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/Constants.h"
36 #include "llvm/DerivedTypes.h"
37 #include "llvm/IntrinsicInst.h"
38 #include "llvm/Instructions.h"
39 #include "llvm/Analysis/AliasAnalysis.h"
40 #include "llvm/Analysis/AliasSetTracker.h"
41 #include "llvm/Analysis/ConstantFolding.h"
42 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/Analysis/LoopPass.h"
44 #include "llvm/Analysis/Dominators.h"
45 #include "llvm/Analysis/ScalarEvolution.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/SSAUpdater.h"
48 #include "llvm/Support/CFG.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/ADT/Statistic.h"
56 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
57 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
58 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
59 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
60 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
63 DisablePromotion("disable-licm-promotion", cl::Hidden,
64 cl::desc("Disable memory promotion in LICM pass"));
67 struct LICM : public LoopPass {
68 static char ID; // Pass identification, replacement for typeid
69 LICM() : LoopPass(ID) {}
71 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
73 /// This transformation requires natural loop information & requires that
74 /// loop preheaders be inserted into the CFG...
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequired<DominatorTree>();
79 AU.addRequired<LoopInfo>();
80 AU.addRequiredID(LoopSimplifyID);
81 AU.addRequired<AliasAnalysis>();
82 AU.addPreserved<AliasAnalysis>();
83 AU.addPreserved<ScalarEvolution>();
84 AU.addPreservedID(LoopSimplifyID);
87 bool doFinalization() {
88 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
93 AliasAnalysis *AA; // Current AliasAnalysis information
94 LoopInfo *LI; // Current LoopInfo
95 DominatorTree *DT; // Dominator Tree for the current Loop.
97 // State that is updated as we process loops.
98 bool Changed; // Set to true when we change anything.
99 BasicBlock *Preheader; // The preheader block of the current loop...
100 Loop *CurLoop; // The current loop we are working on...
101 AliasSetTracker *CurAST; // AliasSet information for the current loop...
102 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
104 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
105 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
107 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
109 void deleteAnalysisValue(Value *V, Loop *L);
111 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
112 /// dominated by the specified block, and that are in the current loop) in
113 /// reverse depth first order w.r.t the DominatorTree. This allows us to
114 /// visit uses before definitions, allowing us to sink a loop body in one
115 /// pass without iteration.
117 void SinkRegion(DomTreeNode *N);
119 /// HoistRegion - Walk the specified region of the CFG (defined by all
120 /// blocks dominated by the specified block, and that are in the current
121 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
122 /// visit definitions before uses, allowing us to hoist a loop body in one
123 /// pass without iteration.
125 void HoistRegion(DomTreeNode *N);
127 /// inSubLoop - Little predicate that returns true if the specified basic
128 /// block is in a subloop of the current one, not the current one itself.
130 bool inSubLoop(BasicBlock *BB) {
131 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
132 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
133 if ((*I)->contains(BB))
134 return true; // A subloop actually contains this block!
138 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
139 /// specified exit block of the loop is dominated by the specified block
140 /// that is in the body of the loop. We use these constraints to
141 /// dramatically limit the amount of the dominator tree that needs to be
143 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
144 BasicBlock *BlockInLoop) const {
145 // If the block in the loop is the loop header, it must be dominated!
146 BasicBlock *LoopHeader = CurLoop->getHeader();
147 if (BlockInLoop == LoopHeader)
150 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop);
151 DomTreeNode *IDom = DT->getNode(ExitBlock);
153 // Because the exit block is not in the loop, we know we have to get _at
154 // least_ its immediate dominator.
155 IDom = IDom->getIDom();
157 while (IDom && IDom != BlockInLoopNode) {
158 // If we have got to the header of the loop, then the instructions block
159 // did not dominate the exit node, so we can't hoist it.
160 if (IDom->getBlock() == LoopHeader)
163 // Get next Immediate Dominator.
164 IDom = IDom->getIDom();
170 /// sink - When an instruction is found to only be used outside of the loop,
171 /// this function moves it to the exit blocks and patches up SSA form as
174 void sink(Instruction &I);
176 /// hoist - When an instruction is found to only use loop invariant operands
177 /// that is safe to hoist, this instruction is called to do the dirty work.
179 void hoist(Instruction &I);
181 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
182 /// is not a trapping instruction or if it is a trapping instruction and is
183 /// guaranteed to execute.
185 bool isSafeToExecuteUnconditionally(Instruction &I);
187 /// pointerInvalidatedByLoop - Return true if the body of this loop may
188 /// store into the memory location pointed to by V.
190 bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
191 // Check to see if any of the basic blocks in CurLoop invalidate *V.
192 return CurAST->getAliasSetForPointer(V, Size).isMod();
195 bool canSinkOrHoistInst(Instruction &I);
196 bool isNotUsedInLoop(Instruction &I);
198 void PromoteAliasSet(AliasSet &AS);
203 INITIALIZE_PASS(LICM, "licm", "Loop Invariant Code Motion", false, false);
205 Pass *llvm::createLICMPass() { return new LICM(); }
207 /// Hoist expressions out of the specified loop. Note, alias info for inner
208 /// loop is not preserved so it is not a good idea to run LICM multiple
209 /// times on one loop.
211 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
214 // Get our Loop and Alias Analysis information...
215 LI = &getAnalysis<LoopInfo>();
216 AA = &getAnalysis<AliasAnalysis>();
217 DT = &getAnalysis<DominatorTree>();
219 CurAST = new AliasSetTracker(*AA);
220 // Collect Alias info from subloops.
221 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
222 LoopItr != LoopItrE; ++LoopItr) {
223 Loop *InnerL = *LoopItr;
224 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
225 assert(InnerAST && "Where is my AST?");
227 // What if InnerLoop was modified by other passes ?
228 CurAST->add(*InnerAST);
230 // Once we've incorporated the inner loop's AST into ours, we don't need the
231 // subloop's anymore.
233 LoopToAliasSetMap.erase(InnerL);
238 // Get the preheader block to move instructions into...
239 Preheader = L->getLoopPreheader();
241 // Loop over the body of this loop, looking for calls, invokes, and stores.
242 // Because subloops have already been incorporated into AST, we skip blocks in
245 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
248 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
249 CurAST->add(*BB); // Incorporate the specified basic block
252 // We want to visit all of the instructions in this loop... that are not parts
253 // of our subloops (they have already had their invariants hoisted out of
254 // their loop, into this loop, so there is no need to process the BODIES of
257 // Traverse the body of the loop in depth first order on the dominator tree so
258 // that we are guaranteed to see definitions before we see uses. This allows
259 // us to sink instructions in one pass, without iteration. After sinking
260 // instructions, we perform another pass to hoist them out of the loop.
262 if (L->hasDedicatedExits())
263 SinkRegion(DT->getNode(L->getHeader()));
265 HoistRegion(DT->getNode(L->getHeader()));
267 // Now that all loop invariants have been removed from the loop, promote any
268 // memory references to scalars that we can.
269 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
270 // Loop over all of the alias sets in the tracker object.
271 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
276 // Clear out loops state information for the next iteration
280 // If this loop is nested inside of another one, save the alias information
281 // for when we process the outer loop.
282 if (L->getParentLoop())
283 LoopToAliasSetMap[L] = CurAST;
289 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
290 /// dominated by the specified block, and that are in the current loop) in
291 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
292 /// uses before definitions, allowing us to sink a loop body in one pass without
295 void LICM::SinkRegion(DomTreeNode *N) {
296 assert(N != 0 && "Null dominator tree node?");
297 BasicBlock *BB = N->getBlock();
299 // If this subregion is not in the top level loop at all, exit.
300 if (!CurLoop->contains(BB)) return;
302 // We are processing blocks in reverse dfo, so process children first.
303 const std::vector<DomTreeNode*> &Children = N->getChildren();
304 for (unsigned i = 0, e = Children.size(); i != e; ++i)
305 SinkRegion(Children[i]);
307 // Only need to process the contents of this block if it is not part of a
308 // subloop (which would already have been processed).
309 if (inSubLoop(BB)) return;
311 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
312 Instruction &I = *--II;
314 // If the instruction is dead, we would try to sink it because it isn't used
315 // in the loop, instead, just delete it.
316 if (isInstructionTriviallyDead(&I)) {
317 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
319 CurAST->deleteValue(&I);
325 // Check to see if we can sink this instruction to the exit blocks
326 // of the loop. We can do this if the all users of the instruction are
327 // outside of the loop. In this case, it doesn't even matter if the
328 // operands of the instruction are loop invariant.
330 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
337 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
338 /// dominated by the specified block, and that are in the current loop) in depth
339 /// first order w.r.t the DominatorTree. This allows us to visit definitions
340 /// before uses, allowing us to hoist a loop body in one pass without iteration.
342 void LICM::HoistRegion(DomTreeNode *N) {
343 assert(N != 0 && "Null dominator tree node?");
344 BasicBlock *BB = N->getBlock();
346 // If this subregion is not in the top level loop at all, exit.
347 if (!CurLoop->contains(BB)) return;
349 // Only need to process the contents of this block if it is not part of a
350 // subloop (which would already have been processed).
352 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
353 Instruction &I = *II++;
355 // Try constant folding this instruction. If all the operands are
356 // constants, it is technically hoistable, but it would be better to just
358 if (Constant *C = ConstantFoldInstruction(&I)) {
359 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
360 CurAST->copyValue(&I, C);
361 CurAST->deleteValue(&I);
362 I.replaceAllUsesWith(C);
367 // Try hoisting the instruction out to the preheader. We can only do this
368 // if all of the operands of the instruction are loop invariant and if it
369 // is safe to hoist the instruction.
371 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) &&
372 isSafeToExecuteUnconditionally(I))
376 const std::vector<DomTreeNode*> &Children = N->getChildren();
377 for (unsigned i = 0, e = Children.size(); i != e; ++i)
378 HoistRegion(Children[i]);
381 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
384 bool LICM::canSinkOrHoistInst(Instruction &I) {
385 // Loads have extra constraints we have to verify before we can hoist them.
386 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
387 if (LI->isVolatile())
388 return false; // Don't hoist volatile loads!
390 // Loads from constant memory are always safe to move, even if they end up
391 // in the same alias set as something that ends up being modified.
392 if (AA->pointsToConstantMemory(LI->getOperand(0)))
395 // Don't hoist loads which have may-aliased stores in loop.
397 if (LI->getType()->isSized())
398 Size = AA->getTypeStoreSize(LI->getType());
399 return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
400 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
401 // Handle obvious cases efficiently.
402 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
403 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
405 else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
406 // If this call only reads from memory and there are no writes to memory
407 // in the loop, we can hoist or sink the call as appropriate.
408 bool FoundMod = false;
409 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
412 if (!AS.isForwardingAliasSet() && AS.isMod()) {
417 if (!FoundMod) return true;
420 // FIXME: This should use mod/ref information to see if we can hoist or sink
426 // Otherwise these instructions are hoistable/sinkable
427 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
428 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
429 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
430 isa<ShuffleVectorInst>(I);
433 /// isNotUsedInLoop - Return true if the only users of this instruction are
434 /// outside of the loop. If this is true, we can sink the instruction to the
435 /// exit blocks of the loop.
437 bool LICM::isNotUsedInLoop(Instruction &I) {
438 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
439 Instruction *User = cast<Instruction>(*UI);
440 if (PHINode *PN = dyn_cast<PHINode>(User)) {
441 // PHI node uses occur in predecessor blocks!
442 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
443 if (PN->getIncomingValue(i) == &I)
444 if (CurLoop->contains(PN->getIncomingBlock(i)))
446 } else if (CurLoop->contains(User)) {
454 /// sink - When an instruction is found to only be used outside of the loop,
455 /// this function moves it to the exit blocks and patches up SSA form as needed.
456 /// This method is guaranteed to remove the original instruction from its
457 /// position, and may either delete it or move it to outside of the loop.
459 void LICM::sink(Instruction &I) {
460 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
462 SmallVector<BasicBlock*, 8> ExitBlocks;
463 CurLoop->getUniqueExitBlocks(ExitBlocks);
465 if (isa<LoadInst>(I)) ++NumMovedLoads;
466 else if (isa<CallInst>(I)) ++NumMovedCalls;
470 // The case where there is only a single exit node of this loop is common
471 // enough that we handle it as a special (more efficient) case. It is more
472 // efficient to handle because there are no PHI nodes that need to be placed.
473 if (ExitBlocks.size() == 1) {
474 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
475 // Instruction is not used, just delete it.
476 CurAST->deleteValue(&I);
477 // If I has users in unreachable blocks, eliminate.
478 // If I is not void type then replaceAllUsesWith undef.
479 // This allows ValueHandlers and custom metadata to adjust itself.
481 I.replaceAllUsesWith(UndefValue::get(I.getType()));
484 // Move the instruction to the start of the exit block, after any PHI
486 I.moveBefore(ExitBlocks[0]->getFirstNonPHI());
488 // This instruction is no longer in the AST for the current loop, because
489 // we just sunk it out of the loop. If we just sunk it into an outer
490 // loop, we will rediscover the operation when we process it.
491 CurAST->deleteValue(&I);
496 if (ExitBlocks.empty()) {
497 // The instruction is actually dead if there ARE NO exit blocks.
498 CurAST->deleteValue(&I);
499 // If I has users in unreachable blocks, eliminate.
500 // If I is not void type then replaceAllUsesWith undef.
501 // This allows ValueHandlers and custom metadata to adjust itself.
503 I.replaceAllUsesWith(UndefValue::get(I.getType()));
508 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the
509 // hard work of inserting PHI nodes as necessary.
510 SmallVector<PHINode*, 8> NewPHIs;
511 SSAUpdater SSA(&NewPHIs);
514 SSA.Initialize(I.getType(), I.getName());
516 // Insert a copy of the instruction in each exit block of the loop that is
517 // dominated by the instruction. Each exit block is known to only be in the
518 // ExitBlocks list once.
519 BasicBlock *InstOrigBB = I.getParent();
520 unsigned NumInserted = 0;
522 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
523 BasicBlock *ExitBlock = ExitBlocks[i];
525 if (!isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB))
528 // Insert the code after the last PHI node.
529 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI();
531 // If this is the first exit block processed, just move the original
532 // instruction, otherwise clone the original instruction and insert
535 if (NumInserted++ == 0) {
536 I.moveBefore(InsertPt);
540 if (!I.getName().empty())
541 New->setName(I.getName()+".le");
542 ExitBlock->getInstList().insert(InsertPt, New);
545 // Now that we have inserted the instruction, inform SSAUpdater.
547 SSA.AddAvailableValue(ExitBlock, New);
550 // If the instruction doesn't dominate any exit blocks, it must be dead.
551 if (NumInserted == 0) {
552 CurAST->deleteValue(&I);
554 I.replaceAllUsesWith(UndefValue::get(I.getType()));
559 // Next, rewrite uses of the instruction, inserting PHI nodes as needed.
560 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) {
561 // Grab the use before incrementing the iterator.
562 Use &U = UI.getUse();
563 // Increment the iterator before removing the use from the list.
565 SSA.RewriteUseAfterInsertions(U);
568 // Update CurAST for NewPHIs if I had pointer type.
569 if (I.getType()->isPointerTy())
570 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
571 CurAST->copyValue(&I, NewPHIs[i]);
573 // Finally, remove the instruction from CurAST. It is no longer in the loop.
574 CurAST->deleteValue(&I);
577 /// hoist - When an instruction is found to only use loop invariant operands
578 /// that is safe to hoist, this instruction is called to do the dirty work.
580 void LICM::hoist(Instruction &I) {
581 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
584 // Move the new node to the Preheader, before its terminator.
585 I.moveBefore(Preheader->getTerminator());
587 if (isa<LoadInst>(I)) ++NumMovedLoads;
588 else if (isa<CallInst>(I)) ++NumMovedCalls;
593 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
594 /// not a trapping instruction or if it is a trapping instruction and is
595 /// guaranteed to execute.
597 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
598 // If it is not a trapping instruction, it is always safe to hoist.
599 if (Inst.isSafeToSpeculativelyExecute())
602 // Otherwise we have to check to make sure that the instruction dominates all
603 // of the exit blocks. If it doesn't, then there is a path out of the loop
604 // which does not execute this instruction, so we can't hoist it.
606 // If the instruction is in the header block for the loop (which is very
607 // common), it is always guaranteed to dominate the exit blocks. Since this
608 // is a common case, and can save some work, check it now.
609 if (Inst.getParent() == CurLoop->getHeader())
612 // Get the exit blocks for the current loop.
613 SmallVector<BasicBlock*, 8> ExitBlocks;
614 CurLoop->getExitBlocks(ExitBlocks);
616 // For each exit block, get the DT node and walk up the DT until the
617 // instruction's basic block is found or we exit the loop.
618 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
619 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
625 /// PromoteAliasSet - Try to promote memory values to scalars by sinking
626 /// stores out of the loop and moving loads to before the loop. We do this by
627 /// looping over the stores in the loop, looking for stores to Must pointers
628 /// which are loop invariant.
630 void LICM::PromoteAliasSet(AliasSet &AS) {
631 // We can promote this alias set if it has a store, if it is a "Must" alias
632 // set, if the pointer is loop invariant, and if we are not eliminating any
633 // volatile loads or stores.
634 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
635 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
638 assert(!AS.empty() &&
639 "Must alias set should have at least one pointer element in it!");
640 Value *SomePtr = AS.begin()->getValue();
642 // It isn't safe to promote a load/store from the loop if the load/store is
643 // conditional. For example, turning:
645 // for () { if (c) *P += 1; }
649 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
651 // is not safe, because *P may only be valid to access if 'c' is true.
653 // It is safe to promote P if all uses are direct load/stores and if at
654 // least one is guaranteed to be executed.
655 bool GuaranteedToExecute = false;
657 SmallVector<Instruction*, 64> LoopUses;
658 SmallPtrSet<Value*, 4> PointerMustAliases;
660 // Check that all of the pointers in the alias set have the same type. We
661 // cannot (yet) promote a memory location that is loaded and stored in
663 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
664 Value *ASIV = ASI->getValue();
665 PointerMustAliases.insert(ASIV);
667 // Check that all of the pointers in the alias set have the same type. We
668 // cannot (yet) promote a memory location that is loaded and stored in
670 if (SomePtr->getType() != ASIV->getType())
673 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end();
675 // Ignore instructions that are outside the loop.
676 Instruction *Use = dyn_cast<Instruction>(*UI);
677 if (!Use || !CurLoop->contains(Use))
680 // If there is an non-load/store instruction in the loop, we can't promote
682 if (isa<LoadInst>(Use))
683 assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken");
684 else if (isa<StoreInst>(Use)) {
685 assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken");
686 if (Use->getOperand(0) == ASIV) return;
688 return; // Not a load or store.
690 if (!GuaranteedToExecute)
691 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
693 LoopUses.push_back(Use);
697 // If there isn't a guaranteed-to-execute instruction, we can't promote.
698 if (!GuaranteedToExecute)
701 // Otherwise, this is safe to promote, lets do it!
702 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
706 // We use the SSAUpdater interface to insert phi nodes as required.
707 SmallVector<PHINode*, 16> NewPHIs;
708 SSAUpdater SSA(&NewPHIs);
710 // It wants to know some value of the same type as what we'll be inserting.
712 if (isa<LoadInst>(LoopUses[0]))
713 SomeValue = LoopUses[0];
715 SomeValue = cast<StoreInst>(LoopUses[0])->getOperand(0);
716 SSA.Initialize(SomeValue->getType(), SomeValue->getName());
718 // First step: bucket up uses of the pointers by the block they occur in.
719 // This is important because we have to handle multiple defs/uses in a block
720 // ourselves: SSAUpdater is purely for cross-block references.
721 // FIXME: Want a TinyVector<Instruction*> since there is usually 0/1 element.
722 DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
723 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
724 Instruction *User = LoopUses[i];
725 UsesByBlock[User->getParent()].push_back(User);
728 // Okay, now we can iterate over all the blocks in the loop with uses,
729 // processing them. Keep track of which loads are loading a live-in value.
730 SmallVector<LoadInst*, 32> LiveInLoads;
731 DenseMap<Value*, Value*> ReplacedLoads;
733 for (unsigned LoopUse = 0, e = LoopUses.size(); LoopUse != e; ++LoopUse) {
734 Instruction *User = LoopUses[LoopUse];
735 std::vector<Instruction*> &BlockUses = UsesByBlock[User->getParent()];
737 // If this block has already been processed, ignore this repeat use.
738 if (BlockUses.empty()) continue;
740 // Okay, this is the first use in the block. If this block just has a
741 // single user in it, we can rewrite it trivially.
742 if (BlockUses.size() == 1) {
743 // If it is a store, it is a trivial def of the value in the block.
744 if (isa<StoreInst>(User)) {
745 SSA.AddAvailableValue(User->getParent(),
746 cast<StoreInst>(User)->getOperand(0));
748 // Otherwise it is a load, queue it to rewrite as a live-in load.
749 LiveInLoads.push_back(cast<LoadInst>(User));
755 // Otherwise, check to see if this block is all loads. If so, we can queue
756 // them all as live in loads.
757 bool HasStore = false;
758 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
759 if (isa<StoreInst>(BlockUses[i])) {
766 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
767 LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
772 // Otherwise, we have mixed loads and stores (or just a bunch of stores).
773 // Since SSAUpdater is purely for cross-block values, we need to determine
774 // the order of these instructions in the block. If the first use in the
775 // block is a load, then it uses the live in value. The last store defines
776 // the live out value. We handle this by doing a linear scan of the block.
777 BasicBlock *BB = User->getParent();
778 Value *StoredValue = 0;
779 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
780 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
781 // If this is a load from an unrelated pointer, ignore it.
782 if (!PointerMustAliases.count(L->getOperand(0))) continue;
784 // If we haven't seen a store yet, this is a live in use, otherwise
785 // use the stored value.
787 L->replaceAllUsesWith(StoredValue);
788 ReplacedLoads[L] = StoredValue;
790 LiveInLoads.push_back(L);
795 if (StoreInst *S = dyn_cast<StoreInst>(II)) {
796 // If this is a store to an unrelated pointer, ignore it.
797 if (!PointerMustAliases.count(S->getOperand(1))) continue;
799 // Remember that this is the active value in the block.
800 StoredValue = S->getOperand(0);
804 // The last stored value that happened is the live-out for the block.
805 assert(StoredValue && "Already checked that there is a store in block");
806 SSA.AddAvailableValue(BB, StoredValue);
810 // Now that all the intra-loop values are classified, set up the preheader.
811 // It gets a load of the pointer we're promoting, and it is the live-out value
812 // from the preheader.
813 LoadInst *PreheaderLoad = new LoadInst(SomePtr,SomePtr->getName()+".promoted",
814 Preheader->getTerminator());
815 SSA.AddAvailableValue(Preheader, PreheaderLoad);
817 // Now that the preheader is good to go, set up the exit blocks. Each exit
818 // block gets a store of the live-out values that feed them. Since we've
819 // already told the SSA updater about the defs in the loop and the preheader
820 // definition, it is all set and we can start using it.
821 SmallVector<BasicBlock*, 8> ExitBlocks;
822 CurLoop->getUniqueExitBlocks(ExitBlocks);
823 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
824 BasicBlock *ExitBlock = ExitBlocks[i];
825 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
826 Instruction *InsertPos = ExitBlock->getFirstNonPHI();
827 new StoreInst(LiveInValue, SomePtr, InsertPos);
830 // Okay, now we rewrite all loads that use live-in values in the loop,
831 // inserting PHI nodes as necessary.
832 for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
833 LoadInst *ALoad = LiveInLoads[i];
834 Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent());
835 ALoad->replaceAllUsesWith(NewVal);
836 CurAST->copyValue(ALoad, NewVal);
837 ReplacedLoads[ALoad] = NewVal;
840 // If the preheader load is itself a pointer, we need to tell alias analysis
841 // about the new pointer we created in the preheader block and about any PHI
842 // nodes that just got inserted.
843 if (PreheaderLoad->getType()->isPointerTy()) {
844 // Copy any value stored to or loaded from a must-alias of the pointer.
845 CurAST->copyValue(SomeValue, PreheaderLoad);
847 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
848 CurAST->copyValue(SomeValue, NewPHIs[i]);
851 // Now that everything is rewritten, delete the old instructions from the body
852 // of the loop. They should all be dead now.
853 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
854 Instruction *User = LoopUses[i];
856 // If this is a load that still has uses, then the load must have been added
857 // as a live value in the SSAUpdate data structure for a block (e.g. because
858 // the loaded value was stored later). In this case, we need to recursively
859 // propagate the updates until we get to the real value.
860 if (!User->use_empty()) {
861 Value *NewVal = ReplacedLoads[User];
862 assert(NewVal && "not a replaced load?");
864 // Propagate down to the ultimate replacee. The intermediately loads
865 // could theoretically already have been deleted, so we don't want to
866 // dereference the Value*'s.
867 DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal);
868 while (RLI != ReplacedLoads.end()) {
869 NewVal = RLI->second;
870 RLI = ReplacedLoads.find(NewVal);
873 User->replaceAllUsesWith(NewVal);
874 CurAST->copyValue(User, NewVal);
877 CurAST->deleteValue(User);
878 User->eraseFromParent();
885 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
886 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
887 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
891 AST->copyValue(From, To);
894 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
896 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
897 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);