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 static bool hoist(Instruction &I, BasicBlock *Preheader);
80 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
81 const Loop *CurLoop, AliasSetTracker *CurAST );
82 static bool isGuaranteedToExecute(const Instruction &Inst,
83 const DominatorTree *DT,
85 const LICMSafetyInfo *SafetyInfo);
86 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
87 const DominatorTree *DT,
88 const TargetLibraryInfo *TLI,
90 const LICMSafetyInfo *SafetyInfo,
91 const Instruction *CtxI = nullptr);
92 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
93 const AAMDNodes &AAInfo,
94 AliasSetTracker *CurAST);
95 static Instruction *CloneInstructionInExitBlock(const Instruction &I,
96 BasicBlock &ExitBlock,
99 static bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA,
100 DominatorTree *DT, TargetLibraryInfo *TLI,
101 Loop *CurLoop, AliasSetTracker *CurAST,
102 LICMSafetyInfo *SafetyInfo);
105 struct LICM : public LoopPass {
106 static char ID; // Pass identification, replacement for typeid
107 LICM() : LoopPass(ID) {
108 initializeLICMPass(*PassRegistry::getPassRegistry());
111 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
113 /// This transformation requires natural loop information & requires that
114 /// loop preheaders be inserted into the CFG...
116 void getAnalysisUsage(AnalysisUsage &AU) const override {
117 AU.setPreservesCFG();
118 AU.addRequired<DominatorTreeWrapperPass>();
119 AU.addRequired<LoopInfoWrapperPass>();
120 AU.addRequiredID(LoopSimplifyID);
121 AU.addPreservedID(LoopSimplifyID);
122 AU.addRequiredID(LCSSAID);
123 AU.addPreservedID(LCSSAID);
124 AU.addRequired<AAResultsWrapperPass>();
125 AU.addPreserved<AAResultsWrapperPass>();
126 AU.addPreserved<BasicAAWrapperPass>();
127 AU.addPreserved<GlobalsAAWrapperPass>();
128 AU.addPreserved<ScalarEvolutionWrapperPass>();
129 AU.addPreserved<SCEVAAWrapperPass>();
130 AU.addRequired<TargetLibraryInfoWrapperPass>();
133 using llvm::Pass::doFinalization;
135 bool doFinalization() override {
136 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
141 AliasAnalysis *AA; // Current AliasAnalysis information
142 LoopInfo *LI; // Current LoopInfo
143 DominatorTree *DT; // Dominator Tree for the current Loop.
145 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
147 // State that is updated as we process loops.
148 bool Changed; // Set to true when we change anything.
149 BasicBlock *Preheader; // The preheader block of the current loop...
150 Loop *CurLoop; // The current loop we are working on...
151 AliasSetTracker *CurAST; // AliasSet information for the current loop...
152 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
154 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
155 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
158 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
160 void deleteAnalysisValue(Value *V, Loop *L) override;
162 /// Simple Analysis hook. Delete loop L from alias set map.
163 void deleteAnalysisLoop(Loop *L) override;
168 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
169 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
170 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
171 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
172 INITIALIZE_PASS_DEPENDENCY(LCSSA)
173 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
174 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
175 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
176 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
177 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
178 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
179 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
181 Pass *llvm::createLICMPass() { return new LICM(); }
183 /// Hoist expressions out of the specified loop. Note, alias info for inner
184 /// loop is not preserved so it is not a good idea to run LICM multiple
185 /// times on one loop.
187 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
188 if (skipOptnoneFunction(L))
193 // Get our Loop and Alias Analysis information...
194 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
195 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
196 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
198 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
200 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
202 CurAST = new AliasSetTracker(*AA);
203 // Collect Alias info from subloops.
204 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
205 LoopItr != LoopItrE; ++LoopItr) {
206 Loop *InnerL = *LoopItr;
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 (Loop::block_iterator I = L->block_begin(), E = L->block_end();
231 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
232 CurAST->add(*BB); // Incorporate the specified basic block
235 // Compute loop safety information.
236 LICMSafetyInfo SafetyInfo;
237 computeLICMSafetyInfo(&SafetyInfo, CurLoop);
239 // We want to visit all of the instructions in this loop... that are not parts
240 // of our subloops (they have already had their invariants hoisted out of
241 // their loop, into this loop, so there is no need to process the BODIES of
244 // Traverse the body of the loop in depth first order on the dominator tree so
245 // that we are guaranteed to see definitions before we see uses. This allows
246 // us to sink instructions in one pass, without iteration. After sinking
247 // instructions, we perform another pass to hoist them out of the loop.
249 if (L->hasDedicatedExits())
250 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, CurLoop,
251 CurAST, &SafetyInfo);
253 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI,
254 CurLoop, CurAST, &SafetyInfo);
256 // Now that all loop invariants have been removed from the loop, promote any
257 // memory references to scalars that we can.
258 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
259 SmallVector<BasicBlock *, 8> ExitBlocks;
260 SmallVector<Instruction *, 8> InsertPts;
261 PredIteratorCache PIC;
263 // Loop over all of the alias sets in the tracker object.
264 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
266 Changed |= promoteLoopAccessesToScalars(*I, ExitBlocks, InsertPts,
267 PIC, LI, DT, CurLoop,
268 CurAST, &SafetyInfo);
270 // Once we have promoted values across the loop body we have to recursively
271 // reform LCSSA as any nested loop may now have values defined within the
272 // loop used in the outer loop.
273 // FIXME: This is really heavy handed. It would be a bit better to use an
274 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
277 auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
278 formLCSSARecursively(*L, *DT, LI, SEWP ? &SEWP->getSE() : nullptr);
282 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
283 // specifically moving instructions across the loop boundary and so it is
284 // especially in need of sanity checking here.
285 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
286 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
287 "Parent loop not left in LCSSA form after LICM!");
289 // Clear out loops state information for the next iteration
293 // If this loop is nested inside of another one, save the alias information
294 // for when we process the outer loop.
295 if (L->getParentLoop())
296 LoopToAliasSetMap[L] = CurAST;
302 /// Walk the specified region of the CFG (defined by all blocks dominated by
303 /// the specified block, and that are in the current loop) in reverse depth
304 /// first order w.r.t the DominatorTree. This allows us to visit uses before
305 /// definitions, allowing us to sink a loop body in one pass without iteration.
307 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
308 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
309 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
312 assert(N != nullptr && AA != nullptr && LI != nullptr &&
313 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
314 SafetyInfo != nullptr && "Unexpected input to sinkRegion");
316 // Set changed as false.
317 bool Changed = false;
319 BasicBlock *BB = N->getBlock();
320 // If this subregion is not in the top level loop at all, exit.
321 if (!CurLoop->contains(BB)) return Changed;
323 // We are processing blocks in reverse dfo, so process children first.
324 const std::vector<DomTreeNode*> &Children = N->getChildren();
325 for (unsigned i = 0, e = Children.size(); i != e; ++i)
327 sinkRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
328 // Only need to process the contents of this block if it is not part of a
329 // subloop (which would already have been processed).
330 if (inSubLoop(BB,CurLoop,LI)) return Changed;
332 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
333 Instruction &I = *--II;
335 // If the instruction is dead, we would try to sink it because it isn't used
336 // in the loop, instead, just delete it.
337 if (isInstructionTriviallyDead(&I, TLI)) {
338 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
340 CurAST->deleteValue(&I);
346 // Check to see if we can sink this instruction to the exit blocks
347 // of the loop. We can do this if the all users of the instruction are
348 // outside of the loop. In this case, it doesn't even matter if the
349 // operands of the instruction are loop invariant.
351 if (isNotUsedInLoop(I, CurLoop) &&
352 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo)) {
354 Changed |= sink(I, LI, DT, CurLoop, CurAST);
360 /// Walk the specified region of the CFG (defined by all blocks dominated by
361 /// the specified block, and that are in the current loop) in depth first
362 /// order w.r.t the DominatorTree. This allows us to visit definitions before
363 /// uses, allowing us to hoist a loop body in one pass without iteration.
365 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
366 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
367 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
369 assert(N != nullptr && AA != nullptr && LI != nullptr &&
370 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
371 SafetyInfo != nullptr && "Unexpected input to hoistRegion");
372 // Set changed as false.
373 bool Changed = false;
375 BasicBlock *BB = N->getBlock();
376 // If this subregion is not in the top level loop at all, exit.
377 if (!CurLoop->contains(BB)) return Changed;
378 // Only need to process the contents of this block if it is not part of a
379 // subloop (which would already have been processed).
380 if (!inSubLoop(BB, CurLoop, LI))
381 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
382 Instruction &I = *II++;
383 // Try constant folding this instruction. If all the operands are
384 // constants, it is technically hoistable, but it would be better to just
386 if (Constant *C = ConstantFoldInstruction(
387 &I, I.getModule()->getDataLayout(), TLI)) {
388 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
389 CurAST->copyValue(&I, C);
390 CurAST->deleteValue(&I);
391 I.replaceAllUsesWith(C);
396 // Try hoisting the instruction out to the preheader. We can only do this
397 // if all of the operands of the instruction are loop invariant and if it
398 // is safe to hoist the instruction.
400 if (CurLoop->hasLoopInvariantOperands(&I) &&
401 canSinkOrHoistInst(I, AA, DT, TLI, CurLoop, CurAST, SafetyInfo) &&
402 isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
403 CurLoop->getLoopPreheader()->getTerminator()))
404 Changed |= hoist(I, CurLoop->getLoopPreheader());
407 const std::vector<DomTreeNode*> &Children = N->getChildren();
408 for (unsigned i = 0, e = Children.size(); i != e; ++i)
410 hoistRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
414 /// Computes loop safety information, checks loop body & header
415 /// for the possibility of may throw exception.
417 void llvm::computeLICMSafetyInfo(LICMSafetyInfo * SafetyInfo, Loop * CurLoop) {
418 assert(CurLoop != nullptr && "CurLoop cant be null");
419 BasicBlock *Header = CurLoop->getHeader();
420 // Setting default safety values.
421 SafetyInfo->MayThrow = false;
422 SafetyInfo->HeaderMayThrow = false;
423 // Iterate over header and compute safety info.
424 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
425 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
426 SafetyInfo->HeaderMayThrow |= I->mayThrow();
428 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
429 // Iterate over loop instructions and compute safety info.
430 for (Loop::block_iterator BB = CurLoop->block_begin(),
431 BBE = CurLoop->block_end(); (BB != BBE) && !SafetyInfo->MayThrow ; ++BB)
432 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
433 (I != E) && !SafetyInfo->MayThrow; ++I)
434 SafetyInfo->MayThrow |= I->mayThrow();
437 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
440 bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA, DominatorTree *DT,
441 TargetLibraryInfo *TLI, Loop *CurLoop,
442 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
443 // Loads have extra constraints we have to verify before we can hoist them.
444 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
445 if (!LI->isUnordered())
446 return false; // Don't hoist volatile/atomic loads!
448 // Loads from constant memory are always safe to move, even if they end up
449 // in the same alias set as something that ends up being modified.
450 if (AA->pointsToConstantMemory(LI->getOperand(0)))
452 if (LI->getMetadata(LLVMContext::MD_invariant_load))
455 // Don't hoist loads which have may-aliased stores in loop.
457 if (LI->getType()->isSized())
458 Size = I.getModule()->getDataLayout().getTypeStoreSize(LI->getType());
461 LI->getAAMetadata(AAInfo);
463 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
464 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
465 // Don't sink or hoist dbg info; it's legal, but not useful.
466 if (isa<DbgInfoIntrinsic>(I))
469 // Handle simple cases by querying alias analysis.
470 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
471 if (Behavior == FMRB_DoesNotAccessMemory)
473 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
474 // If this call only reads from memory and there are no writes to memory
475 // in the loop, we can hoist or sink the call as appropriate.
476 bool FoundMod = false;
477 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
480 if (!AS.isForwardingAliasSet() && AS.isMod()) {
485 if (!FoundMod) return true;
488 // FIXME: This should use mod/ref information to see if we can hoist or
494 // Only these instructions are hoistable/sinkable.
495 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
496 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
497 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
498 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
499 !isa<InsertValueInst>(I))
502 // TODO: Plumb the context instruction through to make hoisting and sinking
503 // more powerful. Hoisting of loads already works due to the special casing
505 return isSafeToExecuteUnconditionally(I, DT, TLI, CurLoop, SafetyInfo,
509 /// Returns true if a PHINode is a trivially replaceable with an
511 /// This is true when all incoming values are that instruction.
512 /// This pattern occurs most often with LCSSA PHI nodes.
514 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
515 for (const Value *IncValue : PN.incoming_values())
522 /// Return true if the only users of this instruction are outside of
523 /// the loop. If this is true, we can sink the instruction to the exit
524 /// blocks of the loop.
526 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop) {
527 for (const User *U : I.users()) {
528 const Instruction *UI = cast<Instruction>(U);
529 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
530 // A PHI node where all of the incoming values are this instruction are
531 // special -- they can just be RAUW'ed with the instruction and thus
532 // don't require a use in the predecessor. This is a particular important
533 // special case because it is the pattern found in LCSSA form.
534 if (isTriviallyReplacablePHI(*PN, I)) {
535 if (CurLoop->contains(PN))
541 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
542 // values. Check for such a use being inside the loop.
543 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
544 if (PN->getIncomingValue(i) == &I)
545 if (CurLoop->contains(PN->getIncomingBlock(i)))
551 if (CurLoop->contains(UI))
557 static Instruction *CloneInstructionInExitBlock(const Instruction &I,
558 BasicBlock &ExitBlock,
560 const LoopInfo *LI) {
561 Instruction *New = I.clone();
562 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
563 if (!I.getName().empty()) New->setName(I.getName() + ".le");
565 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
566 // particularly cheap because we can rip off the PHI node that we're
567 // replacing for the number and blocks of the predecessors.
568 // OPT: If this shows up in a profile, we can instead finish sinking all
569 // invariant instructions, and then walk their operands to re-establish
570 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
571 // sinking bottom-up.
572 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
574 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
575 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
576 if (!OLoop->contains(&PN)) {
578 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
579 OInst->getName() + ".lcssa", ExitBlock.begin());
580 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
581 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
587 /// When an instruction is found to only be used outside of the loop, this
588 /// function moves it to the exit blocks and patches up SSA form as needed.
589 /// This method is guaranteed to remove the original instruction from its
590 /// position, and may either delete it or move it to outside of the loop.
592 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
593 const Loop *CurLoop, AliasSetTracker *CurAST ) {
594 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
595 bool Changed = false;
596 if (isa<LoadInst>(I)) ++NumMovedLoads;
597 else if (isa<CallInst>(I)) ++NumMovedCalls;
602 SmallVector<BasicBlock *, 32> ExitBlocks;
603 CurLoop->getUniqueExitBlocks(ExitBlocks);
604 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
608 // Clones of this instruction. Don't create more than one per exit block!
609 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
611 // If this instruction is only used outside of the loop, then all users are
612 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
614 while (!I.use_empty()) {
615 Value::user_iterator UI = I.user_begin();
616 auto *User = cast<Instruction>(*UI);
617 if (!DT->isReachableFromEntry(User->getParent())) {
618 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
621 // The user must be a PHI node.
622 PHINode *PN = cast<PHINode>(User);
624 // Surprisingly, instructions can be used outside of loops without any
625 // exits. This can only happen in PHI nodes if the incoming block is
627 Use &U = UI.getUse();
628 BasicBlock *BB = PN->getIncomingBlock(U);
629 if (!DT->isReachableFromEntry(BB)) {
630 U = UndefValue::get(I.getType());
634 BasicBlock *ExitBlock = PN->getParent();
635 assert(ExitBlockSet.count(ExitBlock) &&
636 "The LCSSA PHI is not in an exit block!");
639 auto It = SunkCopies.find(ExitBlock);
640 if (It != SunkCopies.end())
643 New = SunkCopies[ExitBlock] =
644 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI);
646 PN->replaceAllUsesWith(New);
647 PN->eraseFromParent();
650 CurAST->deleteValue(&I);
655 /// When an instruction is found to only use loop invariant operands that
656 /// is safe to hoist, this instruction is called to do the dirty work.
658 static bool hoist(Instruction &I, BasicBlock *Preheader) {
659 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
661 // Move the new node to the Preheader, before its terminator.
662 I.moveBefore(Preheader->getTerminator());
664 if (isa<LoadInst>(I)) ++NumMovedLoads;
665 else if (isa<CallInst>(I)) ++NumMovedCalls;
670 /// Only sink or hoist an instruction if it is not a trapping instruction,
671 /// or if the instruction is known not to trap when moved to the preheader.
672 /// or if it is a trapping instruction and is guaranteed to execute.
673 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
674 const DominatorTree *DT,
675 const TargetLibraryInfo *TLI,
677 const LICMSafetyInfo *SafetyInfo,
678 const Instruction *CtxI) {
679 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT, TLI))
682 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
685 static bool isGuaranteedToExecute(const Instruction &Inst,
686 const DominatorTree *DT,
688 const LICMSafetyInfo * SafetyInfo) {
690 // We have to check to make sure that the instruction dominates all
691 // of the exit blocks. If it doesn't, then there is a path out of the loop
692 // which does not execute this instruction, so we can't hoist it.
694 // If the instruction is in the header block for the loop (which is very
695 // common), it is always guaranteed to dominate the exit blocks. Since this
696 // is a common case, and can save some work, check it now.
697 if (Inst.getParent() == CurLoop->getHeader())
698 // If there's a throw in the header block, we can't guarantee we'll reach
700 return !SafetyInfo->HeaderMayThrow;
702 // Somewhere in this loop there is an instruction which may throw and make us
704 if (SafetyInfo->MayThrow)
707 // Get the exit blocks for the current loop.
708 SmallVector<BasicBlock*, 8> ExitBlocks;
709 CurLoop->getExitBlocks(ExitBlocks);
711 // Verify that the block dominates each of the exit blocks of the loop.
712 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
713 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
716 // As a degenerate case, if the loop is statically infinite then we haven't
717 // proven anything since there are no exit blocks.
718 if (ExitBlocks.empty())
725 class LoopPromoter : public LoadAndStorePromoter {
726 Value *SomePtr; // Designated pointer to store to.
727 SmallPtrSetImpl<Value*> &PointerMustAliases;
728 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
729 SmallVectorImpl<Instruction*> &LoopInsertPts;
730 PredIteratorCache &PredCache;
731 AliasSetTracker &AST;
737 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
738 if (Instruction *I = dyn_cast<Instruction>(V))
739 if (Loop *L = LI.getLoopFor(I->getParent()))
740 if (!L->contains(BB)) {
741 // We need to create an LCSSA PHI node for the incoming value and
743 PHINode *PN = PHINode::Create(
744 I->getType(), PredCache.size(BB),
745 I->getName() + ".lcssa", BB->begin());
746 for (BasicBlock *Pred : PredCache.get(BB))
747 PN->addIncoming(I, Pred);
754 LoopPromoter(Value *SP,
755 ArrayRef<const Instruction *> Insts,
756 SSAUpdater &S, SmallPtrSetImpl<Value *> &PMA,
757 SmallVectorImpl<BasicBlock *> &LEB,
758 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
759 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
760 const AAMDNodes &AATags)
761 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
762 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
763 LI(li), DL(dl), Alignment(alignment), AATags(AATags) {}
765 bool isInstInList(Instruction *I,
766 const SmallVectorImpl<Instruction*> &) const override {
768 if (LoadInst *LI = dyn_cast<LoadInst>(I))
769 Ptr = LI->getOperand(0);
771 Ptr = cast<StoreInst>(I)->getPointerOperand();
772 return PointerMustAliases.count(Ptr);
775 void doExtraRewritesBeforeFinalDeletion() const override {
776 // Insert stores after in the loop exit blocks. Each exit block gets a
777 // store of the live-out values that feed them. Since we've already told
778 // the SSA updater about the defs in the loop and the preheader
779 // definition, it is all set and we can start using it.
780 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
781 BasicBlock *ExitBlock = LoopExitBlocks[i];
782 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
783 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
784 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
785 Instruction *InsertPos = LoopInsertPts[i];
786 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
787 NewSI->setAlignment(Alignment);
788 NewSI->setDebugLoc(DL);
789 if (AATags) NewSI->setAAMetadata(AATags);
793 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
794 // Update alias analysis.
795 AST.copyValue(LI, V);
797 void instructionDeleted(Instruction *I) const override {
801 } // end anon namespace
803 /// Try to promote memory values to scalars by sinking stores out of the
804 /// loop and moving loads to before the loop. We do this by looping over
805 /// the stores in the loop, looking for stores to Must pointers which are
808 bool llvm::promoteLoopAccessesToScalars(AliasSet &AS,
809 SmallVectorImpl<BasicBlock*>&ExitBlocks,
810 SmallVectorImpl<Instruction*>&InsertPts,
811 PredIteratorCache &PIC, LoopInfo *LI,
812 DominatorTree *DT, Loop *CurLoop,
813 AliasSetTracker *CurAST,
814 LICMSafetyInfo * SafetyInfo) {
816 assert(LI != nullptr && DT != nullptr &&
817 CurLoop != nullptr && CurAST != nullptr &&
818 SafetyInfo != nullptr &&
819 "Unexpected Input to promoteLoopAccessesToScalars");
820 // Initially set Changed status to false.
821 bool Changed = false;
822 // We can promote this alias set if it has a store, if it is a "Must" alias
823 // set, if the pointer is loop invariant, and if we are not eliminating any
824 // volatile loads or stores.
825 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
826 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
829 assert(!AS.empty() &&
830 "Must alias set should have at least one pointer element in it!");
832 Value *SomePtr = AS.begin()->getValue();
833 BasicBlock * Preheader = CurLoop->getLoopPreheader();
835 // It isn't safe to promote a load/store from the loop if the load/store is
836 // conditional. For example, turning:
838 // for () { if (c) *P += 1; }
842 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
844 // is not safe, because *P may only be valid to access if 'c' is true.
846 // It is safe to promote P if all uses are direct load/stores and if at
847 // least one is guaranteed to be executed.
848 bool GuaranteedToExecute = false;
850 SmallVector<Instruction*, 64> LoopUses;
851 SmallPtrSet<Value*, 4> PointerMustAliases;
853 // We start with an alignment of one and try to find instructions that allow
854 // us to prove better alignment.
855 unsigned Alignment = 1;
857 bool HasDedicatedExits = CurLoop->hasDedicatedExits();
859 // Check that all of the pointers in the alias set have the same type. We
860 // cannot (yet) promote a memory location that is loaded and stored in
861 // different sizes. While we are at it, collect alignment and AA info.
862 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
863 Value *ASIV = ASI->getValue();
864 PointerMustAliases.insert(ASIV);
866 // Check that all of the pointers in the alias set have the same type. We
867 // cannot (yet) promote a memory location that is loaded and stored in
869 if (SomePtr->getType() != ASIV->getType())
872 for (User *U : ASIV->users()) {
873 // Ignore instructions that are outside the loop.
874 Instruction *UI = dyn_cast<Instruction>(U);
875 if (!UI || !CurLoop->contains(UI))
878 // If there is an non-load/store instruction in the loop, we can't promote
880 if (const LoadInst *load = dyn_cast<LoadInst>(UI)) {
881 assert(!load->isVolatile() && "AST broken");
882 if (!load->isSimple())
884 } else if (const StoreInst *store = dyn_cast<StoreInst>(UI)) {
885 // Stores *of* the pointer are not interesting, only stores *to* the
887 if (UI->getOperand(1) != ASIV)
889 assert(!store->isVolatile() && "AST broken");
890 if (!store->isSimple())
892 // Don't sink stores from loops without dedicated block exits. Exits
893 // containing indirect branches are not transformed by loop simplify,
894 // make sure we catch that. An additional load may be generated in the
895 // preheader for SSA updater, so also avoid sinking when no preheader
897 if (!HasDedicatedExits || !Preheader)
900 // Note that we only check GuaranteedToExecute inside the store case
901 // so that we do not introduce stores where they did not exist before
902 // (which would break the LLVM concurrency model).
904 // If the alignment of this instruction allows us to specify a more
905 // restrictive (and performant) alignment and if we are sure this
906 // instruction will be executed, update the alignment.
907 // Larger is better, with the exception of 0 being the best alignment.
908 unsigned InstAlignment = store->getAlignment();
909 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
910 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
911 GuaranteedToExecute = true;
912 Alignment = InstAlignment;
915 if (!GuaranteedToExecute)
916 GuaranteedToExecute = isGuaranteedToExecute(*UI, DT,
917 CurLoop, SafetyInfo);
920 return Changed; // Not a load or store.
922 // Merge the AA tags.
923 if (LoopUses.empty()) {
924 // On the first load/store, just take its AA tags.
925 UI->getAAMetadata(AATags);
927 UI->getAAMetadata(AATags, /* Merge = */ true);
930 LoopUses.push_back(UI);
934 // If there isn't a guaranteed-to-execute instruction, we can't promote.
935 if (!GuaranteedToExecute)
938 // Otherwise, this is safe to promote, lets do it!
939 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
943 // Grab a debug location for the inserted loads/stores; given that the
944 // inserted loads/stores have little relation to the original loads/stores,
945 // this code just arbitrarily picks a location from one, since any debug
946 // location is better than none.
947 DebugLoc DL = LoopUses[0]->getDebugLoc();
949 // Figure out the loop exits and their insertion points, if this is the
951 if (ExitBlocks.empty()) {
952 CurLoop->getUniqueExitBlocks(ExitBlocks);
953 InsertPts.resize(ExitBlocks.size());
954 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
955 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
958 // We use the SSAUpdater interface to insert phi nodes as required.
959 SmallVector<PHINode*, 16> NewPHIs;
960 SSAUpdater SSA(&NewPHIs);
961 LoopPromoter Promoter(SomePtr, LoopUses, SSA,
962 PointerMustAliases, ExitBlocks,
963 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
965 // Set up the preheader to have a definition of the value. It is the live-out
966 // value from the preheader that uses in the loop will use.
967 LoadInst *PreheaderLoad =
968 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
969 Preheader->getTerminator());
970 PreheaderLoad->setAlignment(Alignment);
971 PreheaderLoad->setDebugLoc(DL);
972 if (AATags) PreheaderLoad->setAAMetadata(AATags);
973 SSA.AddAvailableValue(Preheader, PreheaderLoad);
975 // Rewrite all the loads in the loop and remember all the definitions from
976 // stores in the loop.
977 Promoter.run(LoopUses);
979 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
980 if (PreheaderLoad->use_empty())
981 PreheaderLoad->eraseFromParent();
986 /// Simple analysis hook. Clone alias set info.
988 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
989 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
993 AST->copyValue(From, To);
996 /// Simple Analysis hook. Delete value V from alias set
998 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
999 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1003 AST->deleteValue(V);
1006 /// Simple Analysis hook. Delete value L from alias set map.
1008 void LICM::deleteAnalysisLoop(Loop *L) {
1009 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
1014 LoopToAliasSetMap.erase(L);
1018 /// Return true if the body of this loop may store into the memory
1019 /// location pointed to by V.
1021 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
1022 const AAMDNodes &AAInfo,
1023 AliasSetTracker *CurAST) {
1024 // Check to see if any of the basic blocks in CurLoop invalidate *V.
1025 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
1028 /// Little predicate that returns true if the specified basic block is in
1029 /// a subloop of the current one, not the current one itself.
1031 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1032 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1033 return LI->getLoopFor(BB) != CurLoop;