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/ConstantFolding.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Analysis/ScalarEvolution.h"
41 #include "llvm/Analysis/TargetLibraryInfo.h"
42 #include "llvm/Analysis/ValueTracking.h"
43 #include "llvm/IR/CFG.h"
44 #include "llvm/IR/Constants.h"
45 #include "llvm/IR/DataLayout.h"
46 #include "llvm/IR/DerivedTypes.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/Metadata.h"
52 #include "llvm/IR/PredIteratorCache.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/LoopUtils.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
62 #define DEBUG_TYPE "licm"
64 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
65 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
66 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
67 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
68 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
71 DisablePromotion("disable-licm-promotion", cl::Hidden,
72 cl::desc("Disable memory promotion in LICM pass"));
74 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
75 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop);
76 static bool hoist(Instruction &I, BasicBlock *Preheader);
77 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
78 const Loop *CurLoop, AliasSetTracker *CurAST );
79 static bool isGuaranteedToExecute(const Instruction &Inst,
80 const DominatorTree *DT,
82 const LICMSafetyInfo *SafetyInfo);
83 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
84 const DominatorTree *DT,
86 const LICMSafetyInfo *SafetyInfo);
87 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
88 const AAMDNodes &AAInfo,
89 AliasSetTracker *CurAST);
90 static Instruction *CloneInstructionInExitBlock(const Instruction &I,
91 BasicBlock &ExitBlock,
94 static bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA,
95 DominatorTree *DT, Loop *CurLoop,
96 AliasSetTracker *CurAST,
97 LICMSafetyInfo *SafetyInfo);
100 struct LICM : public LoopPass {
101 static char ID; // Pass identification, replacement for typeid
102 LICM() : LoopPass(ID) {
103 initializeLICMPass(*PassRegistry::getPassRegistry());
106 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
108 /// This transformation requires natural loop information & requires that
109 /// loop preheaders be inserted into the CFG...
111 void getAnalysisUsage(AnalysisUsage &AU) const override {
112 AU.setPreservesCFG();
113 AU.addRequired<DominatorTreeWrapperPass>();
114 AU.addRequired<LoopInfoWrapperPass>();
115 AU.addRequiredID(LoopSimplifyID);
116 AU.addPreservedID(LoopSimplifyID);
117 AU.addRequiredID(LCSSAID);
118 AU.addPreservedID(LCSSAID);
119 AU.addRequired<AliasAnalysis>();
120 AU.addPreserved<AliasAnalysis>();
121 AU.addPreserved<ScalarEvolution>();
122 AU.addRequired<TargetLibraryInfoWrapperPass>();
125 using llvm::Pass::doFinalization;
127 bool doFinalization() override {
128 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
133 AliasAnalysis *AA; // Current AliasAnalysis information
134 LoopInfo *LI; // Current LoopInfo
135 DominatorTree *DT; // Dominator Tree for the current Loop.
137 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
139 // State that is updated as we process loops.
140 bool Changed; // Set to true when we change anything.
141 BasicBlock *Preheader; // The preheader block of the current loop...
142 Loop *CurLoop; // The current loop we are working on...
143 AliasSetTracker *CurAST; // AliasSet information for the current loop...
144 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
146 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
147 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
150 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
152 void deleteAnalysisValue(Value *V, Loop *L) override;
154 /// Simple Analysis hook. Delete loop L from alias set map.
155 void deleteAnalysisLoop(Loop *L) override;
160 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
161 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
162 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
163 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
164 INITIALIZE_PASS_DEPENDENCY(LCSSA)
165 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
166 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
167 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
168 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
170 Pass *llvm::createLICMPass() { return new LICM(); }
172 /// Hoist expressions out of the specified loop. Note, alias info for inner
173 /// loop is not preserved so it is not a good idea to run LICM multiple
174 /// times on one loop.
176 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
177 if (skipOptnoneFunction(L))
182 // Get our Loop and Alias Analysis information...
183 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
184 AA = &getAnalysis<AliasAnalysis>();
185 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
187 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
189 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
191 CurAST = new AliasSetTracker(*AA);
192 // Collect Alias info from subloops.
193 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
194 LoopItr != LoopItrE; ++LoopItr) {
195 Loop *InnerL = *LoopItr;
196 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
197 assert(InnerAST && "Where is my AST?");
199 // What if InnerLoop was modified by other passes ?
200 CurAST->add(*InnerAST);
202 // Once we've incorporated the inner loop's AST into ours, we don't need the
203 // subloop's anymore.
205 LoopToAliasSetMap.erase(InnerL);
210 // Get the preheader block to move instructions into...
211 Preheader = L->getLoopPreheader();
213 // Loop over the body of this loop, looking for calls, invokes, and stores.
214 // Because subloops have already been incorporated into AST, we skip blocks in
217 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
220 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
221 CurAST->add(*BB); // Incorporate the specified basic block
224 // Compute loop safety information.
225 LICMSafetyInfo SafetyInfo;
226 computeLICMSafetyInfo(&SafetyInfo, CurLoop);
228 // We want to visit all of the instructions in this loop... that are not parts
229 // of our subloops (they have already had their invariants hoisted out of
230 // their loop, into this loop, so there is no need to process the BODIES of
233 // Traverse the body of the loop in depth first order on the dominator tree so
234 // that we are guaranteed to see definitions before we see uses. This allows
235 // us to sink instructions in one pass, without iteration. After sinking
236 // instructions, we perform another pass to hoist them out of the loop.
238 if (L->hasDedicatedExits())
239 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, CurLoop,
240 CurAST, &SafetyInfo);
242 Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI,
243 CurLoop, CurAST, &SafetyInfo);
245 // Now that all loop invariants have been removed from the loop, promote any
246 // memory references to scalars that we can.
247 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
248 SmallVector<BasicBlock *, 8> ExitBlocks;
249 SmallVector<Instruction *, 8> InsertPts;
250 PredIteratorCache PIC;
252 // Loop over all of the alias sets in the tracker object.
253 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
255 Changed |= promoteLoopAccessesToScalars(*I, ExitBlocks, InsertPts,
256 PIC, LI, DT, CurLoop,
257 CurAST, &SafetyInfo);
259 // Once we have promoted values across the loop body we have to recursively
260 // reform LCSSA as any nested loop may now have values defined within the
261 // loop used in the outer loop.
262 // FIXME: This is really heavy handed. It would be a bit better to use an
263 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
266 formLCSSARecursively(*L, *DT, LI,
267 getAnalysisIfAvailable<ScalarEvolution>());
270 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
271 // specifically moving instructions across the loop boundary and so it is
272 // especially in need of sanity checking here.
273 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
274 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
275 "Parent loop not left in LCSSA form after LICM!");
277 // Clear out loops state information for the next iteration
281 // If this loop is nested inside of another one, save the alias information
282 // for when we process the outer loop.
283 if (L->getParentLoop())
284 LoopToAliasSetMap[L] = CurAST;
290 /// Walk the specified region of the CFG (defined by all blocks dominated by
291 /// the specified block, and that are in the current loop) in reverse depth
292 /// first order w.r.t the DominatorTree. This allows us to visit uses before
293 /// definitions, allowing us to sink a loop body in one pass without iteration.
295 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
296 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
297 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
300 assert(N != nullptr && AA != nullptr && LI != nullptr &&
301 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
302 SafetyInfo != nullptr && "Unexpected input to sinkRegion");
304 // Set changed as false.
305 bool Changed = false;
307 BasicBlock *BB = N->getBlock();
308 // If this subregion is not in the top level loop at all, exit.
309 if (!CurLoop->contains(BB)) return Changed;
311 // We are processing blocks in reverse dfo, so process children first.
312 const std::vector<DomTreeNode*> &Children = N->getChildren();
313 for (unsigned i = 0, e = Children.size(); i != e; ++i)
315 sinkRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
316 // Only need to process the contents of this block if it is not part of a
317 // subloop (which would already have been processed).
318 if (inSubLoop(BB,CurLoop,LI)) return Changed;
320 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
321 Instruction &I = *--II;
323 // If the instruction is dead, we would try to sink it because it isn't used
324 // in the loop, instead, just delete it.
325 if (isInstructionTriviallyDead(&I, TLI)) {
326 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
328 CurAST->deleteValue(&I);
334 // Check to see if we can sink this instruction to the exit blocks
335 // of the loop. We can do this if the all users of the instruction are
336 // outside of the loop. In this case, it doesn't even matter if the
337 // operands of the instruction are loop invariant.
339 if (isNotUsedInLoop(I, CurLoop) &&
340 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo)) {
342 Changed |= sink(I, LI, DT, CurLoop, CurAST);
348 /// Walk the specified region of the CFG (defined by all blocks dominated by
349 /// the specified block, and that are in the current loop) in depth first
350 /// order w.r.t the DominatorTree. This allows us to visit definitions before
351 /// uses, allowing us to hoist a loop body in one pass without iteration.
353 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
354 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
355 AliasSetTracker *CurAST, LICMSafetyInfo *SafetyInfo) {
357 assert(N != nullptr && AA != nullptr && LI != nullptr &&
358 DT != nullptr && CurLoop != nullptr && CurAST != nullptr &&
359 SafetyInfo != nullptr && "Unexpected input to hoistRegion");
360 // Set changed as false.
361 bool Changed = false;
363 BasicBlock *BB = N->getBlock();
364 // If this subregion is not in the top level loop at all, exit.
365 if (!CurLoop->contains(BB)) return Changed;
366 // Only need to process the contents of this block if it is not part of a
367 // subloop (which would already have been processed).
368 if (!inSubLoop(BB, CurLoop, LI))
369 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
370 Instruction &I = *II++;
371 // Try constant folding this instruction. If all the operands are
372 // constants, it is technically hoistable, but it would be better to just
374 if (Constant *C = ConstantFoldInstruction(
375 &I, I.getModule()->getDataLayout(), TLI)) {
376 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
377 CurAST->copyValue(&I, C);
378 CurAST->deleteValue(&I);
379 I.replaceAllUsesWith(C);
384 // Try hoisting the instruction out to the preheader. We can only do this
385 // if all of the operands of the instruction are loop invariant and if it
386 // is safe to hoist the instruction.
388 if (CurLoop->hasLoopInvariantOperands(&I) &&
389 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, SafetyInfo) &&
390 isSafeToExecuteUnconditionally(I, DT, CurLoop, SafetyInfo))
391 Changed |= hoist(I, CurLoop->getLoopPreheader());
394 const std::vector<DomTreeNode*> &Children = N->getChildren();
395 for (unsigned i = 0, e = Children.size(); i != e; ++i)
397 hoistRegion(Children[i], AA, LI, DT, TLI, CurLoop, CurAST, SafetyInfo);
401 /// Computes loop safety information, checks loop body & header
402 /// for the possiblity of may throw exception.
404 void llvm::computeLICMSafetyInfo(LICMSafetyInfo * SafetyInfo, Loop * CurLoop) {
405 assert(CurLoop != nullptr && "CurLoop cant be null");
406 BasicBlock *Header = CurLoop->getHeader();
407 // Setting default safety values.
408 SafetyInfo->MayThrow = false;
409 SafetyInfo->HeaderMayThrow = false;
410 // Iterate over header and compute dafety info.
411 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
412 (I != E) && !SafetyInfo->HeaderMayThrow; ++I)
413 SafetyInfo->HeaderMayThrow |= I->mayThrow();
415 SafetyInfo->MayThrow = SafetyInfo->HeaderMayThrow;
416 // Iterate over loop instructions and compute safety info.
417 for (Loop::block_iterator BB = CurLoop->block_begin(),
418 BBE = CurLoop->block_end(); (BB != BBE) && !SafetyInfo->MayThrow ; ++BB)
419 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
420 (I != E) && !SafetyInfo->MayThrow; ++I)
421 SafetyInfo->MayThrow |= I->mayThrow();
424 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
427 bool canSinkOrHoistInst(Instruction &I, AliasAnalysis *AA, DominatorTree *DT,
428 Loop *CurLoop, AliasSetTracker *CurAST,
429 LICMSafetyInfo *SafetyInfo) {
430 // Loads have extra constraints we have to verify before we can hoist them.
431 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
432 if (!LI->isUnordered())
433 return false; // Don't hoist volatile/atomic loads!
435 // Loads from constant memory are always safe to move, even if they end up
436 // in the same alias set as something that ends up being modified.
437 if (AA->pointsToConstantMemory(LI->getOperand(0)))
439 if (LI->getMetadata(LLVMContext::MD_invariant_load))
442 // Don't hoist loads which have may-aliased stores in loop.
444 if (LI->getType()->isSized())
445 Size = AA->getTypeStoreSize(LI->getType());
448 LI->getAAMetadata(AAInfo);
450 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo, CurAST);
451 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
452 // Don't sink or hoist dbg info; it's legal, but not useful.
453 if (isa<DbgInfoIntrinsic>(I))
456 // Handle simple cases by querying alias analysis.
457 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
458 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
460 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
461 // If this call only reads from memory and there are no writes to memory
462 // in the loop, we can hoist or sink the call as appropriate.
463 bool FoundMod = false;
464 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
467 if (!AS.isForwardingAliasSet() && AS.isMod()) {
472 if (!FoundMod) return true;
475 // FIXME: This should use mod/ref information to see if we can hoist or
481 // Only these instructions are hoistable/sinkable.
482 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
483 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
484 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
485 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
486 !isa<InsertValueInst>(I))
489 return isSafeToExecuteUnconditionally(I, DT, CurLoop, SafetyInfo);
492 /// Returns true if a PHINode is a trivially replaceable with an
494 /// This is true when all incoming values are that instruction.
495 /// This pattern occurs most often with LCSSA PHI nodes.
497 static bool isTriviallyReplacablePHI(const PHINode &PN, const Instruction &I) {
498 for (const Value *IncValue : PN.incoming_values())
505 /// Return true if the only users of this instruction are outside of
506 /// the loop. If this is true, we can sink the instruction to the exit
507 /// blocks of the loop.
509 static bool isNotUsedInLoop(const Instruction &I, const Loop *CurLoop) {
510 for (const User *U : I.users()) {
511 const Instruction *UI = cast<Instruction>(U);
512 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
513 // A PHI node where all of the incoming values are this instruction are
514 // special -- they can just be RAUW'ed with the instruction and thus
515 // don't require a use in the predecessor. This is a particular important
516 // special case because it is the pattern found in LCSSA form.
517 if (isTriviallyReplacablePHI(*PN, I)) {
518 if (CurLoop->contains(PN))
524 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
525 // values. Check for such a use being inside the loop.
526 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
527 if (PN->getIncomingValue(i) == &I)
528 if (CurLoop->contains(PN->getIncomingBlock(i)))
534 if (CurLoop->contains(UI))
540 static Instruction *CloneInstructionInExitBlock(const Instruction &I,
541 BasicBlock &ExitBlock,
543 const LoopInfo *LI) {
544 Instruction *New = I.clone();
545 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
546 if (!I.getName().empty()) New->setName(I.getName() + ".le");
548 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
549 // particularly cheap because we can rip off the PHI node that we're
550 // replacing for the number and blocks of the predecessors.
551 // OPT: If this shows up in a profile, we can instead finish sinking all
552 // invariant instructions, and then walk their operands to re-establish
553 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
554 // sinking bottom-up.
555 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
557 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
558 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
559 if (!OLoop->contains(&PN)) {
561 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
562 OInst->getName() + ".lcssa", ExitBlock.begin());
563 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
564 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
570 /// When an instruction is found to only be used outside of the loop, this
571 /// function moves it to the exit blocks and patches up SSA form as needed.
572 /// This method is guaranteed to remove the original instruction from its
573 /// position, and may either delete it or move it to outside of the loop.
575 static bool sink(Instruction &I, const LoopInfo *LI, const DominatorTree *DT,
576 const Loop *CurLoop, AliasSetTracker *CurAST ) {
577 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
578 bool Changed = false;
579 if (isa<LoadInst>(I)) ++NumMovedLoads;
580 else if (isa<CallInst>(I)) ++NumMovedCalls;
585 SmallVector<BasicBlock *, 32> ExitBlocks;
586 CurLoop->getUniqueExitBlocks(ExitBlocks);
587 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
591 // Clones of this instruction. Don't create more than one per exit block!
592 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
594 // If this instruction is only used outside of the loop, then all users are
595 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
597 while (!I.use_empty()) {
598 Instruction *User = I.user_back();
599 if (!DT->isReachableFromEntry(User->getParent())) {
600 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
603 // The user must be a PHI node.
604 PHINode *PN = cast<PHINode>(User);
606 BasicBlock *ExitBlock = PN->getParent();
607 assert(ExitBlockSet.count(ExitBlock) &&
608 "The LCSSA PHI is not in an exit block!");
611 auto It = SunkCopies.find(ExitBlock);
612 if (It != SunkCopies.end())
615 New = SunkCopies[ExitBlock] =
616 CloneInstructionInExitBlock(I, *ExitBlock, *PN, LI);
618 PN->replaceAllUsesWith(New);
619 PN->eraseFromParent();
622 CurAST->deleteValue(&I);
627 /// When an instruction is found to only use loop invariant operands that
628 /// is safe to hoist, this instruction is called to do the dirty work.
630 static bool hoist(Instruction &I, BasicBlock *Preheader) {
631 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
633 // Move the new node to the Preheader, before its terminator.
634 I.moveBefore(Preheader->getTerminator());
636 if (isa<LoadInst>(I)) ++NumMovedLoads;
637 else if (isa<CallInst>(I)) ++NumMovedCalls;
642 /// Only sink or hoist an instruction if it is not a trapping instruction
643 /// or if it is a trapping instruction and is guaranteed to execute.
645 static bool isSafeToExecuteUnconditionally(const Instruction &Inst,
646 const DominatorTree *DT,
648 const LICMSafetyInfo *SafetyInfo) {
649 // If it is not a trapping instruction, it is always safe to hoist.
650 if (isSafeToSpeculativelyExecute(&Inst))
653 return isGuaranteedToExecute(Inst, DT, CurLoop, SafetyInfo);
656 static bool isGuaranteedToExecute(const Instruction &Inst,
657 const DominatorTree *DT,
659 const LICMSafetyInfo * SafetyInfo) {
661 // We have to check to make sure that the instruction dominates all
662 // of the exit blocks. If it doesn't, then there is a path out of the loop
663 // which does not execute this instruction, so we can't hoist it.
665 // If the instruction is in the header block for the loop (which is very
666 // common), it is always guaranteed to dominate the exit blocks. Since this
667 // is a common case, and can save some work, check it now.
668 if (Inst.getParent() == CurLoop->getHeader())
669 // If there's a throw in the header block, we can't guarantee we'll reach
671 return !SafetyInfo->HeaderMayThrow;
673 // Somewhere in this loop there is an instruction which may throw and make us
675 if (SafetyInfo->MayThrow)
678 // Get the exit blocks for the current loop.
679 SmallVector<BasicBlock*, 8> ExitBlocks;
680 CurLoop->getExitBlocks(ExitBlocks);
682 // Verify that the block dominates each of the exit blocks of the loop.
683 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
684 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
687 // As a degenerate case, if the loop is statically infinite then we haven't
688 // proven anything since there are no exit blocks.
689 if (ExitBlocks.empty())
696 class LoopPromoter : public LoadAndStorePromoter {
697 Value *SomePtr; // Designated pointer to store to.
698 SmallPtrSetImpl<Value*> &PointerMustAliases;
699 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
700 SmallVectorImpl<Instruction*> &LoopInsertPts;
701 PredIteratorCache &PredCache;
702 AliasSetTracker &AST;
708 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
709 if (Instruction *I = dyn_cast<Instruction>(V))
710 if (Loop *L = LI.getLoopFor(I->getParent()))
711 if (!L->contains(BB)) {
712 // We need to create an LCSSA PHI node for the incoming value and
714 PHINode *PN = PHINode::Create(
715 I->getType(), PredCache.size(BB),
716 I->getName() + ".lcssa", BB->begin());
717 for (BasicBlock *Pred : PredCache.get(BB))
718 PN->addIncoming(I, Pred);
725 LoopPromoter(Value *SP,
726 ArrayRef<const Instruction *> Insts,
727 SSAUpdater &S, SmallPtrSetImpl<Value *> &PMA,
728 SmallVectorImpl<BasicBlock *> &LEB,
729 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
730 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
731 const AAMDNodes &AATags)
732 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
733 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
734 LI(li), DL(dl), Alignment(alignment), AATags(AATags) {}
736 bool isInstInList(Instruction *I,
737 const SmallVectorImpl<Instruction*> &) const override {
739 if (LoadInst *LI = dyn_cast<LoadInst>(I))
740 Ptr = LI->getOperand(0);
742 Ptr = cast<StoreInst>(I)->getPointerOperand();
743 return PointerMustAliases.count(Ptr);
746 void doExtraRewritesBeforeFinalDeletion() const override {
747 // Insert stores after in the loop exit blocks. Each exit block gets a
748 // store of the live-out values that feed them. Since we've already told
749 // the SSA updater about the defs in the loop and the preheader
750 // definition, it is all set and we can start using it.
751 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
752 BasicBlock *ExitBlock = LoopExitBlocks[i];
753 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
754 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
755 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
756 Instruction *InsertPos = LoopInsertPts[i];
757 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
758 NewSI->setAlignment(Alignment);
759 NewSI->setDebugLoc(DL);
760 if (AATags) NewSI->setAAMetadata(AATags);
764 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
765 // Update alias analysis.
766 AST.copyValue(LI, V);
768 void instructionDeleted(Instruction *I) const override {
772 } // end anon namespace
774 /// Try to promote memory values to scalars by sinking stores out of the
775 /// loop and moving loads to before the loop. We do this by looping over
776 /// the stores in the loop, looking for stores to Must pointers which are
779 bool llvm::promoteLoopAccessesToScalars(AliasSet &AS,
780 SmallVectorImpl<BasicBlock*>&ExitBlocks,
781 SmallVectorImpl<Instruction*>&InsertPts,
782 PredIteratorCache &PIC, LoopInfo *LI,
783 DominatorTree *DT, Loop *CurLoop,
784 AliasSetTracker *CurAST,
785 LICMSafetyInfo * SafetyInfo) {
787 assert(LI != nullptr && DT != nullptr &&
788 CurLoop != nullptr && CurAST != nullptr &&
789 SafetyInfo != nullptr &&
790 "Unexpected Input to promoteLoopAccessesToScalars");
791 // Initially set Changed status to false.
792 bool Changed = false;
793 // We can promote this alias set if it has a store, if it is a "Must" alias
794 // set, if the pointer is loop invariant, and if we are not eliminating any
795 // volatile loads or stores.
796 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
797 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
800 assert(!AS.empty() &&
801 "Must alias set should have at least one pointer element in it!");
803 Value *SomePtr = AS.begin()->getValue();
804 BasicBlock * Preheader = CurLoop->getLoopPreheader();
806 // It isn't safe to promote a load/store from the loop if the load/store is
807 // conditional. For example, turning:
809 // for () { if (c) *P += 1; }
813 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
815 // is not safe, because *P may only be valid to access if 'c' is true.
817 // It is safe to promote P if all uses are direct load/stores and if at
818 // least one is guaranteed to be executed.
819 bool GuaranteedToExecute = false;
821 SmallVector<Instruction*, 64> LoopUses;
822 SmallPtrSet<Value*, 4> PointerMustAliases;
824 // We start with an alignment of one and try to find instructions that allow
825 // us to prove better alignment.
826 unsigned Alignment = 1;
828 bool HasDedicatedExits = CurLoop->hasDedicatedExits();
830 // Check that all of the pointers in the alias set have the same type. We
831 // cannot (yet) promote a memory location that is loaded and stored in
832 // different sizes. While we are at it, collect alignment and AA info.
833 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
834 Value *ASIV = ASI->getValue();
835 PointerMustAliases.insert(ASIV);
837 // Check that all of the pointers in the alias set have the same type. We
838 // cannot (yet) promote a memory location that is loaded and stored in
840 if (SomePtr->getType() != ASIV->getType())
843 for (User *U : ASIV->users()) {
844 // Ignore instructions that are outside the loop.
845 Instruction *UI = dyn_cast<Instruction>(U);
846 if (!UI || !CurLoop->contains(UI))
849 // If there is an non-load/store instruction in the loop, we can't promote
851 if (const LoadInst *load = dyn_cast<LoadInst>(UI)) {
852 assert(!load->isVolatile() && "AST broken");
853 if (!load->isSimple())
855 } else if (const StoreInst *store = dyn_cast<StoreInst>(UI)) {
856 // Stores *of* the pointer are not interesting, only stores *to* the
858 if (UI->getOperand(1) != ASIV)
860 assert(!store->isVolatile() && "AST broken");
861 if (!store->isSimple())
863 // Don't sink stores from loops without dedicated block exits. Exits
864 // containing indirect branches are not transformed by loop simplify,
865 // make sure we catch that. An additional load may be generated in the
866 // preheader for SSA updater, so also avoid sinking when no preheader
868 if (!HasDedicatedExits || !Preheader)
871 // Note that we only check GuaranteedToExecute inside the store case
872 // so that we do not introduce stores where they did not exist before
873 // (which would break the LLVM concurrency model).
875 // If the alignment of this instruction allows us to specify a more
876 // restrictive (and performant) alignment and if we are sure this
877 // instruction will be executed, update the alignment.
878 // Larger is better, with the exception of 0 being the best alignment.
879 unsigned InstAlignment = store->getAlignment();
880 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
881 if (isGuaranteedToExecute(*UI, DT, CurLoop, SafetyInfo)) {
882 GuaranteedToExecute = true;
883 Alignment = InstAlignment;
886 if (!GuaranteedToExecute)
887 GuaranteedToExecute = isGuaranteedToExecute(*UI, DT,
888 CurLoop, SafetyInfo);
891 return Changed; // Not a load or store.
893 // Merge the AA tags.
894 if (LoopUses.empty()) {
895 // On the first load/store, just take its AA tags.
896 UI->getAAMetadata(AATags);
898 UI->getAAMetadata(AATags, /* Merge = */ true);
901 LoopUses.push_back(UI);
905 // If there isn't a guaranteed-to-execute instruction, we can't promote.
906 if (!GuaranteedToExecute)
909 // Otherwise, this is safe to promote, lets do it!
910 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
914 // Grab a debug location for the inserted loads/stores; given that the
915 // inserted loads/stores have little relation to the original loads/stores,
916 // this code just arbitrarily picks a location from one, since any debug
917 // location is better than none.
918 DebugLoc DL = LoopUses[0]->getDebugLoc();
920 // Figure out the loop exits and their insertion points, if this is the
922 if (ExitBlocks.empty()) {
923 CurLoop->getUniqueExitBlocks(ExitBlocks);
924 InsertPts.resize(ExitBlocks.size());
925 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
926 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
929 // We use the SSAUpdater interface to insert phi nodes as required.
930 SmallVector<PHINode*, 16> NewPHIs;
931 SSAUpdater SSA(&NewPHIs);
932 LoopPromoter Promoter(SomePtr, LoopUses, SSA,
933 PointerMustAliases, ExitBlocks,
934 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
936 // Set up the preheader to have a definition of the value. It is the live-out
937 // value from the preheader that uses in the loop will use.
938 LoadInst *PreheaderLoad =
939 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
940 Preheader->getTerminator());
941 PreheaderLoad->setAlignment(Alignment);
942 PreheaderLoad->setDebugLoc(DL);
943 if (AATags) PreheaderLoad->setAAMetadata(AATags);
944 SSA.AddAvailableValue(Preheader, PreheaderLoad);
946 // Rewrite all the loads in the loop and remember all the definitions from
947 // stores in the loop.
948 Promoter.run(LoopUses);
950 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
951 if (PreheaderLoad->use_empty())
952 PreheaderLoad->eraseFromParent();
957 /// Simple Analysis hook. Clone alias set info.
959 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
960 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
964 AST->copyValue(From, To);
967 /// Simple Analysis hook. Delete value V from alias set
969 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
970 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
977 /// Simple Analysis hook. Delete value L from alias set map.
979 void LICM::deleteAnalysisLoop(Loop *L) {
980 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
985 LoopToAliasSetMap.erase(L);
989 /// Return true if the body of this loop may store into the memory
990 /// location pointed to by V.
992 static bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
993 const AAMDNodes &AAInfo,
994 AliasSetTracker *CurAST) {
995 // Check to see if any of the basic blocks in CurLoop invalidate *V.
996 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
999 /// Little predicate that returns true if the specified basic block is in
1000 /// a subloop of the current one, not the current one itself.
1002 static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
1003 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
1004 return LI->getLoopFor(BB) != CurLoop;