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/ValueTracking.h"
42 #include "llvm/IR/CFG.h"
43 #include "llvm/IR/Constants.h"
44 #include "llvm/IR/DataLayout.h"
45 #include "llvm/IR/DerivedTypes.h"
46 #include "llvm/IR/Dominators.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/PredIteratorCache.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Target/TargetLibraryInfo.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"));
75 struct LICM : public LoopPass {
76 static char ID; // Pass identification, replacement for typeid
77 LICM() : LoopPass(ID) {
78 initializeLICMPass(*PassRegistry::getPassRegistry());
81 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
83 /// This transformation requires natural loop information & requires that
84 /// loop preheaders be inserted into the CFG...
86 void getAnalysisUsage(AnalysisUsage &AU) const override {
88 AU.addRequired<DominatorTreeWrapperPass>();
89 AU.addRequired<LoopInfo>();
90 AU.addRequiredID(LoopSimplifyID);
91 AU.addPreservedID(LoopSimplifyID);
92 AU.addRequiredID(LCSSAID);
93 AU.addPreservedID(LCSSAID);
94 AU.addRequired<AliasAnalysis>();
95 AU.addPreserved<AliasAnalysis>();
96 AU.addPreserved<ScalarEvolution>();
97 AU.addRequired<TargetLibraryInfo>();
100 using llvm::Pass::doFinalization;
102 bool doFinalization() override {
103 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
108 AliasAnalysis *AA; // Current AliasAnalysis information
109 LoopInfo *LI; // Current LoopInfo
110 DominatorTree *DT; // Dominator Tree for the current Loop.
112 const DataLayout *DL; // DataLayout for constant folding.
113 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
115 // State that is updated as we process loops.
116 bool Changed; // Set to true when we change anything.
117 BasicBlock *Preheader; // The preheader block of the current loop...
118 Loop *CurLoop; // The current loop we are working on...
119 AliasSetTracker *CurAST; // AliasSet information for the current loop...
120 bool MayThrow; // The current loop contains an instruction which
121 // may throw, thus preventing code motion of
122 // instructions with side effects.
123 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
125 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
126 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
129 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
131 void deleteAnalysisValue(Value *V, Loop *L) override;
133 /// Simple Analysis hook. Delete loop L from alias set map.
134 void deleteAnalysisLoop(Loop *L) override;
136 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
137 /// dominated by the specified block, and that are in the current loop) in
138 /// reverse depth first order w.r.t the DominatorTree. This allows us to
139 /// visit uses before definitions, allowing us to sink a loop body in one
140 /// pass without iteration.
142 void SinkRegion(DomTreeNode *N);
144 /// HoistRegion - Walk the specified region of the CFG (defined by all
145 /// blocks dominated by the specified block, and that are in the current
146 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
147 /// visit definitions before uses, allowing us to hoist a loop body in one
148 /// pass without iteration.
150 void HoistRegion(DomTreeNode *N);
152 /// inSubLoop - Little predicate that returns true if the specified basic
153 /// block is in a subloop of the current one, not the current one itself.
155 bool inSubLoop(BasicBlock *BB) {
156 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
157 return LI->getLoopFor(BB) != CurLoop;
160 /// sink - When an instruction is found to only be used outside of the loop,
161 /// this function moves it to the exit blocks and patches up SSA form as
164 void sink(Instruction &I);
166 /// hoist - When an instruction is found to only use loop invariant operands
167 /// that is safe to hoist, this instruction is called to do the dirty work.
169 void hoist(Instruction &I);
171 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
172 /// is not a trapping instruction or if it is a trapping instruction and is
173 /// guaranteed to execute.
175 bool isSafeToExecuteUnconditionally(Instruction &I);
177 /// isGuaranteedToExecute - Check that the instruction is guaranteed to
180 bool isGuaranteedToExecute(Instruction &I);
182 /// pointerInvalidatedByLoop - Return true if the body of this loop may
183 /// store into the memory location pointed to by V.
185 bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
186 const AAMDNodes &AAInfo) {
187 // Check to see if any of the basic blocks in CurLoop invalidate *V.
188 return CurAST->getAliasSetForPointer(V, Size, AAInfo).isMod();
191 bool canSinkOrHoistInst(Instruction &I);
192 bool isNotUsedInLoop(Instruction &I);
194 void PromoteAliasSet(AliasSet &AS,
195 SmallVectorImpl<BasicBlock*> &ExitBlocks,
196 SmallVectorImpl<Instruction*> &InsertPts,
197 PredIteratorCache &PIC);
199 /// \brief Create a copy of the instruction in the exit block and patch up
201 /// PN is a user of I in ExitBlock that can be used to get the number and
202 /// list of predecessors fast.
203 Instruction *CloneInstructionInExitBlock(Instruction &I,
204 BasicBlock &ExitBlock,
210 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
211 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
212 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
213 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
214 INITIALIZE_PASS_DEPENDENCY(LCSSA)
215 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
216 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
217 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
218 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
220 Pass *llvm::createLICMPass() { return new LICM(); }
222 /// Hoist expressions out of the specified loop. Note, alias info for inner
223 /// loop is not preserved so it is not a good idea to run LICM multiple
224 /// times on one loop.
226 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
227 if (skipOptnoneFunction(L))
232 // Get our Loop and Alias Analysis information...
233 LI = &getAnalysis<LoopInfo>();
234 AA = &getAnalysis<AliasAnalysis>();
235 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
237 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
238 DL = DLP ? &DLP->getDataLayout() : nullptr;
239 TLI = &getAnalysis<TargetLibraryInfo>();
241 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
243 CurAST = new AliasSetTracker(*AA);
244 // Collect Alias info from subloops.
245 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
246 LoopItr != LoopItrE; ++LoopItr) {
247 Loop *InnerL = *LoopItr;
248 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
249 assert(InnerAST && "Where is my AST?");
251 // What if InnerLoop was modified by other passes ?
252 CurAST->add(*InnerAST);
254 // Once we've incorporated the inner loop's AST into ours, we don't need the
255 // subloop's anymore.
257 LoopToAliasSetMap.erase(InnerL);
262 // Get the preheader block to move instructions into...
263 Preheader = L->getLoopPreheader();
265 // Loop over the body of this loop, looking for calls, invokes, and stores.
266 // Because subloops have already been incorporated into AST, we skip blocks in
269 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
272 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
273 CurAST->add(*BB); // Incorporate the specified basic block
277 // TODO: We've already searched for instructions which may throw in subloops.
278 // We may want to reuse this information.
279 for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end();
280 (BB != BBE) && !MayThrow ; ++BB)
281 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
282 (I != E) && !MayThrow; ++I)
283 MayThrow |= I->mayThrow();
285 // We want to visit all of the instructions in this loop... that are not parts
286 // of our subloops (they have already had their invariants hoisted out of
287 // their loop, into this loop, so there is no need to process the BODIES of
290 // Traverse the body of the loop in depth first order on the dominator tree so
291 // that we are guaranteed to see definitions before we see uses. This allows
292 // us to sink instructions in one pass, without iteration. After sinking
293 // instructions, we perform another pass to hoist them out of the loop.
295 if (L->hasDedicatedExits())
296 SinkRegion(DT->getNode(L->getHeader()));
298 HoistRegion(DT->getNode(L->getHeader()));
300 // Now that all loop invariants have been removed from the loop, promote any
301 // memory references to scalars that we can.
302 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
303 SmallVector<BasicBlock *, 8> ExitBlocks;
304 SmallVector<Instruction *, 8> InsertPts;
305 PredIteratorCache PIC;
307 // Loop over all of the alias sets in the tracker object.
308 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
310 PromoteAliasSet(*I, ExitBlocks, InsertPts, PIC);
312 // Once we have promoted values across the loop body we have to recursively
313 // reform LCSSA as any nested loop may now have values defined within the
314 // loop used in the outer loop.
315 // FIXME: This is really heavy handed. It would be a bit better to use an
316 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
319 formLCSSARecursively(*L, *DT, getAnalysisIfAvailable<ScalarEvolution>());
322 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
323 // specifically moving instructions across the loop boundary and so it is
324 // especially in need of sanity checking here.
325 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
326 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
327 "Parent loop not left in LCSSA form after LICM!");
329 // Clear out loops state information for the next iteration
333 // If this loop is nested inside of another one, save the alias information
334 // for when we process the outer loop.
335 if (L->getParentLoop())
336 LoopToAliasSetMap[L] = CurAST;
342 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
343 /// dominated by the specified block, and that are in the current loop) in
344 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
345 /// uses before definitions, allowing us to sink a loop body in one pass without
348 void LICM::SinkRegion(DomTreeNode *N) {
349 assert(N != nullptr && "Null dominator tree node?");
350 BasicBlock *BB = N->getBlock();
352 // If this subregion is not in the top level loop at all, exit.
353 if (!CurLoop->contains(BB)) return;
355 // We are processing blocks in reverse dfo, so process children first.
356 const std::vector<DomTreeNode*> &Children = N->getChildren();
357 for (unsigned i = 0, e = Children.size(); i != e; ++i)
358 SinkRegion(Children[i]);
360 // Only need to process the contents of this block if it is not part of a
361 // subloop (which would already have been processed).
362 if (inSubLoop(BB)) return;
364 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
365 Instruction &I = *--II;
367 // If the instruction is dead, we would try to sink it because it isn't used
368 // in the loop, instead, just delete it.
369 if (isInstructionTriviallyDead(&I, TLI)) {
370 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
372 CurAST->deleteValue(&I);
378 // Check to see if we can sink this instruction to the exit blocks
379 // of the loop. We can do this if the all users of the instruction are
380 // outside of the loop. In this case, it doesn't even matter if the
381 // operands of the instruction are loop invariant.
383 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
390 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
391 /// dominated by the specified block, and that are in the current loop) in depth
392 /// first order w.r.t the DominatorTree. This allows us to visit definitions
393 /// before uses, allowing us to hoist a loop body in one pass without iteration.
395 void LICM::HoistRegion(DomTreeNode *N) {
396 assert(N != nullptr && "Null dominator tree node?");
397 BasicBlock *BB = N->getBlock();
399 // If this subregion is not in the top level loop at all, exit.
400 if (!CurLoop->contains(BB)) return;
402 // Only need to process the contents of this block if it is not part of a
403 // subloop (which would already have been processed).
405 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
406 Instruction &I = *II++;
408 // Try constant folding this instruction. If all the operands are
409 // constants, it is technically hoistable, but it would be better to just
411 if (Constant *C = ConstantFoldInstruction(&I, DL, TLI)) {
412 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
413 CurAST->copyValue(&I, C);
414 CurAST->deleteValue(&I);
415 I.replaceAllUsesWith(C);
420 // Try hoisting the instruction out to the preheader. We can only do this
421 // if all of the operands of the instruction are loop invariant and if it
422 // is safe to hoist the instruction.
424 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) &&
425 isSafeToExecuteUnconditionally(I))
429 const std::vector<DomTreeNode*> &Children = N->getChildren();
430 for (unsigned i = 0, e = Children.size(); i != e; ++i)
431 HoistRegion(Children[i]);
434 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
437 bool LICM::canSinkOrHoistInst(Instruction &I) {
438 // Loads have extra constraints we have to verify before we can hoist them.
439 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
440 if (!LI->isUnordered())
441 return false; // Don't hoist volatile/atomic loads!
443 // Loads from constant memory are always safe to move, even if they end up
444 // in the same alias set as something that ends up being modified.
445 if (AA->pointsToConstantMemory(LI->getOperand(0)))
447 if (LI->getMetadata("invariant.load"))
450 // Don't hoist loads which have may-aliased stores in loop.
452 if (LI->getType()->isSized())
453 Size = AA->getTypeStoreSize(LI->getType());
456 LI->getAAMetadata(AAInfo);
458 return !pointerInvalidatedByLoop(LI->getOperand(0), Size, AAInfo);
459 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
460 // Don't sink or hoist dbg info; it's legal, but not useful.
461 if (isa<DbgInfoIntrinsic>(I))
464 // Handle simple cases by querying alias analysis.
465 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
466 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
468 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
469 // If this call only reads from memory and there are no writes to memory
470 // in the loop, we can hoist or sink the call as appropriate.
471 bool FoundMod = false;
472 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
475 if (!AS.isForwardingAliasSet() && AS.isMod()) {
480 if (!FoundMod) return true;
483 // FIXME: This should use mod/ref information to see if we can hoist or
489 // Only these instructions are hoistable/sinkable.
490 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
491 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
492 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
493 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
494 !isa<InsertValueInst>(I))
497 return isSafeToExecuteUnconditionally(I);
500 /// \brief Returns true if a PHINode is a trivially replaceable with an
503 /// This is true when all incoming values are that instruction. This pattern
504 /// occurs most often with LCSSA PHI nodes.
505 static bool isTriviallyReplacablePHI(PHINode &PN, Instruction &I) {
506 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
507 if (PN.getIncomingValue(i) != &I)
513 /// isNotUsedInLoop - Return true if the only users of this instruction are
514 /// outside of the loop. If this is true, we can sink the instruction to the
515 /// exit blocks of the loop.
517 bool LICM::isNotUsedInLoop(Instruction &I) {
518 for (User *U : I.users()) {
519 Instruction *UI = cast<Instruction>(U);
520 if (PHINode *PN = dyn_cast<PHINode>(UI)) {
521 // A PHI node where all of the incoming values are this instruction are
522 // special -- they can just be RAUW'ed with the instruction and thus
523 // don't require a use in the predecessor. This is a particular important
524 // special case because it is the pattern found in LCSSA form.
525 if (isTriviallyReplacablePHI(*PN, I)) {
526 if (CurLoop->contains(PN))
532 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
533 // values. Check for such a use being inside the loop.
534 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
535 if (PN->getIncomingValue(i) == &I)
536 if (CurLoop->contains(PN->getIncomingBlock(i)))
542 if (CurLoop->contains(UI))
548 Instruction *LICM::CloneInstructionInExitBlock(Instruction &I,
549 BasicBlock &ExitBlock,
551 Instruction *New = I.clone();
552 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
553 if (!I.getName().empty()) New->setName(I.getName() + ".le");
555 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
556 // particularly cheap because we can rip off the PHI node that we're
557 // replacing for the number and blocks of the predecessors.
558 // OPT: If this shows up in a profile, we can instead finish sinking all
559 // invariant instructions, and then walk their operands to re-establish
560 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
561 // sinking bottom-up.
562 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
564 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
565 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
566 if (!OLoop->contains(&PN)) {
568 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
569 OInst->getName() + ".lcssa", ExitBlock.begin());
570 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
571 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
577 /// sink - When an instruction is found to only be used outside of the loop,
578 /// this function moves it to the exit blocks and patches up SSA form as needed.
579 /// This method is guaranteed to remove the original instruction from its
580 /// position, and may either delete it or move it to outside of the loop.
582 void LICM::sink(Instruction &I) {
583 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
585 if (isa<LoadInst>(I)) ++NumMovedLoads;
586 else if (isa<CallInst>(I)) ++NumMovedCalls;
591 SmallVector<BasicBlock *, 32> ExitBlocks;
592 CurLoop->getUniqueExitBlocks(ExitBlocks);
593 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
596 // Clones of this instruction. Don't create more than one per exit block!
597 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
599 // If this instruction is only used outside of the loop, then all users are
600 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
602 while (!I.use_empty()) {
603 Instruction *User = I.user_back();
604 if (!DT->isReachableFromEntry(User->getParent())) {
605 User->replaceUsesOfWith(&I, UndefValue::get(I.getType()));
608 // The user must be a PHI node.
609 PHINode *PN = cast<PHINode>(User);
611 BasicBlock *ExitBlock = PN->getParent();
612 assert(ExitBlockSet.count(ExitBlock) &&
613 "The LCSSA PHI is not in an exit block!");
616 auto It = SunkCopies.find(ExitBlock);
617 if (It != SunkCopies.end())
620 New = SunkCopies[ExitBlock] =
621 CloneInstructionInExitBlock(I, *ExitBlock, *PN);
623 PN->replaceAllUsesWith(New);
624 PN->eraseFromParent();
627 CurAST->deleteValue(&I);
631 /// hoist - When an instruction is found to only use loop invariant operands
632 /// that is safe to hoist, this instruction is called to do the dirty work.
634 void LICM::hoist(Instruction &I) {
635 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
638 // Move the new node to the Preheader, before its terminator.
639 I.moveBefore(Preheader->getTerminator());
641 if (isa<LoadInst>(I)) ++NumMovedLoads;
642 else if (isa<CallInst>(I)) ++NumMovedCalls;
647 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
648 /// not a trapping instruction or if it is a trapping instruction and is
649 /// guaranteed to execute.
651 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
652 // If it is not a trapping instruction, it is always safe to hoist.
653 if (isSafeToSpeculativelyExecute(&Inst, DL))
656 return isGuaranteedToExecute(Inst);
659 bool LICM::isGuaranteedToExecute(Instruction &Inst) {
661 // Somewhere in this loop there is an instruction which may throw and make us
666 // Otherwise we have to check to make sure that the instruction dominates all
667 // of the exit blocks. If it doesn't, then there is a path out of the loop
668 // which does not execute this instruction, so we can't hoist it.
670 // If the instruction is in the header block for the loop (which is very
671 // common), it is always guaranteed to dominate the exit blocks. Since this
672 // is a common case, and can save some work, check it now.
673 if (Inst.getParent() == CurLoop->getHeader())
676 // Get the exit blocks for the current loop.
677 SmallVector<BasicBlock*, 8> ExitBlocks;
678 CurLoop->getExitBlocks(ExitBlocks);
680 // Verify that the block dominates each of the exit blocks of the loop.
681 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
682 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
685 // As a degenerate case, if the loop is statically infinite then we haven't
686 // proven anything since there are no exit blocks.
687 if (ExitBlocks.empty())
694 class LoopPromoter : public LoadAndStorePromoter {
695 Value *SomePtr; // Designated pointer to store to.
696 SmallPtrSetImpl<Value*> &PointerMustAliases;
697 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
698 SmallVectorImpl<Instruction*> &LoopInsertPts;
699 PredIteratorCache &PredCache;
700 AliasSetTracker &AST;
706 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
707 if (Instruction *I = dyn_cast<Instruction>(V))
708 if (Loop *L = LI.getLoopFor(I->getParent()))
709 if (!L->contains(BB)) {
710 // We need to create an LCSSA PHI node for the incoming value and
712 PHINode *PN = PHINode::Create(
713 I->getType(), PredCache.GetNumPreds(BB),
714 I->getName() + ".lcssa", BB->begin());
715 for (BasicBlock **PI = PredCache.GetPreds(BB); *PI; ++PI)
716 PN->addIncoming(I, *PI);
723 LoopPromoter(Value *SP, const SmallVectorImpl<Instruction *> &Insts,
724 SSAUpdater &S, SmallPtrSetImpl<Value *> &PMA,
725 SmallVectorImpl<BasicBlock *> &LEB,
726 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
727 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
728 const AAMDNodes &AATags)
729 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
730 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
731 LI(li), DL(dl), Alignment(alignment), AATags(AATags) {}
733 bool isInstInList(Instruction *I,
734 const SmallVectorImpl<Instruction*> &) const override {
736 if (LoadInst *LI = dyn_cast<LoadInst>(I))
737 Ptr = LI->getOperand(0);
739 Ptr = cast<StoreInst>(I)->getPointerOperand();
740 return PointerMustAliases.count(Ptr);
743 void doExtraRewritesBeforeFinalDeletion() const override {
744 // Insert stores after in the loop exit blocks. Each exit block gets a
745 // store of the live-out values that feed them. Since we've already told
746 // the SSA updater about the defs in the loop and the preheader
747 // definition, it is all set and we can start using it.
748 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
749 BasicBlock *ExitBlock = LoopExitBlocks[i];
750 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
751 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
752 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
753 Instruction *InsertPos = LoopInsertPts[i];
754 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
755 NewSI->setAlignment(Alignment);
756 NewSI->setDebugLoc(DL);
757 if (AATags) NewSI->setAAMetadata(AATags);
761 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
762 // Update alias analysis.
763 AST.copyValue(LI, V);
765 void instructionDeleted(Instruction *I) const override {
769 } // end anon namespace
771 /// PromoteAliasSet - Try to promote memory values to scalars by sinking
772 /// stores out of the loop and moving loads to before the loop. We do this by
773 /// looping over the stores in the loop, looking for stores to Must pointers
774 /// which are loop invariant.
776 void LICM::PromoteAliasSet(AliasSet &AS,
777 SmallVectorImpl<BasicBlock*> &ExitBlocks,
778 SmallVectorImpl<Instruction*> &InsertPts,
779 PredIteratorCache &PIC) {
780 // We can promote this alias set if it has a store, if it is a "Must" alias
781 // set, if the pointer is loop invariant, and if we are not eliminating any
782 // volatile loads or stores.
783 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
784 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
787 assert(!AS.empty() &&
788 "Must alias set should have at least one pointer element in it!");
789 Value *SomePtr = AS.begin()->getValue();
791 // It isn't safe to promote a load/store from the loop if the load/store is
792 // conditional. For example, turning:
794 // for () { if (c) *P += 1; }
798 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
800 // is not safe, because *P may only be valid to access if 'c' is true.
802 // It is safe to promote P if all uses are direct load/stores and if at
803 // least one is guaranteed to be executed.
804 bool GuaranteedToExecute = false;
806 SmallVector<Instruction*, 64> LoopUses;
807 SmallPtrSet<Value*, 4> PointerMustAliases;
809 // We start with an alignment of one and try to find instructions that allow
810 // us to prove better alignment.
811 unsigned Alignment = 1;
814 // Check that all of the pointers in the alias set have the same type. We
815 // cannot (yet) promote a memory location that is loaded and stored in
816 // different sizes. While we are at it, collect alignment and AA info.
817 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
818 Value *ASIV = ASI->getValue();
819 PointerMustAliases.insert(ASIV);
821 // Check that all of the pointers in the alias set have the same type. We
822 // cannot (yet) promote a memory location that is loaded and stored in
824 if (SomePtr->getType() != ASIV->getType())
827 for (User *U : ASIV->users()) {
828 // Ignore instructions that are outside the loop.
829 Instruction *UI = dyn_cast<Instruction>(U);
830 if (!UI || !CurLoop->contains(UI))
833 // If there is an non-load/store instruction in the loop, we can't promote
835 if (LoadInst *load = dyn_cast<LoadInst>(UI)) {
836 assert(!load->isVolatile() && "AST broken");
837 if (!load->isSimple())
839 } else if (StoreInst *store = dyn_cast<StoreInst>(UI)) {
840 // Stores *of* the pointer are not interesting, only stores *to* the
842 if (UI->getOperand(1) != ASIV)
844 assert(!store->isVolatile() && "AST broken");
845 if (!store->isSimple())
848 // Note that we only check GuaranteedToExecute inside the store case
849 // so that we do not introduce stores where they did not exist before
850 // (which would break the LLVM concurrency model).
852 // If the alignment of this instruction allows us to specify a more
853 // restrictive (and performant) alignment and if we are sure this
854 // instruction will be executed, update the alignment.
855 // Larger is better, with the exception of 0 being the best alignment.
856 unsigned InstAlignment = store->getAlignment();
857 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
858 if (isGuaranteedToExecute(*UI)) {
859 GuaranteedToExecute = true;
860 Alignment = InstAlignment;
863 if (!GuaranteedToExecute)
864 GuaranteedToExecute = isGuaranteedToExecute(*UI);
867 return; // Not a load or store.
869 // Merge the AA tags.
870 if (LoopUses.empty()) {
871 // On the first load/store, just take its AA tags.
872 UI->getAAMetadata(AATags);
874 UI->getAAMetadata(AATags, /* Merge = */ true);
877 LoopUses.push_back(UI);
881 // If there isn't a guaranteed-to-execute instruction, we can't promote.
882 if (!GuaranteedToExecute)
885 // Otherwise, this is safe to promote, lets do it!
886 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
890 // Grab a debug location for the inserted loads/stores; given that the
891 // inserted loads/stores have little relation to the original loads/stores,
892 // this code just arbitrarily picks a location from one, since any debug
893 // location is better than none.
894 DebugLoc DL = LoopUses[0]->getDebugLoc();
896 // Figure out the loop exits and their insertion points, if this is the
898 if (ExitBlocks.empty()) {
899 CurLoop->getUniqueExitBlocks(ExitBlocks);
900 InsertPts.resize(ExitBlocks.size());
901 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
902 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
905 // We use the SSAUpdater interface to insert phi nodes as required.
906 SmallVector<PHINode*, 16> NewPHIs;
907 SSAUpdater SSA(&NewPHIs);
908 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
909 InsertPts, PIC, *CurAST, *LI, DL, Alignment, AATags);
911 // Set up the preheader to have a definition of the value. It is the live-out
912 // value from the preheader that uses in the loop will use.
913 LoadInst *PreheaderLoad =
914 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
915 Preheader->getTerminator());
916 PreheaderLoad->setAlignment(Alignment);
917 PreheaderLoad->setDebugLoc(DL);
918 if (AATags) PreheaderLoad->setAAMetadata(AATags);
919 SSA.AddAvailableValue(Preheader, PreheaderLoad);
921 // Rewrite all the loads in the loop and remember all the definitions from
922 // stores in the loop.
923 Promoter.run(LoopUses);
925 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
926 if (PreheaderLoad->use_empty())
927 PreheaderLoad->eraseFromParent();
931 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
932 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
933 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
937 AST->copyValue(From, To);
940 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
942 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
943 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
950 /// Simple Analysis hook. Delete value L from alias set map.
951 void LICM::deleteAnalysisLoop(Loop *L) {
952 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
957 LoopToAliasSetMap.erase(L);