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 mem2reg functionality to construct the appropriate SSA form for the
32 //===----------------------------------------------------------------------===//
34 #define DEBUG_TYPE "licm"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Constants.h"
37 #include "llvm/DerivedTypes.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Instructions.h"
40 #include "llvm/Target/TargetData.h"
41 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Analysis/LoopPass.h"
43 #include "llvm/Analysis/AliasAnalysis.h"
44 #include "llvm/Analysis/AliasSetTracker.h"
45 #include "llvm/Analysis/Dominators.h"
46 #include "llvm/Analysis/ScalarEvolution.h"
47 #include "llvm/Transforms/Utils/PromoteMemToReg.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"));
66 // This feature is currently disabled by default because CodeGen is not yet
67 // capable of rematerializing these constants in PIC mode, so it can lead to
68 // degraded performance. Compile test/CodeGen/X86/remat-constant.ll with
69 // -relocation-model=pic to see an example of this.
71 EnableLICMConstantMotion("enable-licm-constant-variables", cl::Hidden,
72 cl::desc("Enable hoisting/sinking of constant "
76 struct LICM : public LoopPass {
77 static char ID; // Pass identification, replacement for typeid
78 LICM() : LoopPass(&ID) {}
80 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
82 /// This transformation requires natural loop information & requires that
83 /// loop preheaders be inserted into the CFG...
85 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
87 AU.addRequiredID(LoopSimplifyID);
88 AU.addRequired<LoopInfo>();
89 AU.addRequired<DominatorTree>();
90 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
91 AU.addRequired<AliasAnalysis>();
92 AU.addPreserved<ScalarEvolution>();
93 AU.addPreserved<DominanceFrontier>();
94 AU.addPreservedID(LoopSimplifyID);
97 bool doFinalization() {
98 // Free the values stored in the map
99 for (std::map<Loop *, AliasSetTracker *>::iterator
100 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I)
103 LoopToAliasMap.clear();
108 // Various analyses that we use...
109 AliasAnalysis *AA; // Current AliasAnalysis information
110 LoopInfo *LI; // Current LoopInfo
111 DominatorTree *DT; // Dominator Tree for the current Loop...
112 DominanceFrontier *DF; // Current Dominance Frontier
114 // State that is updated as we process loops
115 bool Changed; // Set to true when we change anything.
116 BasicBlock *Preheader; // The preheader block of the current loop...
117 Loop *CurLoop; // The current loop we are working on...
118 AliasSetTracker *CurAST; // AliasSet information for the current loop...
119 std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
121 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
122 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
124 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
126 void deleteAnalysisValue(Value *V, Loop *L);
128 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
129 /// dominated by the specified block, and that are in the current loop) in
130 /// reverse depth first order w.r.t the DominatorTree. This allows us to
131 /// visit uses before definitions, allowing us to sink a loop body in one
132 /// pass without iteration.
134 void SinkRegion(DomTreeNode *N);
136 /// HoistRegion - Walk the specified region of the CFG (defined by all
137 /// blocks dominated by the specified block, and that are in the current
138 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
139 /// visit definitions before uses, allowing us to hoist a loop body in one
140 /// pass without iteration.
142 void HoistRegion(DomTreeNode *N);
144 /// inSubLoop - Little predicate that returns true if the specified basic
145 /// block is in a subloop of the current one, not the current one itself.
147 bool inSubLoop(BasicBlock *BB) {
148 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
149 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
150 if ((*I)->contains(BB))
151 return true; // A subloop actually contains this block!
155 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
156 /// specified exit block of the loop is dominated by the specified block
157 /// that is in the body of the loop. We use these constraints to
158 /// dramatically limit the amount of the dominator tree that needs to be
160 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
161 BasicBlock *BlockInLoop) const {
162 // If the block in the loop is the loop header, it must be dominated!
163 BasicBlock *LoopHeader = CurLoop->getHeader();
164 if (BlockInLoop == LoopHeader)
167 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop);
168 DomTreeNode *IDom = DT->getNode(ExitBlock);
170 // Because the exit block is not in the loop, we know we have to get _at
171 // least_ its immediate dominator.
173 // Get next Immediate Dominator.
174 IDom = IDom->getIDom();
176 // If we have got to the header of the loop, then the instructions block
177 // did not dominate the exit node, so we can't hoist it.
178 if (IDom->getBlock() == LoopHeader)
181 } while (IDom != BlockInLoopNode);
186 /// sink - When an instruction is found to only be used outside of the loop,
187 /// this function moves it to the exit blocks and patches up SSA form as
190 void sink(Instruction &I);
192 /// hoist - When an instruction is found to only use loop invariant operands
193 /// that is safe to hoist, this instruction is called to do the dirty work.
195 void hoist(Instruction &I);
197 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
198 /// is not a trapping instruction or if it is a trapping instruction and is
199 /// guaranteed to execute.
201 bool isSafeToExecuteUnconditionally(Instruction &I);
203 /// pointerInvalidatedByLoop - Return true if the body of this loop may
204 /// store into the memory location pointed to by V.
206 bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
207 // Check to see if any of the basic blocks in CurLoop invalidate *V.
208 return CurAST->getAliasSetForPointer(V, Size).isMod();
211 bool canSinkOrHoistInst(Instruction &I);
212 bool isLoopInvariantInst(Instruction &I);
213 bool isNotUsedInLoop(Instruction &I);
215 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
216 /// to scalars as we can.
218 void PromoteValuesInLoop();
220 /// FindPromotableValuesInLoop - Check the current loop for stores to
221 /// definite pointers, which are not loaded and stored through may aliases.
222 /// If these are found, create an alloca for the value, add it to the
223 /// PromotedValues list, and keep track of the mapping from value to
226 void FindPromotableValuesInLoop(
227 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
228 std::map<Value*, AllocaInst*> &Val2AlMap);
233 static RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
235 Pass *llvm::createLICMPass() { return new LICM(); }
237 /// Hoist expressions out of the specified loop. Note, alias info for inner
238 /// loop is not preserved so it is not a good idea to run LICM multiple
239 /// times on one loop.
241 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
244 // Get our Loop and Alias Analysis information...
245 LI = &getAnalysis<LoopInfo>();
246 AA = &getAnalysis<AliasAnalysis>();
247 DF = &getAnalysis<DominanceFrontier>();
248 DT = &getAnalysis<DominatorTree>();
250 CurAST = new AliasSetTracker(*AA);
251 // Collect Alias info from subloops
252 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
253 LoopItr != LoopItrE; ++LoopItr) {
254 Loop *InnerL = *LoopItr;
255 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
256 assert (InnerAST && "Where is my AST?");
258 // What if InnerLoop was modified by other passes ?
259 CurAST->add(*InnerAST);
264 // Get the preheader block to move instructions into...
265 Preheader = L->getLoopPreheader();
267 // Loop over the body of this loop, looking for calls, invokes, and stores.
268 // Because subloops have already been incorporated into AST, we skip blocks in
271 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
274 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops...
275 CurAST->add(*BB); // Incorporate the specified basic block
278 // We want to visit all of the instructions in this loop... that are not parts
279 // of our subloops (they have already had their invariants hoisted out of
280 // their loop, into this loop, so there is no need to process the BODIES of
283 // Traverse the body of the loop in depth first order on the dominator tree so
284 // that we are guaranteed to see definitions before we see uses. This allows
285 // us to sink instructions in one pass, without iteration. After sinking
286 // instructions, we perform another pass to hoist them out of the loop.
288 if (L->hasDedicatedExits())
289 SinkRegion(DT->getNode(L->getHeader()));
291 HoistRegion(DT->getNode(L->getHeader()));
293 // Now that all loop invariants have been removed from the loop, promote any
294 // memory references to scalars that we can...
295 if (!DisablePromotion && Preheader && L->hasDedicatedExits())
296 PromoteValuesInLoop();
298 // Clear out loops state information for the next iteration
302 LoopToAliasMap[L] = CurAST;
306 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
307 /// dominated by the specified block, and that are in the current loop) in
308 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
309 /// uses before definitions, allowing us to sink a loop body in one pass without
312 void LICM::SinkRegion(DomTreeNode *N) {
313 assert(N != 0 && "Null dominator tree node?");
314 BasicBlock *BB = N->getBlock();
316 // If this subregion is not in the top level loop at all, exit.
317 if (!CurLoop->contains(BB)) return;
319 // We are processing blocks in reverse dfo, so process children first...
320 const std::vector<DomTreeNode*> &Children = N->getChildren();
321 for (unsigned i = 0, e = Children.size(); i != e; ++i)
322 SinkRegion(Children[i]);
324 // Only need to process the contents of this block if it is not part of a
325 // subloop (which would already have been processed).
326 if (inSubLoop(BB)) return;
328 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
329 Instruction &I = *--II;
331 // Check to see if we can sink this instruction to the exit blocks
332 // of the loop. We can do this if the all users of the instruction are
333 // outside of the loop. In this case, it doesn't even matter if the
334 // operands of the instruction are loop invariant.
336 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
343 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
344 /// dominated by the specified block, and that are in the current loop) in depth
345 /// first order w.r.t the DominatorTree. This allows us to visit definitions
346 /// before uses, allowing us to hoist a loop body in one pass without iteration.
348 void LICM::HoistRegion(DomTreeNode *N) {
349 assert(N != 0 && "Null dominator tree node?");
350 BasicBlock *BB = N->getBlock();
352 // If this subregion is not in the top level loop at all, exit.
353 if (!CurLoop->contains(BB)) return;
355 // Only need to process the contents of this block if it is not part of a
356 // subloop (which would already have been processed).
358 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
359 Instruction &I = *II++;
361 // Try hoisting the instruction out to the preheader. We can only do this
362 // if all of the operands of the instruction are loop invariant and if it
363 // is safe to hoist the instruction.
365 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
366 isSafeToExecuteUnconditionally(I))
370 const std::vector<DomTreeNode*> &Children = N->getChildren();
371 for (unsigned i = 0, e = Children.size(); i != e; ++i)
372 HoistRegion(Children[i]);
375 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
378 bool LICM::canSinkOrHoistInst(Instruction &I) {
379 // Loads have extra constraints we have to verify before we can hoist them.
380 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
381 if (LI->isVolatile())
382 return false; // Don't hoist volatile loads!
384 // Loads from constant memory are always safe to move, even if they end up
385 // in the same alias set as something that ends up being modified.
386 if (EnableLICMConstantMotion &&
387 AA->pointsToConstantMemory(LI->getOperand(0)))
390 // Don't hoist loads which have may-aliased stores in loop.
392 if (LI->getType()->isSized())
393 Size = AA->getTypeStoreSize(LI->getType());
394 return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
395 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
396 if (isa<DbgStopPointInst>(CI)) {
397 // Don't hoist/sink dbgstoppoints, we handle them separately
400 // Handle obvious cases efficiently.
401 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
402 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
404 else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
405 // If this call only reads from memory and there are no writes to memory
406 // in the loop, we can hoist or sink the call as appropriate.
407 bool FoundMod = false;
408 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
411 if (!AS.isForwardingAliasSet() && AS.isMod()) {
416 if (!FoundMod) return true;
419 // FIXME: This should use mod/ref information to see if we can hoist or sink
425 // Otherwise these instructions are hoistable/sinkable
426 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
427 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
428 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
429 isa<ShuffleVectorInst>(I);
432 /// isNotUsedInLoop - Return true if the only users of this instruction are
433 /// outside of the loop. If this is true, we can sink the instruction to the
434 /// exit blocks of the loop.
436 bool LICM::isNotUsedInLoop(Instruction &I) {
437 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
438 Instruction *User = cast<Instruction>(*UI);
439 if (PHINode *PN = dyn_cast<PHINode>(User)) {
440 // PHI node uses occur in predecessor blocks!
441 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
442 if (PN->getIncomingValue(i) == &I)
443 if (CurLoop->contains(PN->getIncomingBlock(i)))
445 } else if (CurLoop->contains(User->getParent())) {
453 /// isLoopInvariantInst - Return true if all operands of this instruction are
454 /// loop invariant. We also filter out non-hoistable instructions here just for
457 bool LICM::isLoopInvariantInst(Instruction &I) {
458 // The instruction is loop invariant if all of its operands are loop-invariant
459 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
460 if (!CurLoop->isLoopInvariant(I.getOperand(i)))
463 // If we got this far, the instruction is loop invariant!
467 /// sink - When an instruction is found to only be used outside of the loop,
468 /// this function moves it to the exit blocks and patches up SSA form as needed.
469 /// This method is guaranteed to remove the original instruction from its
470 /// position, and may either delete it or move it to outside of the loop.
472 void LICM::sink(Instruction &I) {
473 DEBUG(errs() << "LICM sinking instruction: " << I);
475 SmallVector<BasicBlock*, 8> ExitBlocks;
476 CurLoop->getExitBlocks(ExitBlocks);
478 if (isa<LoadInst>(I)) ++NumMovedLoads;
479 else if (isa<CallInst>(I)) ++NumMovedCalls;
483 // The case where there is only a single exit node of this loop is common
484 // enough that we handle it as a special (more efficient) case. It is more
485 // efficient to handle because there are no PHI nodes that need to be placed.
486 if (ExitBlocks.size() == 1) {
487 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
488 // Instruction is not used, just delete it.
489 CurAST->deleteValue(&I);
490 // If I has users in unreachable blocks, eliminate.
491 // If I is not void type then replaceAllUsesWith undef.
492 // This allows ValueHandlers and custom metadata to adjust itself.
493 if (!I.getType()->isVoidTy())
494 I.replaceAllUsesWith(UndefValue::get(I.getType()));
497 // Move the instruction to the start of the exit block, after any PHI
499 I.removeFromParent();
500 BasicBlock::iterator InsertPt = ExitBlocks[0]->getFirstNonPHI();
501 ExitBlocks[0]->getInstList().insert(InsertPt, &I);
503 } else if (ExitBlocks.empty()) {
504 // The instruction is actually dead if there ARE NO exit blocks.
505 CurAST->deleteValue(&I);
506 // If I has users in unreachable blocks, eliminate.
507 // If I is not void type then replaceAllUsesWith undef.
508 // This allows ValueHandlers and custom metadata to adjust itself.
509 if (!I.getType()->isVoidTy())
510 I.replaceAllUsesWith(UndefValue::get(I.getType()));
513 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
514 // do all of the hard work of inserting PHI nodes as necessary. We convert
515 // the value into a stack object to get it to do this.
517 // Firstly, we create a stack object to hold the value...
520 if (!I.getType()->isVoidTy()) {
521 AI = new AllocaInst(I.getType(), 0, I.getName(),
522 I.getParent()->getParent()->getEntryBlock().begin());
526 // Secondly, insert load instructions for each use of the instruction
527 // outside of the loop.
528 while (!I.use_empty()) {
529 Instruction *U = cast<Instruction>(I.use_back());
531 // If the user is a PHI Node, we actually have to insert load instructions
532 // in all predecessor blocks, not in the PHI block itself!
533 if (PHINode *UPN = dyn_cast<PHINode>(U)) {
534 // Only insert into each predecessor once, so that we don't have
535 // different incoming values from the same block!
536 std::map<BasicBlock*, Value*> InsertedBlocks;
537 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i)
538 if (UPN->getIncomingValue(i) == &I) {
539 BasicBlock *Pred = UPN->getIncomingBlock(i);
540 Value *&PredVal = InsertedBlocks[Pred];
542 // Insert a new load instruction right before the terminator in
543 // the predecessor block.
544 PredVal = new LoadInst(AI, "", Pred->getTerminator());
545 CurAST->add(cast<LoadInst>(PredVal));
548 UPN->setIncomingValue(i, PredVal);
552 LoadInst *L = new LoadInst(AI, "", U);
553 U->replaceUsesOfWith(&I, L);
558 // Thirdly, insert a copy of the instruction in each exit block of the loop
559 // that is dominated by the instruction, storing the result into the memory
560 // location. Be careful not to insert the instruction into any particular
561 // basic block more than once.
562 std::set<BasicBlock*> InsertedBlocks;
563 BasicBlock *InstOrigBB = I.getParent();
565 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
566 BasicBlock *ExitBlock = ExitBlocks[i];
568 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) {
569 // If we haven't already processed this exit block, do so now.
570 if (InsertedBlocks.insert(ExitBlock).second) {
571 // Insert the code after the last PHI node...
572 BasicBlock::iterator InsertPt = ExitBlock->getFirstNonPHI();
574 // If this is the first exit block processed, just move the original
575 // instruction, otherwise clone the original instruction and insert
578 if (InsertedBlocks.size() == 1) {
579 I.removeFromParent();
580 ExitBlock->getInstList().insert(InsertPt, &I);
584 CurAST->copyValue(&I, New);
585 if (!I.getName().empty())
586 New->setName(I.getName()+".le");
587 ExitBlock->getInstList().insert(InsertPt, New);
590 // Now that we have inserted the instruction, store it into the alloca
591 if (AI) new StoreInst(New, AI, InsertPt);
596 // If the instruction doesn't dominate any exit blocks, it must be dead.
597 if (InsertedBlocks.empty()) {
598 CurAST->deleteValue(&I);
602 // Finally, promote the fine value to SSA form.
604 std::vector<AllocaInst*> Allocas;
605 Allocas.push_back(AI);
606 PromoteMemToReg(Allocas, *DT, *DF, AI->getContext(), CurAST);
611 /// hoist - When an instruction is found to only use loop invariant operands
612 /// that is safe to hoist, this instruction is called to do the dirty work.
614 void LICM::hoist(Instruction &I) {
615 DEBUG(errs() << "LICM hoisting to " << Preheader->getName() << ": "
618 // Remove the instruction from its current basic block... but don't delete the
620 I.removeFromParent();
622 // Insert the new node in Preheader, before the terminator.
623 Preheader->getInstList().insert(Preheader->getTerminator(), &I);
625 if (isa<LoadInst>(I)) ++NumMovedLoads;
626 else if (isa<CallInst>(I)) ++NumMovedCalls;
631 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
632 /// not a trapping instruction or if it is a trapping instruction and is
633 /// guaranteed to execute.
635 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
636 // If it is not a trapping instruction, it is always safe to hoist.
637 if (Inst.isSafeToSpeculativelyExecute())
640 // Otherwise we have to check to make sure that the instruction dominates all
641 // of the exit blocks. If it doesn't, then there is a path out of the loop
642 // which does not execute this instruction, so we can't hoist it.
644 // If the instruction is in the header block for the loop (which is very
645 // common), it is always guaranteed to dominate the exit blocks. Since this
646 // is a common case, and can save some work, check it now.
647 if (Inst.getParent() == CurLoop->getHeader())
650 // Get the exit blocks for the current loop.
651 SmallVector<BasicBlock*, 8> ExitBlocks;
652 CurLoop->getExitBlocks(ExitBlocks);
654 // For each exit block, get the DT node and walk up the DT until the
655 // instruction's basic block is found or we exit the loop.
656 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
657 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
664 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
665 /// stores out of the loop and moving loads to before the loop. We do this by
666 /// looping over the stores in the loop, looking for stores to Must pointers
667 /// which are loop invariant. We promote these memory locations to use allocas
668 /// instead. These allocas can easily be raised to register values by the
669 /// PromoteMem2Reg functionality.
671 void LICM::PromoteValuesInLoop() {
672 // PromotedValues - List of values that are promoted out of the loop. Each
673 // value has an alloca instruction for it, and a canonical version of the
675 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
676 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
678 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
679 if (ValueToAllocaMap.empty()) return; // If there are values to promote.
682 NumPromoted += PromotedValues.size();
684 std::vector<Value*> PointerValueNumbers;
686 // Emit a copy from the value into the alloca'd value in the loop preheader
687 TerminatorInst *LoopPredInst = Preheader->getTerminator();
688 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
689 Value *Ptr = PromotedValues[i].second;
691 // If we are promoting a pointer value, update alias information for the
693 Value *LoadValue = 0;
694 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) {
695 // Locate a load or store through the pointer, and assign the same value
696 // to LI as we are loading or storing. Since we know that the value is
697 // stored in this loop, this will always succeed.
698 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end();
700 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
703 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
704 if (SI->getOperand(1) == Ptr) {
705 LoadValue = SI->getOperand(0);
709 assert(LoadValue && "No store through the pointer found!");
710 PointerValueNumbers.push_back(LoadValue); // Remember this for later.
713 // Load from the memory we are promoting.
714 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst);
716 if (LoadValue) CurAST->copyValue(LoadValue, LI);
718 // Store into the temporary alloca.
719 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
722 // Scan the basic blocks in the loop, replacing uses of our pointers with
723 // uses of the allocas in question.
725 for (Loop::block_iterator I = CurLoop->block_begin(),
726 E = CurLoop->block_end(); I != E; ++I) {
728 // Rewrite all loads and stores in the block of the pointer...
729 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
730 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
731 std::map<Value*, AllocaInst*>::iterator
732 I = ValueToAllocaMap.find(L->getOperand(0));
733 if (I != ValueToAllocaMap.end())
734 L->setOperand(0, I->second); // Rewrite load instruction...
735 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
736 std::map<Value*, AllocaInst*>::iterator
737 I = ValueToAllocaMap.find(S->getOperand(1));
738 if (I != ValueToAllocaMap.end())
739 S->setOperand(1, I->second); // Rewrite store instruction...
744 // Now that the body of the loop uses the allocas instead of the original
745 // memory locations, insert code to copy the alloca value back into the
746 // original memory location on all exits from the loop. Note that we only
747 // want to insert one copy of the code in each exit block, though the loop may
748 // exit to the same block more than once.
750 SmallPtrSet<BasicBlock*, 16> ProcessedBlocks;
752 SmallVector<BasicBlock*, 8> ExitBlocks;
753 CurLoop->getExitBlocks(ExitBlocks);
754 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
755 if (!ProcessedBlocks.insert(ExitBlocks[i]))
758 // Copy all of the allocas into their memory locations.
759 BasicBlock::iterator BI = ExitBlocks[i]->getFirstNonPHI();
760 Instruction *InsertPos = BI;
762 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
763 // Load from the alloca.
764 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
766 // If this is a pointer type, update alias info appropriately.
767 if (isa<PointerType>(LI->getType()))
768 CurAST->copyValue(PointerValueNumbers[PVN++], LI);
770 // Store into the memory we promoted.
771 new StoreInst(LI, PromotedValues[i].second, InsertPos);
775 // Now that we have done the deed, use the mem2reg functionality to promote
776 // all of the new allocas we just created into real SSA registers.
778 std::vector<AllocaInst*> PromotedAllocas;
779 PromotedAllocas.reserve(PromotedValues.size());
780 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
781 PromotedAllocas.push_back(PromotedValues[i].first);
782 PromoteMemToReg(PromotedAllocas, *DT, *DF, Preheader->getContext(), CurAST);
785 /// FindPromotableValuesInLoop - Check the current loop for stores to definite
786 /// pointers, which are not loaded and stored through may aliases and are safe
787 /// for promotion. If these are found, create an alloca for the value, add it
788 /// to the PromotedValues list, and keep track of the mapping from value to
790 void LICM::FindPromotableValuesInLoop(
791 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
792 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
793 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
795 // Loop over all of the alias sets in the tracker object.
796 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
799 // We can promote this alias set if it has a store, if it is a "Must" alias
800 // set, if the pointer is loop invariant, and if we are not eliminating any
801 // volatile loads or stores.
802 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
803 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
806 assert(!AS.empty() &&
807 "Must alias set should have at least one pointer element in it!");
808 Value *V = AS.begin()->getValue();
810 // Check that all of the pointers in the alias set have the same type. We
811 // cannot (yet) promote a memory location that is loaded and stored in
814 bool PointerOk = true;
815 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
816 if (V->getType() != I->getValue()->getType()) {
824 // It isn't safe to promote a load/store from the loop if the load/store is
825 // conditional. For example, turning:
827 // for () { if (c) *P += 1; }
831 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
833 // is not safe, because *P may only be valid to access if 'c' is true.
835 // It is safe to promote P if all uses are direct load/stores and if at
836 // least one is guaranteed to be executed.
837 bool GuaranteedToExecute = false;
838 bool InvalidInst = false;
839 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end();
841 // Ignore instructions not in this loop.
842 Instruction *Use = dyn_cast<Instruction>(*UI);
843 if (!Use || !CurLoop->contains(Use->getParent()))
846 if (!isa<LoadInst>(Use) && !isa<StoreInst>(Use)) {
851 if (!GuaranteedToExecute)
852 GuaranteedToExecute = isSafeToExecuteUnconditionally(*Use);
855 // If there is an non-load/store instruction in the loop, we can't promote
856 // it. If there isn't a guaranteed-to-execute instruction, we can't
858 if (InvalidInst || !GuaranteedToExecute)
861 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
862 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
863 PromotedValues.push_back(std::make_pair(AI, V));
865 // Update the AST and alias analysis.
866 CurAST->copyValue(V, AI);
868 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
869 ValueToAllocaMap.insert(std::make_pair(I->getValue(), AI));
871 DEBUG(errs() << "LICM: Promoting value: " << *V << "\n");
875 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
876 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
877 AliasSetTracker *AST = LoopToAliasMap[L];
881 AST->copyValue(From, To);
884 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
886 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
887 AliasSetTracker *AST = LoopToAliasMap[L];