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::iterator LoopItr = L->begin(), LoopItrE = L->end();
207 LoopItr != LoopItrE; ++LoopItr) {
208 Loop *InnerL = *LoopItr;
209 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
210 assert(InnerAST && "Where is my AST?");
212 // What if InnerLoop was modified by other passes ?
213 CurAST->add(*InnerAST);
215 // Once we've incorporated the inner loop's AST into ours, we don't need the
216 // subloop's anymore.
218 LoopToAliasSetMap.erase(InnerL);
223 // Get the preheader block to move instructions into...
224 Preheader = L->getLoopPreheader();
226 // Loop over the body of this loop, looking for calls, invokes, and stores.
227 // Because subloops have already been incorporated into AST, we skip blocks in
230 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
233 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
234 CurAST->add(*BB); // Incorporate the specified basic block
237 // Compute loop safety information.
238 LICMSafetyInfo SafetyInfo;
239 computeLICMSafetyInfo(&SafetyInfo, CurLoop);
241 // We want to visit all of the instructions in this loop... that are not parts
242 // of our subloops (they have already had their invariants hoisted out of
243 // their loop, into this loop, so there is no need to process the BODIES of
246 // Traverse the body of the loop in depth first order on the dominator tree so
247 // that we are guaranteed to see definitions before we see uses. This allows
248 // us to sink instructions in one pass, without iteration. After sinking
249 // instructions, we perform another pass to hoist them out of the loop.
251 if (L->hasDedicatedExits())
252 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, CurLoop,
253 CurAST, &SafetyInfo);
255 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI,
256 CurLoop, CurAST, &SafetyInfo);
258 // Now that all loop invariants have been removed from the loop, promote any
259 // memory references to scalars that we can.
260 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
261 SmallVector<BasicBlock *, 8> ExitBlocks;
262 SmallVector<Instruction *, 8> InsertPts;
263 PredIteratorCache PIC;
265 // Loop over all of the alias sets in the tracker object.
266 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
268 Changed |= promoteLoopAccessesToScalars(*I, ExitBlocks, InsertPts,
269 PIC, LI, DT, CurLoop,
270 CurAST, &SafetyInfo);
272 // Once we have promoted values across the loop body we have to recursively
273 // reform LCSSA as any nested loop may now have values defined within the
274 // loop used in the outer loop.
275 // FIXME: This is really heavy handed. It would be a bit better to use an
276 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
279 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
280 formLCSSARecursively(*L, *DT, LI, SEWP ? &SEWP->getSE() : nullptr);
284 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
285 // specifically moving instructions across the loop boundary and so it is
286 // especially in need of sanity checking here.
287 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
288 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
289 "Parent loop not left in LCSSA form after LICM!");
291 // Clear out loops state information for the next iteration
295 // If this loop is nested inside of another one, save the alias information
296 // for when we process the outer loop.
297 if (L->getParentLoop())
298 LoopToAliasSetMap[L] = CurAST;
304 /// Walk the specified region of the CFG (defined by all blocks dominated by
305 /// the specified block, and that are in the current loop) in reverse depth
306 /// first order w.r.t the DominatorTree. This allows us to visit uses before
307 /// definitions, allowing us to sink a loop body in one pass without iteration.
309 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
310 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
311 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
314 assert(N != nullptr && AA != nullptr && LI != nullptr &&
315 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
316 SafetyInfo != nullptr && "Unexpected input to sinkRegion");
318 // Set changed as false.
319 bool Changed = false;
321 BasicBlock *BB = N->getBlock();
322 // If this subregion is not in the top level loop at all, exit.
323 if (!CurLoop->contains(BB)) return Changed;
325 // We are processing blocks in reverse dfo, so process children first.
326 const std::vector<DomTreeNode*> &Children = N->getChildren();
327 for (unsigned i = 0, e = Children.size(); i != e; ++i)
329 sinkRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
330 // Only need to process the contents of this block if it is not part of a
331 // subloop (which would already have been processed).
332 if (inSubLoop(BB,CurLoop,LI)) return Changed;
334 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
335 Instruction &I = *--II;
337 // If the instruction is dead, we would try to sink it because it isn't used
338 // in the loop, instead, just delete it.
339 if (isInstructionTriviallyDead(&I, TLI)) {
340 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
342 CurAST->deleteValue(&I);
348 // Check to see if we can sink this instruction to the exit blocks
349 // of the loop. We can do this if the all users of the instruction are
350 // outside of the loop. In this case, it doesn't even matter if the
351 // operands of the instruction are loop invariant.
353 if (isNotUsedInLoop(I, CurLoop, SafetyInfo) &&
354 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo)) {
356 Changed |= sink(I, LI, DT, CurLoop, CurAST, SafetyInfo);
362 /// Walk the specified region of the CFG (defined by all blocks dominated by
363 /// the specified block, and that are in the current loop) in depth first
364 /// order w.r.t the DominatorTree. This allows us to visit definitions before
365 /// uses, allowing us to hoist a loop body in one pass without iteration.
367 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
368 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
369 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
371 assert(N != nullptr && AA != nullptr && LI != nullptr &&
372 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
373 SafetyInfo != nullptr && "Unexpected input to hoistRegion");
374 // Set changed as false.
375 bool Changed = false;
377 BasicBlock *BB = N->getBlock();
378 // If this subregion is not in the top level loop at all, exit.
379 if (!CurLoop->contains(BB)) return Changed;
380 // Only need to process the contents of this block if it is not part of a
381 // subloop (which would already have been processed).
382 if (!inSubLoop(BB, CurLoop, LI))
383 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
384 Instruction &I = *II++;
385 // Try constant folding this instruction. If all the operands are
386 // constants, it is technically hoistable, but it would be better to just
388 if (Constant *C = ConstantFoldInstruction(
389 &I, I.getModule()->getDataLayout(), TLI)) {
390 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
391 CurAST->copyValue(&I, C);
392 CurAST->deleteValue(&I);
393 I.replaceAllUsesWith(C);
398 // Try hoisting the instruction out to the preheader. We can only do this
399 // if all of the operands of the instruction are loop invariant and if it
400 // is safe to hoist the instruction.
402 if (CurLoop->hasLoopInvariantOperands(&I) &&
403 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo) &&
404 isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
405 CurLoop->getLoopPreheader()->getTerminator()))
406 Changed |= hoist(I, CurLoop->getLoopPreheader());
409 const std::vector<DomTreeNode*> &Children = N->getChildren();
410 for (unsigned i = 0, e = Children.size(); i != e; ++i)
412 hoistRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
416 /// Computes loop safety information, checks loop body & header
417 /// for the possibility of may throw exception.
419 void llvm::computeLICMSafetyInfo(LICMSafetyInfo * SafetyInfo, Loop * CurLoop) {
420 assert(CurLoop != nullptr && "CurLoop cant be null");
421 BasicBlock *Header = CurLoop->getHeader();
422 // Setting default safety values.
423 SafetyInfo->MayThrow = false;
424 SafetyInfo->HeaderMayThrow = false;
425 // Iterate over header and compute safety info.
426 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
427 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
428 SafetyInfo->HeaderMayThrow |= I->mayThrow();
430 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
431 // Iterate over loop instructions and compute safety info.
432 for (Loop::block_iterator BB = CurLoop->block_begin(),
433 BBE = CurLoop->block_end(); (BB != BBE) && !SafetyInfo->MayThrow ; ++BB)
434 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
435 (I != E) && !SafetyInfo->MayThrow; ++I)
436 SafetyInfo->MayThrow |= I->mayThrow();
438 // Compute funclet colors if we might sink/hoist in a function with a funclet
439 // personality routine.
440 Function *Fn = CurLoop->getHeader()->getParent();
441 if (Fn->hasPersonalityFn())
442 if (Constant *PersonalityFn = Fn->getPersonalityFn())
443 if (isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)))
444 SafetyInfo->BlockColors = colorEHFunclets(*Fn);
447 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
450 bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA, DominatorTree *DT,
451 TargetLibraryInfo *TLI, Loop *CurLoop,
452 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
453 // Loads have extra constraints we have to verify before we can hoist them.
454 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
455 if (!LI->isUnordered())
456 return false; // Don't hoist volatile/atomic loads!
458 // Loads from constant memory are always safe to move, even if they end up
459 // in the same alias set as something that ends up being modified.
460 if (AA->pointsToConstantMemory(LI->getOperand(0)))
462 if (LI->getMetadata(LLVMContext::MD_invariant_load))
465 // Don't hoist loads which have may-aliased stores in loop.
467 if (LI->getType()->isSized())
468 Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType());
471 LI->getAAMetadata(AAInfo);
473 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
474 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
475 // Don't sink or hoist dbg info; it's legal, but not useful.
476 if (isa<DbgInfoIntrinsic>(I))
479 // Don't sink calls which can throw.
483 // Handle simple cases by querying alias analysis.
484 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
485 if (Behavior == FMRB_DoesNotAccessMemory)
487 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
488 // A readonly argmemonly function only reads from memory pointed to by
489 // it's arguments with arbitrary offsets. If we can prove there are no
490 // writes to this memory in the loop, we can hoist or sink.
491 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
492 for (Value *Op : CI->arg_operands())
493 if (Op->getType()->isPointerTy() &&
494 pointerInvalidatedByLoop(Op, MemoryLocation::UnknownSize,
495 AAMDNodes(), CurAST))
499 // If this call only reads from memory and there are no writes to memory
500 // in the loop, we can hoist or sink the call as appropriate.
501 bool FoundMod = false;
502 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
505 if (!AS.isForwardingAliasSet() && AS.isMod()) {
510 if (!FoundMod) return true;
513 // FIXME: This should use mod/ref information to see if we can hoist or
519 // Only these instructions are hoistable/sinkable.
520 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
521 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
522 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
523 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
524 !isa<InsertValueInst>(I))
527 // TODO: Plumb the context instruction through to make hoisting and sinking
528 // more powerful. Hoisting of loads already works due to the special casing
530 return isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
534 /// Returns true if a PHINode is a trivially replaceable with an
536 /// This is true when all incoming values are that instruction.
537 /// This pattern occurs most often with LCSSA PHI nodes.
539 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
540 for (const Value *IncValue : PN.incoming_values())
547 /// Return true if the only users of this instruction are outside of
548 /// the loop. If this is true, we can sink the instruction to the exit
549 /// blocks of the loop.
551 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop,
552 const LICMSafetyInfo *SafetyInfo) {
553 const auto &BlockColors = SafetyInfo->BlockColors;
554 for (const User *U : I.users()) {
555 const Instruction *UI = cast<Instruction>(U);
556 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
557 const BasicBlock *BB = PN->getParent();
558 // We cannot sink uses in catchswitches.
559 if (isa<CatchSwitchInst>(BB->getTerminator()))
562 // We need to sink a callsite to a unique funclet. Avoid sinking if the
563 // phi use is too muddled.
564 if (isa<CallInst>(I))
565 if (!BlockColors.empty() &&
566 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
569 // A PHI node where all of the incoming values are this instruction are
570 // special -- they can just be RAUW'ed with the instruction and thus
571 // don't require a use in the predecessor. This is a particular important
572 // special case because it is the pattern found in LCSSA form.
573 if (isTriviallyReplacablePHI(*PN, I)) {
574 if (CurLoop->contains(PN))
580 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
581 // values. Check for such a use being inside the loop.
582 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
583 if (PN->getIncomingValue(i) == &I)
584 if (CurLoop->contains(PN->getIncomingBlock(i)))
590 if (CurLoop->contains(UI))
597 CloneInstructionInExitBlock(Instruction &I, BasicBlock &ExitBlock, PHINode &PN,
599 const LICMSafetyInfo *SafetyInfo) {
601 if (auto *CI = dyn_cast<CallInst>(&I)) {
602 const auto &BlockColors = SafetyInfo->BlockColors;
604 // Sinking call-sites need to be handled differently from other
605 // instructions. The cloned call-site needs a funclet bundle operand
606 // appropriate for it's location in the CFG.
607 SmallVector<OperandBundleDef, 1> OpBundles;
608 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
609 BundleIdx != BundleEnd; ++BundleIdx) {
610 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
611 if (Bundle.getTagID() == LLVMContext::OB_funclet)
614 OpBundles.emplace_back(Bundle);
617 if (!BlockColors.empty()) {
618 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
619 assert(CV.size() == 1 && "non-unique color for exit block!");
620 BasicBlock *BBColor = CV.front();
621 Instruction *EHPad = BBColor->getFirstNonPHI();
622 if (EHPad->isEHPad())
623 OpBundles.emplace_back("funclet", EHPad);
626 New = CallInst::Create(CI, OpBundles);
631 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
632 if (!I.getName().empty()) New->setName(I.getName() + ".le");
634 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
635 // particularly cheap because we can rip off the PHI node that we're
636 // replacing for the number and blocks of the predecessors.
637 // OPT: If this shows up in a profile, we can instead finish sinking all
638 // invariant instructions, and then walk their operands to re-establish
639 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
640 // sinking bottom-up.
641 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
643 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
644 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
645 if (!OLoop->contains(&PN)) {
647 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
648 OInst->getName() + ".lcssa", &ExitBlock.front());
649 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
650 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
656 /// When an instruction is found to only be used outside of the loop, this
657 /// function moves it to the exit blocks and patches up SSA form as needed.
658 /// This method is guaranteed to remove the original instruction from its
659 /// position, and may either delete it or move it to outside of the loop.
661 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
662 const Loop *CurLoop, AliasSetTracker *CurAST,
663 const LICMSafetyInfo *SafetyInfo) {
664 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
665 bool Changed = false;
666 if (isa<LoadInst>(I)) ++NumMovedLoads;
667 else if (isa<CallInst>(I)) ++NumMovedCalls;
672 SmallVector<BasicBlock *, 32> ExitBlocks;
673 CurLoop->getUniqueExitBlocks(ExitBlocks);
674 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
678 // Clones of this instruction. Don't create more than one per exit block!
679 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
681 // If this instruction is only used outside of the loop, then all users are
682 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
684 while (!I.use_empty()) {
685 Value::user_iterator UI = I.user_begin();
686 auto *User = cast<Instruction>(*UI);
687 if (!DT->isReachableFromEntry(User->getParent())) {
688 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
691 // The user must be a PHI node.
692 PHINode *PN = cast<PHINode>(User);
694 // Surprisingly, instructions can be used outside of loops without any
695 // exits. This can only happen in PHI nodes if the incoming block is
697 Use &U = UI.getUse();
698 BasicBlock *BB = PN->getIncomingBlock(U);
699 if (!DT->isReachableFromEntry(BB)) {
700 U = UndefValue::get(I.getType());
704 BasicBlock *ExitBlock = PN->getParent();
705 assert(ExitBlockSet.count(ExitBlock) &&
706 "The LCSSA PHI is not in an exit block!");
709 auto It = SunkCopies.find(ExitBlock);
710 if (It != SunkCopies.end())
713 New = SunkCopies[ExitBlock] =
714 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI, SafetyInfo);
716 PN->replaceAllUsesWith(New);
717 PN->eraseFromParent();
720 CurAST->deleteValue(&I);
725 /// When an instruction is found to only use loop invariant operands that
726 /// is safe to hoist, this instruction is called to do the dirty work.
728 static bool hoist(Instruction &I, BasicBlock *Preheader) {
729 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
731 // Move the new node to the Preheader, before its terminator.
732 I.moveBefore(Preheader->getTerminator());
734 // Metadata can be dependent on the condition we are hoisting above.
735 // Conservatively strip all metadata on the instruction.
736 I.dropUnknownNonDebugMetadata();
738 if (isa<LoadInst>(I)) ++NumMovedLoads;
739 else if (isa<CallInst>(I)) ++NumMovedCalls;
744 /// Only sink or hoist an instruction if it is not a trapping instruction,
745 /// or if the instruction is known not to trap when moved to the preheader.
746 /// or if it is a trapping instruction and is guaranteed to execute.
747 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
748 const DominatorTree *DT,
749 const TargetLibraryInfo *TLI,
751 const LICMSafetyInfo *SafetyInfo,
752 const Instruction *CtxI) {
753 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI))
756 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
759 static bool isGuaranteedToExecute(const Instruction &Inst,
760 const DominatorTree *DT,
762 const LICMSafetyInfo * SafetyInfo) {
764 // We have to check to make sure that the instruction dominates all
765 // of the exit blocks. If it doesn't, then there is a path out of the loop
766 // which does not execute this instruction, so we can't hoist it.
768 // If the instruction is in the header block for the loop (which is very
769 // common), it is always guaranteed to dominate the exit blocks. Since this
770 // is a common case, and can save some work, check it now.
771 if (Inst.getParent() == CurLoop->getHeader())
772 // If there's a throw in the header block, we can't guarantee we'll reach
774 return !SafetyInfo->HeaderMayThrow;
776 // Somewhere in this loop there is an instruction which may throw and make us
778 if (SafetyInfo->MayThrow)
781 // Get the exit blocks for the current loop.
782 SmallVector<BasicBlock*, 8> ExitBlocks;
783 CurLoop->getExitBlocks(ExitBlocks);
785 // Verify that the block dominates each of the exit blocks of the loop.
786 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
787 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
790 // As a degenerate case, if the loop is statically infinite then we haven't
791 // proven anything since there are no exit blocks.
792 if (ExitBlocks.empty())
799 class LoopPromoter : public LoadAndStorePromoter {
800 Value *SomePtr; // Designated pointer to store to.
801 SmallPtrSetImpl<Value*> &PointerMustAliases;
802 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
803 SmallVectorImpl<Instruction*> &LoopInsertPts;
804 PredIteratorCache &PredCache;
805 AliasSetTracker &AST;
811 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
812 if (Instruction *I = dyn_cast<Instruction>(V))
813 if (Loop *L = LI.getLoopFor(I->getParent()))
814 if (!L->contains(BB)) {
815 // We need to create an LCSSA PHI node for the incoming value and
818 PHINode::Create(I->getType(), PredCache.size(BB),
819 I->getName() + ".lcssa", &BB->front());
820 for (BasicBlock *Pred : PredCache.get(BB))
821 PN->addIncoming(I, Pred);
828 LoopPromoter(Value *SP,
829 ArrayRef<const Instruction *> Insts,
830 SSAUpdater &S, SmallPtrSetImpl<Value *> &PMA,
831 SmallVectorImpl<BasicBlock *> &LEB,
832 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
833 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
834 const AAMDNodes &AATags)
835 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
836 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
837 LI(li), DL(dl), Alignment(alignment), AATags(AATags) {}
839 bool isInstInList(Instruction *I,
840 const SmallVectorImpl<Instruction*> &) const override {
842 if (LoadInst *LI = dyn_cast<LoadInst>(I))
843 Ptr = LI->getOperand(0);
845 Ptr = cast<StoreInst>(I)->getPointerOperand();
846 return PointerMustAliases.count(Ptr);
849 void doExtraRewritesBeforeFinalDeletion() const override {
850 // Insert stores after in the loop exit blocks. Each exit block gets a
851 // store of the live-out values that feed them. Since we've already told
852 // the SSA updater about the defs in the loop and the preheader
853 // definition, it is all set and we can start using it.
854 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
855 BasicBlock *ExitBlock = LoopExitBlocks[i];
856 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
857 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
858 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
859 Instruction *InsertPos = LoopInsertPts[i];
860 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
861 NewSI->setAlignment(Alignment);
862 NewSI->setDebugLoc(DL);
863 if (AATags) NewSI->setAAMetadata(AATags);
867 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
868 // Update alias analysis.
869 AST.copyValue(LI, V);
871 void instructionDeleted(Instruction *I) const override {
875 } // end anon namespace
877 /// Try to promote memory values to scalars by sinking stores out of the
878 /// loop and moving loads to before the loop. We do this by looping over
879 /// the stores in the loop, looking for stores to Must pointers which are
882 bool llvm::promoteLoopAccessesToScalars(AliasSet &AS,
883 SmallVectorImpl<BasicBlock*>&ExitBlocks,
884 SmallVectorImpl<Instruction*>&InsertPts,
885 PredIteratorCache &PIC, LoopInfo *LI,
886 DominatorTree *DT, Loop *CurLoop,
887 AliasSetTracker *CurAST,
888 LICMSafetyInfo * SafetyInfo) {
890 assert(LI != nullptr && DT != nullptr &&
891 CurLoop != nullptr && CurAST != nullptr &&
892 SafetyInfo != nullptr &&
893 "Unexpected Input to promoteLoopAccessesToScalars");
894 // Initially set Changed status to false.
895 bool Changed = false;
896 // We can promote this alias set if it has a store, if it is a "Must" alias
897 // set, if the pointer is loop invariant, and if we are not eliminating any
898 // volatile loads or stores.
899 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
900 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
903 assert(!AS.empty() &&
904 "Must alias set should have at least one pointer element in it!");
906 Value *SomePtr = AS.begin()->getValue();
907 BasicBlock * Preheader = CurLoop->getLoopPreheader();
909 // It isn't safe to promote a load/store from the loop if the load/store is
910 // conditional. For example, turning:
912 // for () { if (c) *P += 1; }
916 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
918 // is not safe, because *P may only be valid to access if 'c' is true.
920 // It is safe to promote P if all uses are direct load/stores and if at
921 // least one is guaranteed to be executed.
922 bool GuaranteedToExecute = false;
924 SmallVector<Instruction*, 64> LoopUses;
925 SmallPtrSet<Value*, 4> PointerMustAliases;
927 // We start with an alignment of one and try to find instructions that allow
928 // us to prove better alignment.
929 unsigned Alignment = 1;
931 bool HasDedicatedExits = CurLoop->hasDedicatedExits();
933 // Check that all of the pointers in the alias set have the same type. We
934 // cannot (yet) promote a memory location that is loaded and stored in
935 // different sizes. While we are at it, collect alignment and AA info.
936 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
937 Value *ASIV = ASI->getValue();
938 PointerMustAliases.insert(ASIV);
940 // Check that all of the pointers in the alias set have the same type. We
941 // cannot (yet) promote a memory location that is loaded and stored in
943 if (SomePtr->getType() != ASIV->getType())
946 for (User *U : ASIV->users()) {
947 // Ignore instructions that are outside the loop.
948 Instruction *UI = dyn_cast<Instruction>(U);
949 if (!UI || !CurLoop->contains(UI))
952 // If there is an non-load/store instruction in the loop, we can't promote
954 if (const LoadInst *load = dyn_cast<LoadInst>(UI)) {
955 assert(!load->isVolatile() && "AST broken");
956 if (!load->isSimple())
958 } else if (const StoreInst *store = dyn_cast<StoreInst>(UI)) {
959 // Stores *of* the pointer are not interesting, only stores *to* the
961 if (UI->getOperand(1) != ASIV)
963 assert(!store->isVolatile() && "AST broken");
964 if (!store->isSimple())
966 // Don't sink stores from loops without dedicated block exits. Exits
967 // containing indirect branches are not transformed by loop simplify,
968 // make sure we catch that. An additional load may be generated in the
969 // preheader for SSA updater, so also avoid sinking when no preheader
971 if (!HasDedicatedExits || !Preheader)
974 // Note that we only check GuaranteedToExecute inside the store case
975 // so that we do not introduce stores where they did not exist before
976 // (which would break the LLVM concurrency model).
978 // If the alignment of this instruction allows us to specify a more
979 // restrictive (and performant) alignment and if we are sure this
980 // instruction will be executed, update the alignment.
981 // Larger is better, with the exception of 0 being the best alignment.
982 unsigned InstAlignment = store->getAlignment();
983 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
984 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
985 GuaranteedToExecute = true;
986 Alignment = InstAlignment;
989 if (!GuaranteedToExecute)
990 GuaranteedToExecute = isGuaranteedToExecute(*UI, DT,
991 CurLoop, SafetyInfo);
994 return Changed; // Not a load or store.
996 // Merge the AA tags.
997 if (LoopUses.empty()) {
998 // On the first load/store, just take its AA tags.
999 UI->getAAMetadata(AATags);
1000 } else if (AATags) {
1001 UI->getAAMetadata(AATags, /* Merge = */ true);
1004 LoopUses.push_back(UI);
1008 // If there isn't a guaranteed-to-execute instruction, we can't promote.
1009 if (!GuaranteedToExecute)
1012 // Otherwise, this is safe to promote, lets do it!
1013 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
1017 // Grab a debug location for the inserted loads/stores; given that the
1018 // inserted loads/stores have little relation to the original loads/stores,
1019 // this code just arbitrarily picks a location from one, since any debug
1020 // location is better than none.
1021 DebugLoc DL = LoopUses[0]->getDebugLoc();
1023 // Figure out the loop exits and their insertion points, if this is the
1025 if (ExitBlocks.empty()) {
1026 CurLoop->getUniqueExitBlocks(ExitBlocks);
1027 InsertPts.resize(ExitBlocks.size());
1028 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
1029 InsertPts[i] = &*ExitBlocks[i]->getFirstInsertionPt();
1032 // We use the SSAUpdater interface to insert phi nodes as required.
1033 SmallVector<PHINode*, 16> NewPHIs;
1034 SSAUpdater SSA(&NewPHIs);
1035 LoopPromoter Promoter(SomePtr, LoopUses, SSA,
1036 PointerMustAliases, ExitBlocks,
1037 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
1039 // Set up the preheader to have a definition of the value. It is the live-out
1040 // value from the preheader that uses in the loop will use.
1041 LoadInst *PreheaderLoad =
1042 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
1043 Preheader->getTerminator());
1044 PreheaderLoad->setAlignment(Alignment);
1045 PreheaderLoad->setDebugLoc(DL);
1046 if (AATags) PreheaderLoad->setAAMetadata(AATags);
1047 SSA.AddAvailableValue(Preheader, PreheaderLoad);
1049 // Rewrite all the loads in the loop and remember all the definitions from
1050 // stores in the loop.
1051 Promoter.run(LoopUses);
1053 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
1054 if (PreheaderLoad->use_empty())
1055 PreheaderLoad->eraseFromParent();
1060 /// Simple analysis hook. Clone alias set info.
1062 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
1063 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1067 AST->copyValue(From, To);
1070 /// Simple Analysis hook. Delete value V from alias set
1072 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
1073 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1077 AST->deleteValue(V);
1080 /// Simple Analysis hook. Delete value L from alias set map.
1082 void LICM::deleteAnalysisLoop(Loop *L) {
1083 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1088 LoopToAliasSetMap.erase(L);
1092 /// Return true if the body of this loop may store into the memory
1093 /// location pointed to by V.
1095 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
1096 const AAMDNodes &AAInfo,
1097 AliasSetTracker *CurAST) {
1098 // Check to see if any of the basic blocks in CurLoop invalidate *V.
1099 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
1102 /// Little predicate that returns true if the specified basic block is in
1103 /// a subloop of the current one, not the current one itself.
1105 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1106 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1107 return LI->getLoopFor(BB) != CurLoop;