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/LLVMContext.h"
40 #include "llvm/Analysis/AliasAnalysis.h"
41 #include "llvm/Analysis/AliasSetTracker.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/LoopInfo.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Analysis/Dominators.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) {
70 initializeLICMPass(*PassRegistry::getPassRegistry());
73 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
75 /// This transformation requires natural loop information & requires that
76 /// loop preheaders be inserted into the CFG...
78 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
80 AU.addRequired<DominatorTree>();
81 AU.addRequired<LoopInfo>();
82 AU.addRequiredID(LoopSimplifyID);
83 AU.addRequired<AliasAnalysis>();
84 AU.addPreserved<AliasAnalysis>();
85 AU.addPreserved("scalar-evolution");
86 AU.addPreservedID(LoopSimplifyID);
89 bool doFinalization() {
90 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
95 AliasAnalysis *AA; // Current AliasAnalysis information
96 LoopInfo *LI; // Current LoopInfo
97 DominatorTree *DT; // Dominator Tree for the current Loop.
99 // State that is updated as we process loops.
100 bool Changed; // Set to true when we change anything.
101 BasicBlock *Preheader; // The preheader block of the current loop...
102 Loop *CurLoop; // The current loop we are working on...
103 AliasSetTracker *CurAST; // AliasSet information for the current loop...
104 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
106 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
107 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
109 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
111 void deleteAnalysisValue(Value *V, Loop *L);
113 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
114 /// dominated by the specified block, and that are in the current loop) in
115 /// reverse depth first order w.r.t the DominatorTree. This allows us to
116 /// visit uses before definitions, allowing us to sink a loop body in one
117 /// pass without iteration.
119 void SinkRegion(DomTreeNode *N);
121 /// HoistRegion - Walk the specified region of the CFG (defined by all
122 /// blocks dominated by the specified block, and that are in the current
123 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
124 /// visit definitions before uses, allowing us to hoist a loop body in one
125 /// pass without iteration.
127 void HoistRegion(DomTreeNode *N);
129 /// inSubLoop - Little predicate that returns true if the specified basic
130 /// block is in a subloop of the current one, not the current one itself.
132 bool inSubLoop(BasicBlock *BB) {
133 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
134 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
135 if ((*I)->contains(BB))
136 return true; // A subloop actually contains this block!
140 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
141 /// specified exit block of the loop is dominated by the specified block
142 /// that is in the body of the loop. We use these constraints to
143 /// dramatically limit the amount of the dominator tree that needs to be
145 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
146 BasicBlock *BlockInLoop) const {
147 // If the block in the loop is the loop header, it must be dominated!
148 BasicBlock *LoopHeader = CurLoop->getHeader();
149 if (BlockInLoop == LoopHeader)
152 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop);
153 DomTreeNode *IDom = DT->getNode(ExitBlock);
155 // Because the exit block is not in the loop, we know we have to get _at
156 // least_ its immediate dominator.
157 IDom = IDom->getIDom();
159 while (IDom && IDom != BlockInLoopNode) {
160 // If we have got to the header of the loop, then the instructions block
161 // did not dominate the exit node, so we can't hoist it.
162 if (IDom->getBlock() == LoopHeader)
165 // Get next Immediate Dominator.
166 IDom = IDom->getIDom();
172 /// sink - When an instruction is found to only be used outside of the loop,
173 /// this function moves it to the exit blocks and patches up SSA form as
176 void sink(Instruction &I);
178 /// hoist - When an instruction is found to only use loop invariant operands
179 /// that is safe to hoist, this instruction is called to do the dirty work.
181 void hoist(Instruction &I);
183 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
184 /// is not a trapping instruction or if it is a trapping instruction and is
185 /// guaranteed to execute.
187 bool isSafeToExecuteUnconditionally(Instruction &I);
189 /// pointerInvalidatedByLoop - Return true if the body of this loop may
190 /// store into the memory location pointed to by V.
192 bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
193 const MDNode *TBAAInfo) {
194 // Check to see if any of the basic blocks in CurLoop invalidate *V.
195 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod();
198 bool canSinkOrHoistInst(Instruction &I);
199 bool isNotUsedInLoop(Instruction &I);
201 void PromoteAliasSet(AliasSet &AS);
206 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
207 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
208 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
209 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
210 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
211 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
213 Pass *llvm::createLICMPass() { return new LICM(); }
215 /// Hoist expressions out of the specified loop. Note, alias info for inner
216 /// loop is not preserved so it is not a good idea to run LICM multiple
217 /// times on one loop.
219 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
222 // Get our Loop and Alias Analysis information...
223 LI = &getAnalysis<LoopInfo>();
224 AA = &getAnalysis<AliasAnalysis>();
225 DT = &getAnalysis<DominatorTree>();
227 CurAST = new AliasSetTracker(*AA);
228 // Collect Alias info from subloops.
229 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
230 LoopItr != LoopItrE; ++LoopItr) {
231 Loop *InnerL = *LoopItr;
232 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
233 assert(InnerAST && "Where is my AST?");
235 // What if InnerLoop was modified by other passes ?
236 CurAST->add(*InnerAST);
238 // Once we've incorporated the inner loop's AST into ours, we don't need the
239 // subloop's anymore.
241 LoopToAliasSetMap.erase(InnerL);
246 // Get the preheader block to move instructions into...
247 Preheader = L->getLoopPreheader();
249 // Loop over the body of this loop, looking for calls, invokes, and stores.
250 // Because subloops have already been incorporated into AST, we skip blocks in
253 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
256 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
257 CurAST->add(*BB); // Incorporate the specified basic block
260 // We want to visit all of the instructions in this loop... that are not parts
261 // of our subloops (they have already had their invariants hoisted out of
262 // their loop, into this loop, so there is no need to process the BODIES of
265 // Traverse the body of the loop in depth first order on the dominator tree so
266 // that we are guaranteed to see definitions before we see uses. This allows
267 // us to sink instructions in one pass, without iteration. After sinking
268 // instructions, we perform another pass to hoist them out of the loop.
270 if (L->hasDedicatedExits())
271 SinkRegion(DT->getNode(L->getHeader()));
273 HoistRegion(DT->getNode(L->getHeader()));
275 // Now that all loop invariants have been removed from the loop, promote any
276 // memory references to scalars that we can.
277 if (!DisablePromotion && Preheader && L->hasDedicatedExits()) {
278 // Loop over all of the alias sets in the tracker object.
279 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
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)) {
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)) {
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->isVolatile())
396 return false; // Don't hoist volatile 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)))
403 // Don't hoist loads which have may-aliased stores in loop.
405 if (LI->getType()->isSized())
406 Size = AA->getTypeStoreSize(LI->getType());
407 return !pointerInvalidatedByLoop(LI->getOperand(0), Size,
408 LI->getMetadata(LLVMContext::MD_tbaa));
409 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
410 // Handle obvious cases efficiently.
411 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
412 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
414 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
415 // If this call only reads from memory and there are no writes to memory
416 // in the loop, we can hoist or sink the call as appropriate.
417 bool FoundMod = false;
418 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
421 if (!AS.isForwardingAliasSet() && AS.isMod()) {
426 if (!FoundMod) return true;
429 // FIXME: This should use mod/ref information to see if we can hoist or sink
435 // Otherwise these instructions are hoistable/sinkable
436 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
437 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
438 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
439 isa<ShuffleVectorInst>(I);
442 /// isNotUsedInLoop - Return true if the only users of this instruction are
443 /// outside of the loop. If this is true, we can sink the instruction to the
444 /// exit blocks of the loop.
446 bool LICM::isNotUsedInLoop(Instruction &I) {
447 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
448 Instruction *User = cast<Instruction>(*UI);
449 if (PHINode *PN = dyn_cast<PHINode>(User)) {
450 // PHI node uses occur in predecessor blocks!
451 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
452 if (PN->getIncomingValue(i) == &I)
453 if (CurLoop->contains(PN->getIncomingBlock(i)))
455 } else if (CurLoop->contains(User)) {
463 /// sink - When an instruction is found to only be used outside of the loop,
464 /// this function moves it to the exit blocks and patches up SSA form as needed.
465 /// This method is guaranteed to remove the original instruction from its
466 /// position, and may either delete it or move it to outside of the loop.
468 void LICM::sink(Instruction &I) {
469 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
471 SmallVector<BasicBlock*, 8> ExitBlocks;
472 CurLoop->getUniqueExitBlocks(ExitBlocks);
474 if (isa<LoadInst>(I)) ++NumMovedLoads;
475 else if (isa<CallInst>(I)) ++NumMovedCalls;
479 // The case where there is only a single exit node of this loop is common
480 // enough that we handle it as a special (more efficient) case. It is more
481 // efficient to handle because there are no PHI nodes that need to be placed.
482 if (ExitBlocks.size() == 1) {
483 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
484 // Instruction is not used, just delete it.
485 CurAST->deleteValue(&I);
486 // If I has users in unreachable blocks, eliminate.
487 // If I is not void type then replaceAllUsesWith undef.
488 // This allows ValueHandlers and custom metadata to adjust itself.
490 I.replaceAllUsesWith(UndefValue::get(I.getType()));
493 // Move the instruction to the start of the exit block, after any PHI
495 I.moveBefore(ExitBlocks[0]->getFirstNonPHI());
497 // This instruction is no longer in the AST for the current loop, because
498 // we just sunk it out of the loop. If we just sunk it into an outer
499 // loop, we will rediscover the operation when we process it.
500 CurAST->deleteValue(&I);
505 if (ExitBlocks.empty()) {
506 // The instruction is actually dead if there ARE NO exit blocks.
507 CurAST->deleteValue(&I);
508 // If I has users in unreachable blocks, eliminate.
509 // If I is not void type then replaceAllUsesWith undef.
510 // This allows ValueHandlers and custom metadata to adjust itself.
512 I.replaceAllUsesWith(UndefValue::get(I.getType()));
517 // Otherwise, if we have multiple exits, use the SSAUpdater to do all of the
518 // hard work of inserting PHI nodes as necessary.
519 SmallVector<PHINode*, 8> NewPHIs;
520 SSAUpdater SSA(&NewPHIs);
523 SSA.Initialize(I.getType(), I.getName());
525 // Insert a copy of the instruction in each exit block of the loop that is
526 // dominated by the instruction. Each exit block is known to only be in the
527 // ExitBlocks list once.
528 BasicBlock *InstOrigBB = I.getParent();
529 unsigned NumInserted = 0;
531 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
532 BasicBlock *ExitBlock = ExitBlocks[i];
534 if (!isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB))
537 // Insert the code after the last PHI node.
538 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI();
540 // If this is the first exit block processed, just move the original
541 // instruction, otherwise clone the original instruction and insert
544 if (NumInserted++ == 0) {
545 I.moveBefore(InsertPt);
549 if (!I.getName().empty())
550 New->setName(I.getName()+".le");
551 ExitBlock->getInstList().insert(InsertPt, New);
554 // Now that we have inserted the instruction, inform SSAUpdater.
556 SSA.AddAvailableValue(ExitBlock, New);
559 // If the instruction doesn't dominate any exit blocks, it must be dead.
560 if (NumInserted == 0) {
561 CurAST->deleteValue(&I);
563 I.replaceAllUsesWith(UndefValue::get(I.getType()));
568 // Next, rewrite uses of the instruction, inserting PHI nodes as needed.
569 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; ) {
570 // Grab the use before incrementing the iterator.
571 Use &U = UI.getUse();
572 // Increment the iterator before removing the use from the list.
574 SSA.RewriteUseAfterInsertions(U);
577 // Update CurAST for NewPHIs if I had pointer type.
578 if (I.getType()->isPointerTy())
579 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
580 CurAST->copyValue(&I, NewPHIs[i]);
582 // Finally, remove the instruction from CurAST. It is no longer in the loop.
583 CurAST->deleteValue(&I);
586 /// hoist - When an instruction is found to only use loop invariant operands
587 /// that is safe to hoist, this instruction is called to do the dirty work.
589 void LICM::hoist(Instruction &I) {
590 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
593 // Move the new node to the Preheader, before its terminator.
594 I.moveBefore(Preheader->getTerminator());
596 if (isa<LoadInst>(I)) ++NumMovedLoads;
597 else if (isa<CallInst>(I)) ++NumMovedCalls;
602 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
603 /// not a trapping instruction or if it is a trapping instruction and is
604 /// guaranteed to execute.
606 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
607 // If it is not a trapping instruction, it is always safe to hoist.
608 if (Inst.isSafeToSpeculativelyExecute())
611 // Otherwise we have to check to make sure that the instruction dominates all
612 // of the exit blocks. If it doesn't, then there is a path out of the loop
613 // which does not execute this instruction, so we can't hoist it.
615 // If the instruction is in the header block for the loop (which is very
616 // common), it is always guaranteed to dominate the exit blocks. Since this
617 // is a common case, and can save some work, check it now.
618 if (Inst.getParent() == CurLoop->getHeader())
621 // Get the exit blocks for the current loop.
622 SmallVector<BasicBlock*, 8> ExitBlocks;
623 CurLoop->getExitBlocks(ExitBlocks);
625 // For each exit block, get the DT node and walk up the DT until the
626 // instruction's basic block is found or we exit the loop.
627 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
628 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
634 /// PromoteAliasSet - Try to promote memory values to scalars by sinking
635 /// stores out of the loop and moving loads to before the loop. We do this by
636 /// looping over the stores in the loop, looking for stores to Must pointers
637 /// which are loop invariant.
639 void LICM::PromoteAliasSet(AliasSet &AS) {
640 // We can promote this alias set if it has a store, if it is a "Must" alias
641 // set, if the pointer is loop invariant, and if we are not eliminating any
642 // volatile loads or stores.
643 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
644 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
647 assert(!AS.empty() &&
648 "Must alias set should have at least one pointer element in it!");
649 Value *SomePtr = AS.begin()->getValue();
651 // It isn't safe to promote a load/store from the loop if the load/store is
652 // conditional. For example, turning:
654 // for () { if (c) *P += 1; }
658 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
660 // is not safe, because *P may only be valid to access if 'c' is true.
662 // It is safe to promote P if all uses are direct load/stores and if at
663 // least one is guaranteed to be executed.
664 bool GuaranteedToExecute = false;
666 SmallVector<Instruction*, 64> LoopUses;
667 SmallPtrSet<Value*, 4> PointerMustAliases;
669 // Check that all of the pointers in the alias set have the same type. We
670 // cannot (yet) promote a memory location that is loaded and stored in
672 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
673 Value *ASIV = ASI->getValue();
674 PointerMustAliases.insert(ASIV);
676 // Check that all of the pointers in the alias set have the same type. We
677 // cannot (yet) promote a memory location that is loaded and stored in
679 if (SomePtr->getType() != ASIV->getType())
682 for (Value::use_iterator UI = ASIV->use_begin(), UE = ASIV->use_end();
684 // Ignore instructions that are outside the loop.
685 Instruction *Use = dyn_cast<Instruction>(*UI);
686 if (!Use || !CurLoop->contains(Use))
689 // If there is an non-load/store instruction in the loop, we can't promote
691 if (isa<LoadInst>(Use))
692 assert(!cast<LoadInst>(Use)->isVolatile() && "AST broken");
693 else if (isa<StoreInst>(Use)) {
694 assert(!cast<StoreInst>(Use)->isVolatile() && "AST broken");
695 if (Use->getOperand(0) == ASIV) return;
697 return; // Not a load or store.
699 if (!GuaranteedToExecute)
700 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
702 LoopUses.push_back(Use);
706 // If there isn't a guaranteed-to-execute instruction, we can't promote.
707 if (!GuaranteedToExecute)
710 // Otherwise, this is safe to promote, lets do it!
711 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
715 // We use the SSAUpdater interface to insert phi nodes as required.
716 SmallVector<PHINode*, 16> NewPHIs;
717 SSAUpdater SSA(&NewPHIs);
719 // It wants to know some value of the same type as what we'll be inserting.
721 if (isa<LoadInst>(LoopUses[0]))
722 SomeValue = LoopUses[0];
724 SomeValue = cast<StoreInst>(LoopUses[0])->getOperand(0);
725 SSA.Initialize(SomeValue->getType(), SomeValue->getName());
727 // First step: bucket up uses of the pointers by the block they occur in.
728 // This is important because we have to handle multiple defs/uses in a block
729 // ourselves: SSAUpdater is purely for cross-block references.
730 // FIXME: Want a TinyVector<Instruction*> since there is usually 0/1 element.
731 DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
732 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
733 Instruction *User = LoopUses[i];
734 UsesByBlock[User->getParent()].push_back(User);
737 // Okay, now we can iterate over all the blocks in the loop with uses,
738 // processing them. Keep track of which loads are loading a live-in value.
739 SmallVector<LoadInst*, 32> LiveInLoads;
740 DenseMap<Value*, Value*> ReplacedLoads;
742 for (unsigned LoopUse = 0, e = LoopUses.size(); LoopUse != e; ++LoopUse) {
743 Instruction *User = LoopUses[LoopUse];
744 std::vector<Instruction*> &BlockUses = UsesByBlock[User->getParent()];
746 // If this block has already been processed, ignore this repeat use.
747 if (BlockUses.empty()) continue;
749 // Okay, this is the first use in the block. If this block just has a
750 // single user in it, we can rewrite it trivially.
751 if (BlockUses.size() == 1) {
752 // If it is a store, it is a trivial def of the value in the block.
753 if (isa<StoreInst>(User)) {
754 SSA.AddAvailableValue(User->getParent(),
755 cast<StoreInst>(User)->getOperand(0));
757 // Otherwise it is a load, queue it to rewrite as a live-in load.
758 LiveInLoads.push_back(cast<LoadInst>(User));
764 // Otherwise, check to see if this block is all loads. If so, we can queue
765 // them all as live in loads.
766 bool HasStore = false;
767 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
768 if (isa<StoreInst>(BlockUses[i])) {
775 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
776 LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
781 // Otherwise, we have mixed loads and stores (or just a bunch of stores).
782 // Since SSAUpdater is purely for cross-block values, we need to determine
783 // the order of these instructions in the block. If the first use in the
784 // block is a load, then it uses the live in value. The last store defines
785 // the live out value. We handle this by doing a linear scan of the block.
786 BasicBlock *BB = User->getParent();
787 Value *StoredValue = 0;
788 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
789 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
790 // If this is a load from an unrelated pointer, ignore it.
791 if (!PointerMustAliases.count(L->getOperand(0))) continue;
793 // If we haven't seen a store yet, this is a live in use, otherwise
794 // use the stored value.
796 L->replaceAllUsesWith(StoredValue);
797 ReplacedLoads[L] = StoredValue;
799 LiveInLoads.push_back(L);
804 if (StoreInst *S = dyn_cast<StoreInst>(II)) {
805 // If this is a store to an unrelated pointer, ignore it.
806 if (!PointerMustAliases.count(S->getOperand(1))) continue;
808 // Remember that this is the active value in the block.
809 StoredValue = S->getOperand(0);
813 // The last stored value that happened is the live-out for the block.
814 assert(StoredValue && "Already checked that there is a store in block");
815 SSA.AddAvailableValue(BB, StoredValue);
819 // Now that all the intra-loop values are classified, set up the preheader.
820 // It gets a load of the pointer we're promoting, and it is the live-out value
821 // from the preheader.
822 LoadInst *PreheaderLoad = new LoadInst(SomePtr,SomePtr->getName()+".promoted",
823 Preheader->getTerminator());
824 SSA.AddAvailableValue(Preheader, PreheaderLoad);
826 // Now that the preheader is good to go, set up the exit blocks. Each exit
827 // block gets a store of the live-out values that feed them. Since we've
828 // already told the SSA updater about the defs in the loop and the preheader
829 // definition, it is all set and we can start using it.
830 SmallVector<BasicBlock*, 8> ExitBlocks;
831 CurLoop->getUniqueExitBlocks(ExitBlocks);
832 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
833 BasicBlock *ExitBlock = ExitBlocks[i];
834 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
835 Instruction *InsertPos = ExitBlock->getFirstNonPHI();
836 new StoreInst(LiveInValue, SomePtr, InsertPos);
839 // Okay, now we rewrite all loads that use live-in values in the loop,
840 // inserting PHI nodes as necessary.
841 for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
842 LoadInst *ALoad = LiveInLoads[i];
843 Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent());
844 ALoad->replaceAllUsesWith(NewVal);
845 CurAST->copyValue(ALoad, NewVal);
846 ReplacedLoads[ALoad] = NewVal;
849 // If the preheader load is itself a pointer, we need to tell alias analysis
850 // about the new pointer we created in the preheader block and about any PHI
851 // nodes that just got inserted.
852 if (PreheaderLoad->getType()->isPointerTy()) {
853 // Copy any value stored to or loaded from a must-alias of the pointer.
854 CurAST->copyValue(SomeValue, PreheaderLoad);
856 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
857 CurAST->copyValue(SomeValue, NewPHIs[i]);
860 // Now that everything is rewritten, delete the old instructions from the body
861 // of the loop. They should all be dead now.
862 for (unsigned i = 0, e = LoopUses.size(); i != e; ++i) {
863 Instruction *User = LoopUses[i];
865 // If this is a load that still has uses, then the load must have been added
866 // as a live value in the SSAUpdate data structure for a block (e.g. because
867 // the loaded value was stored later). In this case, we need to recursively
868 // propagate the updates until we get to the real value.
869 if (!User->use_empty()) {
870 Value *NewVal = ReplacedLoads[User];
871 assert(NewVal && "not a replaced load?");
873 // Propagate down to the ultimate replacee. The intermediately loads
874 // could theoretically already have been deleted, so we don't want to
875 // dereference the Value*'s.
876 DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal);
877 while (RLI != ReplacedLoads.end()) {
878 NewVal = RLI->second;
879 RLI = ReplacedLoads.find(NewVal);
882 User->replaceAllUsesWith(NewVal);
883 CurAST->copyValue(User, NewVal);
886 CurAST->deleteValue(User);
887 User->eraseFromParent();
894 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
895 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
896 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
900 AST->copyValue(From, To);
903 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
905 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
906 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);