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 #define DEBUG_TYPE "licm"
34 #include "llvm/Transforms/Scalar.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/AliasSetTracker.h"
38 #include "llvm/Analysis/ConstantFolding.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Analysis/LoopInfo.h"
41 #include "llvm/Analysis/LoopPass.h"
42 #include "llvm/Analysis/ValueTracking.h"
43 #include "llvm/Constants.h"
44 #include "llvm/DataLayout.h"
45 #include "llvm/DerivedTypes.h"
46 #include "llvm/Instructions.h"
47 #include "llvm/IntrinsicInst.h"
48 #include "llvm/LLVMContext.h"
49 #include "llvm/Support/CFG.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetLibraryInfo.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include "llvm/Transforms/Utils/SSAUpdater.h"
59 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
60 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
61 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
62 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
63 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
66 DisablePromotion("disable-licm-promotion", cl::Hidden,
67 cl::desc("Disable memory promotion in LICM pass"));
70 struct LICM : public LoopPass {
71 static char ID; // Pass identification, replacement for typeid
72 LICM() : LoopPass(ID) {
73 initializeLICMPass(*PassRegistry::getPassRegistry());
76 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
78 /// This transformation requires natural loop information & requires that
79 /// loop preheaders be inserted into the CFG...
81 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
83 AU.addRequired<DominatorTree>();
84 AU.addRequired<LoopInfo>();
85 AU.addRequiredID(LoopSimplifyID);
86 AU.addRequired<AliasAnalysis>();
87 AU.addPreserved<AliasAnalysis>();
88 AU.addPreserved("scalar-evolution");
89 AU.addPreservedID(LoopSimplifyID);
90 AU.addRequired<TargetLibraryInfo>();
93 using llvm::Pass::doFinalization;
95 bool doFinalization() {
96 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
101 AliasAnalysis *AA; // Current AliasAnalysis information
102 LoopInfo *LI; // Current LoopInfo
103 DominatorTree *DT; // Dominator Tree for the current Loop.
105 DataLayout *TD; // DataLayout for constant folding.
106 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
108 // State that is updated as we process loops.
109 bool Changed; // Set to true when we change anything.
110 BasicBlock *Preheader; // The preheader block of the current loop...
111 Loop *CurLoop; // The current loop we are working on...
112 AliasSetTracker *CurAST; // AliasSet information for the current loop...
113 bool MayThrow; // The current loop contains an instruction which
114 // may throw, thus preventing code motion of
115 // instructions with side effects.
116 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
118 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
119 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
121 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
123 void deleteAnalysisValue(Value *V, Loop *L);
125 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
126 /// dominated by the specified block, and that are in the current loop) in
127 /// reverse depth first order w.r.t the DominatorTree. This allows us to
128 /// visit uses before definitions, allowing us to sink a loop body in one
129 /// pass without iteration.
131 void SinkRegion(DomTreeNode *N);
133 /// HoistRegion - Walk the specified region of the CFG (defined by all
134 /// blocks dominated by the specified block, and that are in the current
135 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
136 /// visit definitions before uses, allowing us to hoist a loop body in one
137 /// pass without iteration.
139 void HoistRegion(DomTreeNode *N);
141 /// inSubLoop - Little predicate that returns true if the specified basic
142 /// block is in a subloop of the current one, not the current one itself.
144 bool inSubLoop(BasicBlock *BB) {
145 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
146 return LI->getLoopFor(BB) != CurLoop;
149 /// sink - When an instruction is found to only be used outside of the loop,
150 /// this function moves it to the exit blocks and patches up SSA form as
153 void sink(Instruction &I);
155 /// hoist - When an instruction is found to only use loop invariant operands
156 /// that is safe to hoist, this instruction is called to do the dirty work.
158 void hoist(Instruction &I);
160 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
161 /// is not a trapping instruction or if it is a trapping instruction and is
162 /// guaranteed to execute.
164 bool isSafeToExecuteUnconditionally(Instruction &I);
166 /// isGuaranteedToExecute - Check that the instruction is guaranteed to
169 bool isGuaranteedToExecute(Instruction &I);
171 /// pointerInvalidatedByLoop - Return true if the body of this loop may
172 /// store into the memory location pointed to by V.
174 bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
175 const MDNode *TBAAInfo) {
176 // Check to see if any of the basic blocks in CurLoop invalidate *V.
177 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod();
180 bool canSinkOrHoistInst(Instruction &I);
181 bool isNotUsedInLoop(Instruction &I);
183 void PromoteAliasSet(AliasSet &AS,
184 SmallVectorImpl<BasicBlock*> &ExitBlocks,
185 SmallVectorImpl<Instruction*> &InsertPts);
190 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
191 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
192 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
193 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
194 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
195 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
196 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
198 Pass *llvm::createLICMPass() { return new LICM(); }
200 /// Hoist expressions out of the specified loop. Note, alias info for inner
201 /// loop is not preserved so it is not a good idea to run LICM multiple
202 /// times on one loop.
204 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
207 // Get our Loop and Alias Analysis information...
208 LI = &getAnalysis<LoopInfo>();
209 AA = &getAnalysis<AliasAnalysis>();
210 DT = &getAnalysis<DominatorTree>();
212 TD = getAnalysisIfAvailable<DataLayout>();
213 TLI = &getAnalysis<TargetLibraryInfo>();
215 CurAST = new AliasSetTracker(*AA);
216 // Collect Alias info from subloops.
217 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
218 LoopItr != LoopItrE; ++LoopItr) {
219 Loop *InnerL = *LoopItr;
220 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
221 assert(InnerAST && "Where is my AST?");
223 // What if InnerLoop was modified by other passes ?
224 CurAST->add(*InnerAST);
226 // Once we've incorporated the inner loop's AST into ours, we don't need the
227 // subloop's anymore.
229 LoopToAliasSetMap.erase(InnerL);
234 // Get the preheader block to move instructions into...
235 Preheader = L->getLoopPreheader();
237 // Loop over the body of this loop, looking for calls, invokes, and stores.
238 // Because subloops have already been incorporated into AST, we skip blocks in
241 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
244 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
245 CurAST->add(*BB); // Incorporate the specified basic block
249 // TODO: We've already searched for instructions which may throw in subloops.
250 // We may want to reuse this information.
251 for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end();
252 (BB != BBE) && !MayThrow ; ++BB)
253 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
254 (I != E) && !MayThrow; ++I)
255 MayThrow |= I->mayThrow();
257 // We want to visit all of the instructions in this loop... that are not parts
258 // of our subloops (they have already had their invariants hoisted out of
259 // their loop, into this loop, so there is no need to process the BODIES of
262 // Traverse the body of the loop in depth first order on the dominator tree so
263 // that we are guaranteed to see definitions before we see uses. This allows
264 // us to sink instructions in one pass, without iteration. After sinking
265 // instructions, we perform another pass to hoist them out of the loop.
267 if (L->hasDedicatedExits())
268 SinkRegion(DT->getNode(L->getHeader()));
270 HoistRegion(DT->getNode(L->getHeader()));
272 // Now that all loop invariants have been removed from the loop, promote any
273 // memory references to scalars that we can.
274 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
275 SmallVector<BasicBlock *, 8> ExitBlocks;
276 SmallVector<Instruction *, 8> InsertPts;
278 // Loop over all of the alias sets in the tracker object.
279 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
281 PromoteAliasSet(*I, ExitBlocks, InsertPts);
284 // Clear out loops state information for the next iteration
288 // If this loop is nested inside of another one, save the alias information
289 // for when we process the outer loop.
290 if (L->getParentLoop())
291 LoopToAliasSetMap[L] = CurAST;
297 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
298 /// dominated by the specified block, and that are in the current loop) in
299 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
300 /// uses before definitions, allowing us to sink a loop body in one pass without
303 void LICM::SinkRegion(DomTreeNode *N) {
304 assert(N != 0 && "Null dominator tree node?");
305 BasicBlock *BB = N->getBlock();
307 // If this subregion is not in the top level loop at all, exit.
308 if (!CurLoop->contains(BB)) return;
310 // We are processing blocks in reverse dfo, so process children first.
311 const std::vector<DomTreeNode*> &Children = N->getChildren();
312 for (unsigned i = 0, e = Children.size(); i != e; ++i)
313 SinkRegion(Children[i]);
315 // Only need to process the contents of this block if it is not part of a
316 // subloop (which would already have been processed).
317 if (inSubLoop(BB)) return;
319 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
320 Instruction &I = *--II;
322 // If the instruction is dead, we would try to sink it because it isn't used
323 // in the loop, instead, just delete it.
324 if (isInstructionTriviallyDead(&I, TLI)) {
325 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
327 CurAST->deleteValue(&I);
333 // Check to see if we can sink this instruction to the exit blocks
334 // of the loop. We can do this if the all users of the instruction are
335 // outside of the loop. In this case, it doesn't even matter if the
336 // operands of the instruction are loop invariant.
338 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
345 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
346 /// dominated by the specified block, and that are in the current loop) in depth
347 /// first order w.r.t the DominatorTree. This allows us to visit definitions
348 /// before uses, allowing us to hoist a loop body in one pass without iteration.
350 void LICM::HoistRegion(DomTreeNode *N) {
351 assert(N != 0 && "Null dominator tree node?");
352 BasicBlock *BB = N->getBlock();
354 // If this subregion is not in the top level loop at all, exit.
355 if (!CurLoop->contains(BB)) return;
357 // Only need to process the contents of this block if it is not part of a
358 // subloop (which would already have been processed).
360 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
361 Instruction &I = *II++;
363 // Try constant folding this instruction. If all the operands are
364 // constants, it is technically hoistable, but it would be better to just
366 if (Constant *C = ConstantFoldInstruction(&I, TD, TLI)) {
367 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
368 CurAST->copyValue(&I, C);
369 CurAST->deleteValue(&I);
370 I.replaceAllUsesWith(C);
375 // Try hoisting the instruction out to the preheader. We can only do this
376 // if all of the operands of the instruction are loop invariant and if it
377 // is safe to hoist the instruction.
379 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) &&
380 isSafeToExecuteUnconditionally(I))
384 const std::vector<DomTreeNode*> &Children = N->getChildren();
385 for (unsigned i = 0, e = Children.size(); i != e; ++i)
386 HoistRegion(Children[i]);
389 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
392 bool LICM::canSinkOrHoistInst(Instruction &I) {
393 // Loads have extra constraints we have to verify before we can hoist them.
394 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
395 if (!LI->isUnordered())
396 return false; // Don't hoist volatile/atomic loads!
398 // Loads from constant memory are always safe to move, even if they end up
399 // in the same alias set as something that ends up being modified.
400 if (AA->pointsToConstantMemory(LI->getOperand(0)))
402 if (LI->getMetadata("invariant.load"))
405 // Don't hoist loads which have may-aliased stores in loop.
407 if (LI->getType()->isSized())
408 Size = AA->getTypeStoreSize(LI->getType());
409 return !pointerInvalidatedByLoop(LI->getOperand(0), Size,
410 LI->getMetadata(LLVMContext::MD_tbaa));
411 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
412 // Don't sink or hoist dbg info; it's legal, but not useful.
413 if (isa<DbgInfoIntrinsic>(I))
416 // Handle simple cases by querying alias analysis.
417 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
418 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
420 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
421 // If this call only reads from memory and there are no writes to memory
422 // in the loop, we can hoist or sink the call as appropriate.
423 bool FoundMod = false;
424 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
427 if (!AS.isForwardingAliasSet() && AS.isMod()) {
432 if (!FoundMod) return true;
435 // FIXME: This should use mod/ref information to see if we can hoist or
441 // Only these instructions are hoistable/sinkable.
442 bool HoistableKind = (isa<BinaryOperator>(I) || isa<CastInst>(I) ||
443 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) ||
444 isa<CmpInst>(I) || isa<InsertElementInst>(I) ||
445 isa<ExtractElementInst>(I) ||
446 isa<ShuffleVectorInst>(I));
450 return isSafeToExecuteUnconditionally(I);
453 /// isNotUsedInLoop - Return true if the only users of this instruction are
454 /// outside of the loop. If this is true, we can sink the instruction to the
455 /// exit blocks of the loop.
457 bool LICM::isNotUsedInLoop(Instruction &I) {
458 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
459 Instruction *User = cast<Instruction>(*UI);
460 if (PHINode *PN = dyn_cast<PHINode>(User)) {
461 // PHI node uses occur in predecessor blocks!
462 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
463 if (PN->getIncomingValue(i) == &I)
464 if (CurLoop->contains(PN->getIncomingBlock(i)))
466 } else if (CurLoop->contains(User)) {
474 /// sink - When an instruction is found to only be used outside of the loop,
475 /// this function moves it to the exit blocks and patches up SSA form as needed.
476 /// This method is guaranteed to remove the original instruction from its
477 /// position, and may either delete it or move it to outside of the loop.
479 void LICM::sink(Instruction &I) {
480 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
482 SmallVector<BasicBlock*, 8> ExitBlocks;
483 CurLoop->getUniqueExitBlocks(ExitBlocks);
485 if (isa<LoadInst>(I)) ++NumMovedLoads;
486 else if (isa<CallInst>(I)) ++NumMovedCalls;
490 // The case where there is only a single exit node of this loop is common
491 // enough that we handle it as a special (more efficient) case. It is more
492 // efficient to handle because there are no PHI nodes that need to be placed.
493 if (ExitBlocks.size() == 1) {
494 if (!DT->dominates(I.getParent(), ExitBlocks[0])) {
495 // Instruction is not used, just delete it.
496 CurAST->deleteValue(&I);
497 // If I has users in unreachable blocks, eliminate.
498 // If I is not void type then replaceAllUsesWith undef.
499 // This allows ValueHandlers and custom metadata to adjust itself.
501 I.replaceAllUsesWith(UndefValue::get(I.getType()));
504 // Move the instruction to the start of the exit block, after any PHI
506 I.moveBefore(ExitBlocks[0]->getFirstInsertionPt());
508 // This instruction is no longer in the AST for the current loop, because
509 // we just sunk it out of the loop. If we just sunk it into an outer
510 // loop, we will rediscover the operation when we process it.
511 CurAST->deleteValue(&I);
516 if (ExitBlocks.empty()) {
517 // The instruction is actually dead if there ARE NO exit blocks.
518 CurAST->deleteValue(&I);
519 // If I has users in unreachable blocks, eliminate.
520 // If I is not void type then replaceAllUsesWith undef.
521 // This allows ValueHandlers and custom metadata to adjust itself.
523 I.replaceAllUsesWith(UndefValue::get(I.getType()));
528 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the
529 // hard work of inserting PHI nodes as necessary.
530 SmallVector<PHINode*, 8> NewPHIs;
531 SSAUpdater SSA(&NewPHIs);
534 SSA.Initialize(I.getType(), I.getName());
536 // Insert a copy of the instruction in each exit block of the loop that is
537 // dominated by the instruction. Each exit block is known to only be in the
538 // ExitBlocks list once.
539 BasicBlock *InstOrigBB = I.getParent();
540 unsigned NumInserted = 0;
542 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
543 BasicBlock *ExitBlock = ExitBlocks[i];
545 if (!DT->dominates(InstOrigBB, ExitBlock))
548 // Insert the code after the last PHI node.
549 BasicBlock::iterator InsertPt = ExitBlock->getFirstInsertionPt();
551 // If this is the first exit block processed, just move the original
552 // instruction, otherwise clone the original instruction and insert
555 if (NumInserted++ == 0) {
556 I.moveBefore(InsertPt);
560 if (!I.getName().empty())
561 New->setName(I.getName()+".le");
562 ExitBlock->getInstList().insert(InsertPt, New);
565 // Now that we have inserted the instruction, inform SSAUpdater.
567 SSA.AddAvailableValue(ExitBlock, New);
570 // If the instruction doesn't dominate any exit blocks, it must be dead.
571 if (NumInserted == 0) {
572 CurAST->deleteValue(&I);
574 I.replaceAllUsesWith(UndefValue::get(I.getType()));
579 // Next, rewrite uses of the instruction, inserting PHI nodes as needed.
580 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) {
581 // Grab the use before incrementing the iterator.
582 Use &U = UI.getUse();
583 // Increment the iterator before removing the use from the list.
585 SSA.RewriteUseAfterInsertions(U);
588 // Update CurAST for NewPHIs if I had pointer type.
589 if (I.getType()->isPointerTy())
590 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
591 CurAST->copyValue(&I, NewPHIs[i]);
593 // Finally, remove the instruction from CurAST. It is no longer in the loop.
594 CurAST->deleteValue(&I);
597 /// hoist - When an instruction is found to only use loop invariant operands
598 /// that is safe to hoist, this instruction is called to do the dirty work.
600 void LICM::hoist(Instruction &I) {
601 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
604 // Move the new node to the Preheader, before its terminator.
605 I.moveBefore(Preheader->getTerminator());
607 if (isa<LoadInst>(I)) ++NumMovedLoads;
608 else if (isa<CallInst>(I)) ++NumMovedCalls;
613 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
614 /// not a trapping instruction or if it is a trapping instruction and is
615 /// guaranteed to execute.
617 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
618 // If it is not a trapping instruction, it is always safe to hoist.
619 if (isSafeToSpeculativelyExecute(&Inst))
622 return isGuaranteedToExecute(Inst);
625 bool LICM::isGuaranteedToExecute(Instruction &Inst) {
627 // Somewhere in this loop there is an instruction which may throw and make us
632 // Otherwise we have to check to make sure that the instruction dominates all
633 // of the exit blocks. If it doesn't, then there is a path out of the loop
634 // which does not execute this instruction, so we can't hoist it.
636 // If the instruction is in the header block for the loop (which is very
637 // common), it is always guaranteed to dominate the exit blocks. Since this
638 // is a common case, and can save some work, check it now.
639 if (Inst.getParent() == CurLoop->getHeader())
642 // Get the exit blocks for the current loop.
643 SmallVector<BasicBlock*, 8> ExitBlocks;
644 CurLoop->getExitBlocks(ExitBlocks);
646 // Verify that the block dominates each of the exit blocks of the loop.
647 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
648 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
651 // As a degenerate case, if the loop is statically infinite then we haven't
652 // proven anything since there are no exit blocks.
653 if (ExitBlocks.empty())
660 class LoopPromoter : public LoadAndStorePromoter {
661 Value *SomePtr; // Designated pointer to store to.
662 SmallPtrSet<Value*, 4> &PointerMustAliases;
663 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
664 SmallVectorImpl<Instruction*> &LoopInsertPts;
665 AliasSetTracker &AST;
669 LoopPromoter(Value *SP,
670 const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
671 SmallPtrSet<Value*, 4> &PMA,
672 SmallVectorImpl<BasicBlock*> &LEB,
673 SmallVectorImpl<Instruction*> &LIP,
674 AliasSetTracker &ast, DebugLoc dl, int alignment)
675 : LoadAndStorePromoter(Insts, S), SomePtr(SP),
676 PointerMustAliases(PMA), LoopExitBlocks(LEB), LoopInsertPts(LIP),
677 AST(ast), DL(dl), Alignment(alignment) {}
679 virtual bool isInstInList(Instruction *I,
680 const SmallVectorImpl<Instruction*> &) const {
682 if (LoadInst *LI = dyn_cast<LoadInst>(I))
683 Ptr = LI->getOperand(0);
685 Ptr = cast<StoreInst>(I)->getPointerOperand();
686 return PointerMustAliases.count(Ptr);
689 virtual void doExtraRewritesBeforeFinalDeletion() const {
690 // Insert stores after in the loop exit blocks. Each exit block gets a
691 // store of the live-out values that feed them. Since we've already told
692 // the SSA updater about the defs in the loop and the preheader
693 // definition, it is all set and we can start using it.
694 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
695 BasicBlock *ExitBlock = LoopExitBlocks[i];
696 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
697 Instruction *InsertPos = LoopInsertPts[i];
698 StoreInst *NewSI = new StoreInst(LiveInValue, SomePtr, InsertPos);
699 NewSI->setAlignment(Alignment);
700 NewSI->setDebugLoc(DL);
704 virtual void replaceLoadWithValue(LoadInst *LI, Value *V) const {
705 // Update alias analysis.
706 AST.copyValue(LI, V);
708 virtual void instructionDeleted(Instruction *I) const {
712 } // end anon namespace
714 /// PromoteAliasSet - Try to promote memory values to scalars by sinking
715 /// stores out of the loop and moving loads to before the loop. We do this by
716 /// looping over the stores in the loop, looking for stores to Must pointers
717 /// which are loop invariant.
719 void LICM::PromoteAliasSet(AliasSet &AS,
720 SmallVectorImpl<BasicBlock*> &ExitBlocks,
721 SmallVectorImpl<Instruction*> &InsertPts) {
722 // We can promote this alias set if it has a store, if it is a "Must" alias
723 // set, if the pointer is loop invariant, and if we are not eliminating any
724 // volatile loads or stores.
725 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
726 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
729 assert(!AS.empty() &&
730 "Must alias set should have at least one pointer element in it!");
731 Value *SomePtr = AS.begin()->getValue();
733 // It isn't safe to promote a load/store from the loop if the load/store is
734 // conditional. For example, turning:
736 // for () { if (c) *P += 1; }
740 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
742 // is not safe, because *P may only be valid to access if 'c' is true.
744 // It is safe to promote P if all uses are direct load/stores and if at
745 // least one is guaranteed to be executed.
746 bool GuaranteedToExecute = false;
748 SmallVector<Instruction*, 64> LoopUses;
749 SmallPtrSet<Value*, 4> PointerMustAliases;
751 // We start with an alignment of one and try to find instructions that allow
752 // us to prove better alignment.
753 unsigned Alignment = 1;
755 // Check that all of the pointers in the alias set have the same type. We
756 // cannot (yet) promote a memory location that is loaded and stored in
758 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
759 Value *ASIV = ASI->getValue();
760 PointerMustAliases.insert(ASIV);
762 // Check that all of the pointers in the alias set have the same type. We
763 // cannot (yet) promote a memory location that is loaded and stored in
765 if (SomePtr->getType() != ASIV->getType())
768 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end();
770 // Ignore instructions that are outside the loop.
771 Instruction *Use = dyn_cast<Instruction>(*UI);
772 if (!Use || !CurLoop->contains(Use))
775 // If there is an non-load/store instruction in the loop, we can't promote
777 if (LoadInst *load = dyn_cast<LoadInst>(Use)) {
778 assert(!load->isVolatile() && "AST broken");
779 if (!load->isSimple())
781 } else if (StoreInst *store = dyn_cast<StoreInst>(Use)) {
782 // Stores *of* the pointer are not interesting, only stores *to* the
784 if (Use->getOperand(1) != ASIV)
786 assert(!store->isVolatile() && "AST broken");
787 if (!store->isSimple())
790 // Note that we only check GuaranteedToExecute inside the store case
791 // so that we do not introduce stores where they did not exist before
792 // (which would break the LLVM concurrency model).
794 // If the alignment of this instruction allows us to specify a more
795 // restrictive (and performant) alignment and if we are sure this
796 // instruction will be executed, update the alignment.
797 // Larger is better, with the exception of 0 being the best alignment.
798 unsigned InstAlignment = store->getAlignment();
799 if ((InstAlignment > Alignment || InstAlignment == 0)
801 if (isGuaranteedToExecute(*Use)) {
802 GuaranteedToExecute = true;
803 Alignment = InstAlignment;
806 if (!GuaranteedToExecute)
807 GuaranteedToExecute = isGuaranteedToExecute(*Use);
810 return; // Not a load or store.
812 LoopUses.push_back(Use);
816 // If there isn't a guaranteed-to-execute instruction, we can't promote.
817 if (!GuaranteedToExecute)
820 // Otherwise, this is safe to promote, lets do it!
821 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
825 // Grab a debug location for the inserted loads/stores; given that the
826 // inserted loads/stores have little relation to the original loads/stores,
827 // this code just arbitrarily picks a location from one, since any debug
828 // location is better than none.
829 DebugLoc DL = LoopUses[0]->getDebugLoc();
831 // Figure out the loop exits and their insertion points, if this is the
833 if (ExitBlocks.empty()) {
834 CurLoop->getUniqueExitBlocks(ExitBlocks);
835 InsertPts.resize(ExitBlocks.size());
836 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
837 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
840 // We use the SSAUpdater interface to insert phi nodes as required.
841 SmallVector<PHINode*, 16> NewPHIs;
842 SSAUpdater SSA(&NewPHIs);
843 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
844 InsertPts, *CurAST, DL, Alignment);
846 // Set up the preheader to have a definition of the value. It is the live-out
847 // value from the preheader that uses in the loop will use.
848 LoadInst *PreheaderLoad =
849 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
850 Preheader->getTerminator());
851 PreheaderLoad->setAlignment(Alignment);
852 PreheaderLoad->setDebugLoc(DL);
853 SSA.AddAvailableValue(Preheader, PreheaderLoad);
855 // Rewrite all the loads in the loop and remember all the definitions from
856 // stores in the loop.
857 Promoter.run(LoopUses);
859 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
860 if (PreheaderLoad->use_empty())
861 PreheaderLoad->eraseFromParent();
865 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
866 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
867 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
871 AST->copyValue(From, To);
874 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
876 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
877 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);