequiredID(LoopSimplifyID);
123 AU.addPreservedID(LoopSimplifyID);
124 AU.addRequiredID(LCSSAID);
125 AU.addPreservedID(LCSSAID);
126 AU.addRequired<AAResultsWrapperPass>();
127 AU.addPreserved<AAResultsWrapperPass>();
128 AU.addPreserved<BasicAAWrapperPass>();
129 AU.addPreserved<GlobalsAAWrapperPass>();
130 AU.addPreserved<ScalarEvolutionWrapperPass>();
131 AU.addPreserved<SCEVAAWrapperPass>();
132 AU.addRequired<TargetLibraryInfoWrapperPass>();
135 using llvm::Pass::doFinalization;
137 bool doFinalization() override {
138 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
143 AliasAnalysis *AA; // Current AliasAnalysis information
144 LoopInfo *LI; // Current LoopInfo
145 DominatorTree *DT; // Dominator Tree for the current Loop.
147 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
149 // State that is updated as we process loops.
150 bool Changed; // Set to true when we change anything.
151 BasicBlock *Preheader; // The preheader block of the current loop...
152 Loop *CurLoop; // The current loop we are working on...
153 AliasSetTracker *CurAST; // AliasSet information for the current loop...
154 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
156 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
157 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
160 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
162 void deleteAnalysisValue(Value *V, Loop *L) override;
164 /// Simple Analysis hook. Delete loop L from alias set map.
165 void deleteAnalysisLoop(Loop *L) override;
170 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
171 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
172 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
173 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
174 INITIALIZE_PASS_DEPENDENCY(LCSSA)
175 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
176 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
177 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
178 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
179 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
180 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
181 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
183 Pass *llvm::createLICMPass() { return new LICM(); }
185 /// Hoist expressions out of the specified loop. Note, alias info for inner
186 /// loop is not preserved so it is not a good idea to run LICM multiple
187 /// times on one loop.
189 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
190 if (skipOptnoneFunction(L))
195 // Get our Loop and Alias Analysis information...
196 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
197 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
198 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
200 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
202 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
204 CurAST = new AliasSetTracker(*AA);
205 // Collect Alias info from subloops.
206 for (Loop *InnerL : L->getSubLoops()) {
207 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
208 assert(InnerAST && "Where is my AST?");
210 // What if InnerLoop was modified by other passes ?
211 CurAST->add(*InnerAST);
213 // Once we've incorporated the inner loop's AST into ours, we don't need the
214 // subloop's anymore.
216 LoopToAliasSetMap.erase(InnerL);
221 // Get the preheader block to move instructions into...
222 Preheader = L->getLoopPreheader();
224 // Loop over the body of this loop, looking for calls, invokes, and stores.
225 // Because subloops have already been incorporated into AST, we skip blocks in
228 for (BasicBlock *BB : L->blocks()) {
229 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
230 CurAST->add(*BB); // Incorporate the specified basic block
233 // Compute loop safety information.
234 LICMSafetyInfo SafetyInfo;
235 computeLICMSafetyInfo(&SafetyInfo, CurLoop);
237 // We want to visit all of the instructions in this loop... that are not parts
238 // of our subloops (they have already had their invariants hoisted out of
239 // their loop, into this loop, so there is no need to process the BODIES of
242 // Traverse the body of the loop in depth first order on the dominator tree so
243 // that we are guaranteed to see definitions before we see uses. This allows
244 // us to sink instructions in one pass, without iteration. After sinking
245 // instructions, we perform another pass to hoist them out of the loop.
247 if (L->hasDedicatedExits())
248 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, CurLoop,
249 CurAST, &SafetyInfo);
251 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI,
252 CurLoop, CurAST, &SafetyInfo);
254 // Now that all loop invariants have been removed from the loop, promote any
255 // memory references to scalars that we can.
256 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
257 SmallVector<BasicBlock *, 8> ExitBlocks;
258 SmallVector<Instruction *, 8> InsertPts;
259 PredIteratorCache PIC;
261 // Loop over all of the alias sets in the tracker object.
262 for (AliasSet &AS : *CurAST)
263 Changed |= promoteLoopAccessesToScalars(AS, ExitBlocks, InsertPts,
264 PIC, LI, DT, CurLoop,
265 CurAST, &SafetyInfo);
267 // Once we have promoted values across the loop body we have to recursively
268 // reform LCSSA as any nested loop may now have values defined within the
269 // loop used in the outer loop.
270 // FIXME: This is really heavy handed. It would be a bit better to use an
271 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
274 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
275 formLCSSARecursively(*L, *DT, LI, SEWP ? &SEWP->getSE() : nullptr);
279 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
280 // specifically moving instructions across the loop boundary and so it is
281 // especially in need of sanity checking here.
282 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
283 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
284 "Parent loop not left in LCSSA form after LICM!");
286 // Clear out loops state information for the next iteration
290 // If this loop is nested inside of another one, save the alias information
291 // for when we process the outer loop.
292 if (L->getParentLoop())
293 LoopToAliasSetMap[L] = CurAST;
299 /// Walk the specified region of the CFG (defined by all blocks dominated by
300 /// the specified block, and that are in the current loop) in reverse depth
301 /// first order w.r.t the DominatorTree. This allows us to visit uses before
302 /// definitions, allowing us to sink a loop body in one pass without iteration.
304 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
305 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
306 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
309 assert(N != nullptr && AA != nullptr && LI != nullptr &&
310 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
311 SafetyInfo != nullptr && "Unexpected input to sinkRegion");
313 // Set changed as false.
314 bool Changed = false;
316 BasicBlock *BB = N->getBlock();
317 // If this subregion is not in the top level loop at all, exit.
318 if (!CurLoop->contains(BB)) return Changed;
320 // We are processing blocks in reverse dfo, so process children first.
321 const std::vector<DomTreeNode*> &Children = N->getChildren();
322 for (DomTreeNode *Child : Children)
323 Changed |= sinkRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
325 // Only need to process the contents of this block if it is not part of a
326 // subloop (which would already have been processed).
327 if (inSubLoop(BB,CurLoop,LI)) return Changed;
329 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
330 Instruction &I = *--II;
332 // If the instruction is dead, we would try to sink it because it isn't used
333 // in the loop, instead, just delete it.
334 if (isInstructionTriviallyDead(&I, TLI)) {
335 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
337 CurAST->deleteValue(&I);
343 // Check to see if we can sink this instruction to the exit blocks
344 // of the loop. We can do this if the all users of the instruction are
345 // outside of the loop. In this case, it doesn't even matter if the
346 // operands of the instruction are loop invariant.
348 if (isNotUsedInLoop(I, CurLoop, SafetyInfo) &&
349 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo)) {
351 Changed |= sink(I, LI, DT, CurLoop, CurAST, SafetyInfo);
357 /// Walk the specified region of the CFG (defined by all blocks dominated by
358 /// the specified block, and that are in the current loop) in depth first
359 /// order w.r.t the DominatorTree. This allows us to visit definitions before
360 /// uses, allowing us to hoist a loop body in one pass without iteration.
362 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
363 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
364 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
366 assert(N != nullptr && AA != nullptr && LI != nullptr &&
367 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
368 SafetyInfo != nullptr && "Unexpected input to hoistRegion");
369 // Set changed as false.
370 bool Changed = false;
372 BasicBlock *BB = N->getBlock();
373 // If this subregion is not in the top level loop at all, exit.
374 if (!CurLoop->contains(BB)) return Changed;
375 // Only need to process the contents of this block if it is not part of a
376 // subloop (which would already have been processed).
377 if (!inSubLoop(BB, CurLoop, LI))
378 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
379 Instruction &I = *II++;
380 // Try constant folding this instruction. If all the operands are
381 // constants, it is technically hoistable, but it would be better to just
383 if (Constant *C = ConstantFoldInstruction(
384 &I, I.getModule()->getDataLayout(), TLI)) {
385 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
386 CurAST->copyValue(&I, C);
387 CurAST->deleteValue(&I);
388 I.replaceAllUsesWith(C);
393 // Try hoisting the instruction out to the preheader. We can only do this
394 // if all of the operands of the instruction are loop invariant and if it
395 // is safe to hoist the instruction.
397 if (CurLoop->hasLoopInvariantOperands(&I) &&
398 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo) &&
399 isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
400 CurLoop->getLoopPreheader()->getTerminator()))
401 Changed |= hoist(I, CurLoop->getLoopPreheader());
404 const std::vector<DomTreeNode*> &Children = N->getChildren();
405 for (DomTreeNode *Child : Children)
406 Changed |= hoistRegion(Child, AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
410 /// Computes loop safety information, checks loop body & header
411 /// for the possibility of may throw exception.
413 void llvm::computeLICMSafetyInfo(LICMSafetyInfo * SafetyInfo, Loop * CurLoop) {
414 assert(CurLoop != nullptr && "CurLoop cant be null");
415 BasicBlock *Header = CurLoop->getHeader();
416 // Setting default safety values.
417 SafetyInfo->MayThrow = false;
418 SafetyInfo->HeaderMayThrow = false;
419 // Iterate over header and compute safety info.
420 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
421 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
422 SafetyInfo->HeaderMayThrow |= I->mayThrow();
424 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
425 // Iterate over loop instructions and compute safety info.
426 for (Loop::block_iterator BB = CurLoop->block_begin(),
427 BBE = CurLoop->block_end(); (BB != BBE) && !SafetyInfo->MayThrow ; ++BB)
428 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
429 (I != E) && !SafetyInfo->MayThrow; ++I)
430 SafetyInfo->MayThrow |= I->mayThrow();
432 // Compute funclet colors if we might sink/hoist in a function with a funclet
433 // personality routine.
434 Function *Fn = CurLoop->getHeader()->getParent();
435 if (Fn->hasPersonalityFn())
436 if (Constant *PersonalityFn = Fn->getPersonalityFn())
437 if (isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)))
438 SafetyInfo->BlockColors = colorEHFunclets(*Fn);
441 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
444 bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA, DominatorTree *DT,
445 TargetLibraryInfo *TLI, Loop *CurLoop,
446 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
447 // Loads have extra constraints we have to verify before we can hoist them.
448 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
449 if (!LI->isUnordered())
450 return false; // Don't hoist volatile/atomic loads!
452 // Loads from constant memory are always safe to move, even if they end up
453 // in the same alias set as something that ends up being modified.
454 if (AA->pointsToConstantMemory(LI->getOperand(0)))
456 if (LI->getMetadata(LLVMContext::MD_invariant_load))
459 // Don't hoist loads which have may-aliased stores in loop.
461 if (LI->getType()->isSized())
462 Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType());
465 LI->getAAMetadata(AAInfo);
467 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
468 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
469 // Don't sink or hoist dbg info; it's legal, but not useful.
470 if (isa<DbgInfoIntrinsic>(I))
473 // Don't sink calls which can throw.
477 // Handle simple cases by querying alias analysis.
478 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
479 if (Behavior == FMRB_DoesNotAccessMemory)
481 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
482 // A readonly argmemonly function only reads from memory pointed to by
483 // it's arguments with arbitrary offsets. If we can prove there are no
484 // writes to this memory in the loop, we can hoist or sink.
485 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
486 for (Value *Op : CI->arg_operands())
487 if (Op->getType()->isPointerTy() &&
488 pointerInvalidatedByLoop(Op, MemoryLocation::UnknownSize,
489 AAMDNodes(), CurAST))
493 // If this call only reads from memory and there are no writes to memory
494 // in the loop, we can hoist or sink the call as appropriate.
495 bool FoundMod = false;
496 for (AliasSet &AS : *CurAST) {
497 if (!AS.isForwardingAliasSet() && AS.isMod()) {
502 if (!FoundMod) return true;
505 // FIXME: This should use mod/ref information to see if we can hoist or
511 // Only these instructions are hoistable/sinkable.
512 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
513 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
514 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
515 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
516 !isa<InsertValueInst>(I))
519 // TODO: Plumb the context instruction through to make hoisting and sinking
520 // more powerful. Hoisting of loads already works due to the special casing
522 return isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
526 /// Returns true if a PHINode is a trivially replaceable with an
528 /// This is true when all incoming values are that instruction.
529 /// This pattern occurs most often with LCSSA PHI nodes.
531 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
532 for (const Value *IncValue : PN.incoming_values())
539 /// Return true if the only users of this instruction are outside of
540 /// the loop. If this is true, we can sink the instruction to the exit
541 /// blocks of the loop.
543 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
544 const LICMSafetyInfo *SafetyInfo) {
545 const auto &BlockColors = SafetyInfo->BlockColors;
546 for (const User *U : I.users()) {
547 const Instruction *UI = cast<Instruction>(U);
548 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
549 const BasicBlock *BB = PN->getParent();
550 // We cannot sink uses in catchswitches.
551 if (isa<CatchSwitchInst>(BB->getTerminator()))
554 // We need to sink a callsite to a unique funclet. Avoid sinking if the
555 // phi use is too muddled.
556 if (isa<CallInst>(I))
557 if (!BlockColors.empty() &&
558 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
561 // A PHI node where all of the incoming values are this instruction are
562 // special -- they can just be RAUW'ed with the instruction and thus
563 // don't require a use in the predecessor. This is a particular important
564 // special case because it is the pattern found in LCSSA form.
565 if (isTriviallyReplacablePHI(*PN, I)) {
566 if (CurLoop->contains(PN))
572 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
573 // values. Check for such a use being inside the loop.
574 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
575 if (PN->getIncomingValue(i) == &I)
576 if (CurLoop->contains(PN->getIncomingBlock(i)))
582 if (CurLoop->contains(UI))
589 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
591 const LICMSafetyInfo *SafetyInfo) {
593 if (auto *CI = dyn_cast<CallInst>(&I)) {
594 const auto &BlockColors = SafetyInfo->BlockColors;
596 // Sinking call-sites need to be handled differently from other
597 // instructions. The cloned call-site needs a funclet bundle operand
598 // appropriate for it's location in the CFG.
599 SmallVector<OperandBundleDef, 1> OpBundles;
600 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
601 BundleIdx != BundleEnd; ++BundleIdx) {
602 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
603 if (Bundle.getTagID() == LLVMContext::OB_funclet)
606 OpBundles.emplace_back(Bundle);
609 if (!BlockColors.empty()) {
610 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
611 assert(CV.size() == 1 && "non-unique color for exit block!");
612 BasicBlock *BBColor = CV.front();
613 Instruction *EHPad = BBColor->getFirstNonPHI();
614 if (EHPad->isEHPad())
615 OpBundles.emplace_back("funclet", EHPad);
618 New = CallInst::Create(CI, OpBundles);
623 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
624 if (!I.getName().empty()) New->setName(I.getName() + ".le");
626 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
627 // particularly cheap because we can rip off the PHI node that we're
628 // replacing for the number and blocks of the predecessors.
629 // OPT: If this shows up in a profile, we can instead finish sinking all
630 // invariant instructions, and then walk their operands to re-establish
631 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
632 // sinking bottom-up.
633 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
635 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
636 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
637 if (!OLoop->contains(&PN)) {
639 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
640 OInst->getName() + ".lcssa", &ExitBlock.front());
641 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
642 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
648 /// When an instruction is found to only be used outside of the loop, this
649 /// function moves it to the exit blocks and patches up SSA form as needed.
650 /// This method is guaranteed to remove the original instruction from its
651 /// position, and may either delete it or move it to outside of the loop.
653 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
654 const Loop *CurLoop, AliasSetTracker *CurAST,
655 const LICMSafetyInfo *SafetyInfo) {
656 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
657 bool Changed = false;
658 if (isa<LoadInst>(I)) ++NumMovedLoads;
659 else if (isa<CallInst>(I)) ++NumMovedCalls;
664 SmallVector<BasicBlock *, 32> ExitBlocks;
665 CurLoop->getUniqueExitBlocks(ExitBlocks);
666 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
670 // Clones of this instruction. Don't create more than one per exit block!
671 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
673 // If this instruction is only used outside of the loop, then all users are
674 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
676 while (!I.use_empty()) {
677 Value::user_iterator UI = I.user_begin();
678 auto *User = cast<Instruction>(*UI);
679 if (!DT->isReachableFromEntry(User->getParent())) {
680 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
683 // The user must be a PHI node.
684 PHINode *PN = cast<PHINode>(User);
686 // Surprisingly, instructions can be used outside of loops without any
687 // exits. This can only happen in PHI nodes if the incoming block is
689 Use &U = UI.getUse();
690 BasicBlock *BB = PN->getIncomingBlock(U);
691 if (!DT->isReachableFromEntry(BB)) {
692 U = UndefValue::get(I.getType());
696 BasicBlock *ExitBlock = PN->getParent();
697 assert(ExitBlockSet.count(ExitBlock) &&
698 "The LCSSA PHI is not in an exit block!");
701 auto It = SunkCopies.find(ExitBlock);
702 if (It != SunkCopies.end())
705 New = SunkCopies[ExitBlock] =
706 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI, SafetyInfo);
708 PN->replaceAllUsesWith(New);
709 PN->eraseFromParent();
712 CurAST->deleteValue(&I);
717 /// When an instruction is found to only use loop invariant operands that
718 /// is safe to hoist, this instruction is called to do the dirty work.
720 static bool hoist(Instruction &I, BasicBlock *Preheader) {
721 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
723 // Move the new node to the Preheader, before its terminator.
724 I.moveBefore(Preheader->getTerminator());
726 // Metadata can be dependent on the condition we are hoisting above.
727 // Conservatively strip all metadata on the instruction.
728 I.dropUnknownNonDebugMetadata();
730 if (isa<LoadInst>(I)) ++NumMovedLoads;
731 else if (isa<CallInst>(I)) ++NumMovedCalls;
736 /// Only sink or hoist an instruction if it is not a trapping instruction,
737 /// or if the instruction is known not to trap when moved to the preheader.
738 /// or if it is a trapping instruction and is guaranteed to execute.
739 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
740 const DominatorTree *DT,
741 const TargetLibraryInfo *TLI,
743 const LICMSafetyInfo *SafetyInfo,
744 const Instruction *CtxI) {
745 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI))
748 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
751 static bool isGuaranteedToExecute(const Instruction &Inst,
752 const DominatorTree *DT,
754 const LICMSafetyInfo * SafetyInfo) {
756 // We have to check to make sure that the instruction dominates all
757 // of the exit blocks. If it doesn't, then there is a path out of the loop
758 // which does not execute this instruction, so we can't hoist it.
760 // If the instruction is in the header block for the loop (which is very
761 // common), it is always guaranteed to dominate the exit blocks. Since this
762 // is a common case, and can save some work, check it now.
763 if (Inst.getParent() == CurLoop->getHeader())
764 // If there's a throw in the header block, we can't guarantee we'll reach
766 return !SafetyInfo->HeaderMayThrow;
768 // Somewhere in this loop there is an instruction which may throw and make us
770 if (SafetyInfo->MayThrow)
773 // Get the exit blocks for the current loop.
774 SmallVector<BasicBlock*, 8> ExitBlocks;
775 CurLoop->getExitBlocks(ExitBlocks);
777 // Verify that the block dominates each of the exit blocks of the loop.
778 for (BasicBlock *ExitBlock : ExitBlocks)
779 if (!DT->dominates(Inst.getParent(), ExitBlock))
782 // As a degenerate case, if the loop is statically infinite then we haven't
783 // proven anything since there are no exit blocks.
784 if (ExitBlocks.empty())
791 class LoopPromoter : public LoadAndStorePromoter {
792 Value *SomePtr; // Designated pointer to store to.
793 SmallPtrSetImpl<Value*> &PointerMustAliases;
794 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
795 SmallVectorImpl<Instruction*> &LoopInsertPts;
796 PredIteratorCache &PredCache;
797 AliasSetTracker &AST;
803 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
804 if (Instruction *I = dyn_cast<Instruction>(V))
805 if (Loop *L = LI.getLoopFor(I->getParent()))
806 if (!L->contains(BB)) {
807 // We need to create an LCSSA PHI node for the incoming value and
810 PHINode::Create(I->getType(), PredCache.size(BB),
811 I->getName() + ".lcssa", &BB->front());
812 for (BasicBlock *Pred : PredCache.get(BB))
813 PN->addIncoming(I, Pred);
820 LoopPromoter(Value *SP,
821 ArrayRef<const Instruction *> Insts,
822 SSAUpdater &S, SmallPtrSetImpl<Value *> &PMA,
823 SmallVectorImpl<BasicBlock *> &LEB,
824 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
825 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
826 const AAMDNodes &AATags)
827 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
828 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
829 LI(li), DL(dl), Alignment(alignment), AATags(AATags) {}
831 bool isInstInList(Instruction *I,
832 const SmallVectorImpl<Instruction*> &) const override {
834 if (LoadInst *LI = dyn_cast<LoadInst>(I))
835 Ptr = LI->getOperand(0);
837 Ptr = cast<StoreInst>(I)->getPointerOperand();
838 return PointerMustAliases.count(Ptr);
841 void doExtraRewritesBeforeFinalDeletion() const override {
842 // Insert stores after in the loop exit blocks. Each exit block gets a
843 // store of the live-out values that feed them. Since we've already told
844 // the SSA updater about the defs in the loop and the preheader
845 // definition, it is all set and we can start using it.
846 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
847 BasicBlock *ExitBlock = LoopExitBlocks[i];
848 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
849 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
850 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
851 Instruction *InsertPos = LoopInsertPts[i];
852 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
853 NewSI->setAlignment(Alignment);
854 NewSI->setDebugLoc(DL);
855 if (AATags) NewSI->setAAMetadata(AATags);
859 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
860 // Update alias analysis.
861 AST.copyValue(LI, V);
863 void instructionDeleted(Instruction *I) const override {
867 } // end anon namespace
869 /// Try to promote memory values to scalars by sinking stores out of the
870 /// loop and moving loads to before the loop. We do this by looping over
871 /// the stores in the loop, looking for stores to Must pointers which are
874 bool llvm::promoteLoopAccessesToScalars(AliasSet &AS,
875 SmallVectorImpl<BasicBlock*>&ExitBlocks,
876 SmallVectorImpl<Instruction*>&InsertPts,
877 PredIteratorCache &PIC, LoopInfo *LI,
878 DominatorTree *DT, Loop *CurLoop,
879 AliasSetTracker *CurAST,
880 LICMSafetyInfo * SafetyInfo) {
882 assert(LI != nullptr && DT != nullptr &&
883 CurLoop != nullptr && CurAST != nullptr &&
884 SafetyInfo != nullptr &&
885 "Unexpected Input to promoteLoopAccessesToScalars");
886 // Initially set Changed status to false.
887 bool Changed = false;
888 // We can promote this alias set if it has a store, if it is a "Must" alias
889 // set, if the pointer is loop invariant, and if we are not eliminating any
890 // volatile loads or stores.
891 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
892 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
895 assert(!AS.empty() &&
896 "Must alias set should have at least one pointer element in it!");
898 Value *SomePtr = AS.begin()->getValue();
899 BasicBlock * Preheader = CurLoop->getLoopPreheader();
901 // It isn't safe to promote a load/store from the loop if the load/store is
902 // conditional. For example, turning:
904 // for () { if (c) *P += 1; }
908 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
910 // is not safe, because *P may only be valid to access if 'c' is true.
912 // It is safe to promote P if all uses are direct load/stores and if at
913 // least one is guaranteed to be executed.
914 bool GuaranteedToExecute = false;
916 SmallVector<Instruction*, 64> LoopUses;
917 SmallPtrSet<Value*, 4> PointerMustAliases;
919 // We start with an alignment of one and try to find instructions that allow
920 // us to prove better alignment.
921 unsigned Alignment = 1;
923 bool HasDedicatedExits = CurLoop->hasDedicatedExits();
925 // Check that all of the pointers in the alias set have the same type. We
926 // cannot (yet) promote a memory location that is loaded and stored in
927 // different sizes. While we are at it, collect alignment and AA info.
928 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
929 Value *ASIV = ASI->getValue();
930 PointerMustAliases.insert(ASIV);
932 // Check that all of the pointers in the alias set have the same type. We
933 // cannot (yet) promote a memory location that is loaded and stored in
935 if (SomePtr->getType() != ASIV->getType())
938 for (User *U : ASIV->users()) {
939 // Ignore instructions that are outside the loop.
940 Instruction *UI = dyn_cast<Instruction>(U);
941 if (!UI || !CurLoop->contains(UI))
944 // If there is an non-load/store instruction in the loop, we can't promote
946 if (const LoadInst *Load = dyn_cast<LoadInst>(UI)) {
947 assert(!Load->isVolatile() && "AST broken");
948 if (!Load->isSimple())
950 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
951 // Stores *of* the pointer are not interesting, only stores *to* the
953 if (UI->getOperand(1) != ASIV)
955 assert(!Store->isVolatile() && "AST broken");
956 if (!Store->isSimple())
958 // Don't sink stores from loops without dedicated block exits. Exits
959 // containing indirect branches are not transformed by loop simplify,
960 // make sure we catch that. An additional load may be generated in the
961 // preheader for SSA updater, so also avoid sinking when no preheader
963 if (!HasDedicatedExits || !Preheader)
966 // Note that we only check GuaranteedToExecute inside the store case
967 // so that we do not introduce stores where they did not exist before
968 // (which would break the LLVM concurrency model).
970 // If the alignment of this instruction allows us to specify a more
971 // restrictive (and performant) alignment and if we are sure this
972 // instruction will be executed, update the alignment.
973 // Larger is better, with the exception of 0 being the best alignment.
974 unsigned InstAlignment = Store->getAlignment();
975 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
976 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
977 GuaranteedToExecute = true;
978 Alignment = InstAlignment;
981 if (!GuaranteedToExecute)
982 GuaranteedToExecute = isGuaranteedToExecute(*UI, DT,
983 CurLoop, SafetyInfo);
986 return Changed; // Not a load or store.
988 // Merge the AA tags.
989 if (LoopUses.empty()) {
990 // On the first load/store, just take its AA tags.
991 UI->getAAMetadata(AATags);
993 UI->getAAMetadata(AATags, /* Merge = */ true);
996 LoopUses.push_back(UI);
1000 // If there isn't a guaranteed-to-execute instruction, we can't promote.
1001 if (!GuaranteedToExecute)
1004 // Figure out the loop exits and their insertion points, if this is the
1006 if (ExitBlocks.empty()) {
1007 CurLoop->getUniqueExitBlocks(ExitBlocks);
1009 InsertPts.reserve(ExitBlocks.size());
1010 for (BasicBlock *ExitBlock : ExitBlocks)
1011 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
1014 // Can't insert into a catchswitch.
1015 for (BasicBlock *ExitBlock : ExitBlocks)
1016 if (isa<CatchSwitchInst>(ExitBlock->getTerminator()))
1019 // Otherwise, this is safe to promote, lets do it!
1020 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
1024 // Grab a debug location for the inserted loads/stores; given that the
1025 // inserted loads/stores have little relation to the original loads/stores,
1026 // this code just arbitrarily picks a location from one, since any debug
1027 // location is better than none.
1028 DebugLoc DL = LoopUses[0]->getDebugLoc();
1030 // We use the SSAUpdater interface to insert phi nodes as required.
1031 SmallVector<PHINode*, 16> NewPHIs;
1032 SSAUpdater SSA(&NewPHIs);
1033 LoopPromoter Promoter(SomePtr, LoopUses, SSA,
1034 PointerMustAliases, ExitBlocks,
1035 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
1037 // Set up the preheader to have a definition of the value. It is the live-out
1038 // value from the preheader that uses in the loop will use.
1039 LoadInst *PreheaderLoad =
1040 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
1041 Preheader->getTerminator());
1042 PreheaderLoad->setAlignment(Alignment);
1043 PreheaderLoad->setDebugLoc(DL);
1044 if (AATags) PreheaderLoad->setAAMetadata(AATags);
1045 SSA.AddAvailableValue(Preheader, PreheaderLoad);
1047 // Rewrite all the loads in the loop and remember all the definitions from
1048 // stores in the loop.
1049 Promoter.run(LoopUses);
1051 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
1052 if (PreheaderLoad->use_empty())
1053 PreheaderLoad->eraseFromParent();
1058 /// Simple analysis hook. Clone alias set info.
1060 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
1061 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1065 AST->copyValue(From, To);
1068 /// Simple Analysis hook. Delete value V from alias set
1070 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
1071 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1075 AST->deleteValue(V);
1078 /// Simple Analysis hook. Delete value L from alias set map.
1080 void LICM::deleteAnalysisLoop(Loop *L) {
1081 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1086 LoopToAliasSetMap.erase(L);
1090 /// Return true if the body of this loop may store into the memory
1091 /// location pointed to by V.
1093 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
1094 const AAMDNodes &AAInfo,
1095 AliasSetTracker *CurAST) {
1096 // Check to see if any of the basic blocks in CurLoop invalidate *V.
1097 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
1100 /// Little predicate that returns true if the specified basic block is in
1101 /// a subloop of the current one, not the current one itself.
1103 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1104 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1105 return LI->getLoopFor(BB) != CurLoop;