1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/Instructions.h"
39 #include "llvm/Target/TargetData.h"
40 #include "llvm/Analysis/LoopInfo.h"
41 #include "llvm/Analysis/LoopPass.h"
42 #include "llvm/Analysis/AliasAnalysis.h"
43 #include "llvm/Analysis/AliasSetTracker.h"
44 #include "llvm/Analysis/Dominators.h"
45 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
46 #include "llvm/Support/CFG.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/ADT/Statistic.h"
54 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
55 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
56 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
57 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
58 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
62 DisablePromotion("disable-licm-promotion", cl::Hidden,
63 cl::desc("Disable memory promotion in LICM pass"));
65 struct VISIBILITY_HIDDEN LICM : public LoopPass {
66 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
68 /// This transformation requires natural loop information & requires that
69 /// loop preheaders be inserted into the CFG...
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
73 AU.addRequiredID(LoopSimplifyID);
74 AU.addRequired<LoopInfo>();
75 AU.addRequired<DominatorTree>();
76 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
77 AU.addRequired<AliasAnalysis>();
81 LoopToAliasMap.clear();
86 // Various analyses that we use...
87 AliasAnalysis *AA; // Current AliasAnalysis information
88 LoopInfo *LI; // Current LoopInfo
89 DominatorTree *DT; // Dominator Tree for the current Loop...
90 DominanceFrontier *DF; // Current Dominance Frontier
92 // State that is updated as we process loops
93 bool Changed; // Set to true when we change anything.
94 BasicBlock *Preheader; // The preheader block of the current loop...
95 Loop *CurLoop; // The current loop we are working on...
96 AliasSetTracker *CurAST; // AliasSet information for the current loop...
97 std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
99 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
100 /// dominated by the specified block, and that are in the current loop) in
101 /// reverse depth first order w.r.t the DominatorTree. This allows us to
102 /// visit uses before definitions, allowing us to sink a loop body in one
103 /// pass without iteration.
105 void SinkRegion(DominatorTree::Node *N);
107 /// HoistRegion - Walk the specified region of the CFG (defined by all
108 /// blocks dominated by the specified block, and that are in the current
109 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
110 /// visit definitions before uses, allowing us to hoist a loop body in one
111 /// pass without iteration.
113 void HoistRegion(DominatorTree::Node *N);
115 /// inSubLoop - Little predicate that returns true if the specified basic
116 /// block is in a subloop of the current one, not the current one itself.
118 bool inSubLoop(BasicBlock *BB) {
119 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
120 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
121 if ((*I)->contains(BB))
122 return true; // A subloop actually contains this block!
126 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
127 /// specified exit block of the loop is dominated by the specified block
128 /// that is in the body of the loop. We use these constraints to
129 /// dramatically limit the amount of the dominator tree that needs to be
131 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
132 BasicBlock *BlockInLoop) const {
133 // If the block in the loop is the loop header, it must be dominated!
134 BasicBlock *LoopHeader = CurLoop->getHeader();
135 if (BlockInLoop == LoopHeader)
138 DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop);
139 DominatorTree::Node *IDom = DT->getNode(ExitBlock);
141 // Because the exit block is not in the loop, we know we have to get _at
142 // least_ its immediate dominator.
144 // Get next Immediate Dominator.
145 IDom = IDom->getIDom();
147 // If we have got to the header of the loop, then the instructions block
148 // did not dominate the exit node, so we can't hoist it.
149 if (IDom->getBlock() == LoopHeader)
152 } while (IDom != BlockInLoopNode);
157 /// sink - When an instruction is found to only be used outside of the loop,
158 /// this function moves it to the exit blocks and patches up SSA form as
161 void sink(Instruction &I);
163 /// hoist - When an instruction is found to only use loop invariant operands
164 /// that is safe to hoist, this instruction is called to do the dirty work.
166 void hoist(Instruction &I);
168 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
169 /// is not a trapping instruction or if it is a trapping instruction and is
170 /// guaranteed to execute.
172 bool isSafeToExecuteUnconditionally(Instruction &I);
174 /// pointerInvalidatedByLoop - Return true if the body of this loop may
175 /// store into the memory location pointed to by V.
177 bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
178 // Check to see if any of the basic blocks in CurLoop invalidate *V.
179 return CurAST->getAliasSetForPointer(V, Size).isMod();
182 bool canSinkOrHoistInst(Instruction &I);
183 bool isLoopInvariantInst(Instruction &I);
184 bool isNotUsedInLoop(Instruction &I);
186 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
187 /// to scalars as we can.
189 void PromoteValuesInLoop();
191 /// FindPromotableValuesInLoop - Check the current loop for stores to
192 /// definite pointers, which are not loaded and stored through may aliases.
193 /// If these are found, create an alloca for the value, add it to the
194 /// PromotedValues list, and keep track of the mapping from value to
197 void FindPromotableValuesInLoop(
198 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
199 std::map<Value*, AllocaInst*> &Val2AlMap);
202 RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
205 LoopPass *llvm::createLICMPass() { return new LICM(); }
207 /// Hoist expressions out of the specified loop...
209 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
212 // Get our Loop and Alias Analysis information...
213 LI = &getAnalysis<LoopInfo>();
214 AA = &getAnalysis<AliasAnalysis>();
215 DF = &getAnalysis<DominanceFrontier>();
216 DT = &getAnalysis<DominatorTree>();
218 CurAST = new AliasSetTracker(*AA);
219 // Collect Alias info frmo subloops
220 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
221 LoopItr != LoopItrE; ++LoopItr) {
222 Loop *InnerL = *LoopItr;
223 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
224 assert (InnerAST && "Where is my AST?");
226 // What if InnerLoop was modified by other passes ?
227 CurAST->add(*InnerAST);
232 // Get the preheader block to move instructions into...
233 Preheader = L->getLoopPreheader();
234 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
236 // Loop over the body of this loop, looking for calls, invokes, and stores.
237 // Because subloops have already been incorporated into AST, we skip blocks in
240 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
241 E = L->getBlocks().end(); I != E; ++I)
242 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
243 CurAST->add(**I); // Incorporate the specified basic block
245 // We want to visit all of the instructions in this loop... that are not parts
246 // of our subloops (they have already had their invariants hoisted out of
247 // their loop, into this loop, so there is no need to process the BODIES of
250 // Traverse the body of the loop in depth first order on the dominator tree so
251 // that we are guaranteed to see definitions before we see uses. This allows
252 // us to sink instructions in one pass, without iteration. AFter sinking
253 // instructions, we perform another pass to hoist them out of the loop.
255 SinkRegion(DT->getNode(L->getHeader()));
256 HoistRegion(DT->getNode(L->getHeader()));
258 // Now that all loop invariants have been removed from the loop, promote any
259 // memory references to scalars that we can...
260 if (!DisablePromotion)
261 PromoteValuesInLoop();
263 // Clear out loops state information for the next iteration
267 LoopToAliasMap[L] = CurAST;
271 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
272 /// dominated by the specified block, and that are in the current loop) in
273 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
274 /// uses before definitions, allowing us to sink a loop body in one pass without
277 void LICM::SinkRegion(DominatorTree::Node *N) {
278 assert(N != 0 && "Null dominator tree node?");
279 BasicBlock *BB = N->getBlock();
281 // If this subregion is not in the top level loop at all, exit.
282 if (!CurLoop->contains(BB)) return;
284 // We are processing blocks in reverse dfo, so process children first...
285 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
286 for (unsigned i = 0, e = Children.size(); i != e; ++i)
287 SinkRegion(Children[i]);
289 // Only need to process the contents of this block if it is not part of a
290 // subloop (which would already have been processed).
291 if (inSubLoop(BB)) return;
293 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
294 Instruction &I = *--II;
296 // Check to see if we can sink this instruction to the exit blocks
297 // of the loop. We can do this if the all users of the instruction are
298 // outside of the loop. In this case, it doesn't even matter if the
299 // operands of the instruction are loop invariant.
301 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
309 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
310 /// dominated by the specified block, and that are in the current loop) in depth
311 /// first order w.r.t the DominatorTree. This allows us to visit definitions
312 /// before uses, allowing us to hoist a loop body in one pass without iteration.
314 void LICM::HoistRegion(DominatorTree::Node *N) {
315 assert(N != 0 && "Null dominator tree node?");
316 BasicBlock *BB = N->getBlock();
318 // If this subregion is not in the top level loop at all, exit.
319 if (!CurLoop->contains(BB)) return;
321 // Only need to process the contents of this block if it is not part of a
322 // subloop (which would already have been processed).
324 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
325 Instruction &I = *II++;
327 // Try hoisting the instruction out to the preheader. We can only do this
328 // if all of the operands of the instruction are loop invariant and if it
329 // is safe to hoist the instruction.
331 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
332 isSafeToExecuteUnconditionally(I))
336 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
337 for (unsigned i = 0, e = Children.size(); i != e; ++i)
338 HoistRegion(Children[i]);
341 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
344 bool LICM::canSinkOrHoistInst(Instruction &I) {
345 // Loads have extra constraints we have to verify before we can hoist them.
346 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
347 if (LI->isVolatile())
348 return false; // Don't hoist volatile loads!
350 // Don't hoist loads which have may-aliased stores in loop.
352 if (LI->getType()->isSized())
353 Size = AA->getTargetData().getTypeSize(LI->getType());
354 return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
355 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
356 // Handle obvious cases efficiently.
357 if (Function *Callee = CI->getCalledFunction()) {
358 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI);
359 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
361 else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
362 // If this call only reads from memory and there are no writes to memory
363 // in the loop, we can hoist or sink the call as appropriate.
364 bool FoundMod = false;
365 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
368 if (!AS.isForwardingAliasSet() && AS.isMod()) {
373 if (!FoundMod) return true;
377 // FIXME: This should use mod/ref information to see if we can hoist or sink
383 // Otherwise these instructions are hoistable/sinkable
384 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
385 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I);
388 /// isNotUsedInLoop - Return true if the only users of this instruction are
389 /// outside of the loop. If this is true, we can sink the instruction to the
390 /// exit blocks of the loop.
392 bool LICM::isNotUsedInLoop(Instruction &I) {
393 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
394 Instruction *User = cast<Instruction>(*UI);
395 if (PHINode *PN = dyn_cast<PHINode>(User)) {
396 // PHI node uses occur in predecessor blocks!
397 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
398 if (PN->getIncomingValue(i) == &I)
399 if (CurLoop->contains(PN->getIncomingBlock(i)))
401 } else if (CurLoop->contains(User->getParent())) {
409 /// isLoopInvariantInst - Return true if all operands of this instruction are
410 /// loop invariant. We also filter out non-hoistable instructions here just for
413 bool LICM::isLoopInvariantInst(Instruction &I) {
414 // The instruction is loop invariant if all of its operands are loop-invariant
415 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
416 if (!CurLoop->isLoopInvariant(I.getOperand(i)))
419 // If we got this far, the instruction is loop invariant!
423 /// sink - When an instruction is found to only be used outside of the loop,
424 /// this function moves it to the exit blocks and patches up SSA form as needed.
425 /// This method is guaranteed to remove the original instruction from its
426 /// position, and may either delete it or move it to outside of the loop.
428 void LICM::sink(Instruction &I) {
429 DOUT << "LICM sinking instruction: " << I;
431 std::vector<BasicBlock*> ExitBlocks;
432 CurLoop->getExitBlocks(ExitBlocks);
434 if (isa<LoadInst>(I)) ++NumMovedLoads;
435 else if (isa<CallInst>(I)) ++NumMovedCalls;
439 // The case where there is only a single exit node of this loop is common
440 // enough that we handle it as a special (more efficient) case. It is more
441 // efficient to handle because there are no PHI nodes that need to be placed.
442 if (ExitBlocks.size() == 1) {
443 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
444 // Instruction is not used, just delete it.
445 CurAST->deleteValue(&I);
446 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
447 I.replaceAllUsesWith(UndefValue::get(I.getType()));
450 // Move the instruction to the start of the exit block, after any PHI
452 I.removeFromParent();
454 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
455 while (isa<PHINode>(InsertPt)) ++InsertPt;
456 ExitBlocks[0]->getInstList().insert(InsertPt, &I);
458 } else if (ExitBlocks.size() == 0) {
459 // The instruction is actually dead if there ARE NO exit blocks.
460 CurAST->deleteValue(&I);
461 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
462 I.replaceAllUsesWith(UndefValue::get(I.getType()));
465 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
466 // do all of the hard work of inserting PHI nodes as necessary. We convert
467 // the value into a stack object to get it to do this.
469 // Firstly, we create a stack object to hold the value...
472 if (I.getType() != Type::VoidTy)
473 AI = new AllocaInst(I.getType(), 0, I.getName(),
474 I.getParent()->getParent()->front().begin());
476 // Secondly, insert load instructions for each use of the instruction
477 // outside of the loop.
478 while (!I.use_empty()) {
479 Instruction *U = cast<Instruction>(I.use_back());
481 // If the user is a PHI Node, we actually have to insert load instructions
482 // in all predecessor blocks, not in the PHI block itself!
483 if (PHINode *UPN = dyn_cast<PHINode>(U)) {
484 // Only insert into each predecessor once, so that we don't have
485 // different incoming values from the same block!
486 std::map<BasicBlock*, Value*> InsertedBlocks;
487 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i)
488 if (UPN->getIncomingValue(i) == &I) {
489 BasicBlock *Pred = UPN->getIncomingBlock(i);
490 Value *&PredVal = InsertedBlocks[Pred];
492 // Insert a new load instruction right before the terminator in
493 // the predecessor block.
494 PredVal = new LoadInst(AI, "", Pred->getTerminator());
497 UPN->setIncomingValue(i, PredVal);
501 LoadInst *L = new LoadInst(AI, "", U);
502 U->replaceUsesOfWith(&I, L);
506 // Thirdly, insert a copy of the instruction in each exit block of the loop
507 // that is dominated by the instruction, storing the result into the memory
508 // location. Be careful not to insert the instruction into any particular
509 // basic block more than once.
510 std::set<BasicBlock*> InsertedBlocks;
511 BasicBlock *InstOrigBB = I.getParent();
513 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
514 BasicBlock *ExitBlock = ExitBlocks[i];
516 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) {
517 // If we haven't already processed this exit block, do so now.
518 if (InsertedBlocks.insert(ExitBlock).second) {
519 // Insert the code after the last PHI node...
520 BasicBlock::iterator InsertPt = ExitBlock->begin();
521 while (isa<PHINode>(InsertPt)) ++InsertPt;
523 // If this is the first exit block processed, just move the original
524 // instruction, otherwise clone the original instruction and insert
527 if (InsertedBlocks.size() == 1) {
528 I.removeFromParent();
529 ExitBlock->getInstList().insert(InsertPt, &I);
533 CurAST->copyValue(&I, New);
534 if (!I.getName().empty())
535 New->setName(I.getName()+".le");
536 ExitBlock->getInstList().insert(InsertPt, New);
539 // Now that we have inserted the instruction, store it into the alloca
540 if (AI) new StoreInst(New, AI, InsertPt);
545 // If the instruction doesn't dominate any exit blocks, it must be dead.
546 if (InsertedBlocks.empty()) {
547 CurAST->deleteValue(&I);
551 // Finally, promote the fine value to SSA form.
553 std::vector<AllocaInst*> Allocas;
554 Allocas.push_back(AI);
555 PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData(), CurAST);
560 /// hoist - When an instruction is found to only use loop invariant operands
561 /// that is safe to hoist, this instruction is called to do the dirty work.
563 void LICM::hoist(Instruction &I) {
564 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I;
566 // Remove the instruction from its current basic block... but don't delete the
568 I.removeFromParent();
570 // Insert the new node in Preheader, before the terminator.
571 Preheader->getInstList().insert(Preheader->getTerminator(), &I);
573 if (isa<LoadInst>(I)) ++NumMovedLoads;
574 else if (isa<CallInst>(I)) ++NumMovedCalls;
579 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
580 /// not a trapping instruction or if it is a trapping instruction and is
581 /// guaranteed to execute.
583 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
584 // If it is not a trapping instruction, it is always safe to hoist.
585 if (!Inst.isTrapping()) return true;
587 // Otherwise we have to check to make sure that the instruction dominates all
588 // of the exit blocks. If it doesn't, then there is a path out of the loop
589 // which does not execute this instruction, so we can't hoist it.
591 // If the instruction is in the header block for the loop (which is very
592 // common), it is always guaranteed to dominate the exit blocks. Since this
593 // is a common case, and can save some work, check it now.
594 if (Inst.getParent() == CurLoop->getHeader())
597 // It's always safe to load from a global or alloca.
598 if (isa<LoadInst>(Inst))
599 if (isa<AllocationInst>(Inst.getOperand(0)) ||
600 isa<GlobalVariable>(Inst.getOperand(0)))
603 // Get the exit blocks for the current loop.
604 std::vector<BasicBlock*> ExitBlocks;
605 CurLoop->getExitBlocks(ExitBlocks);
607 // For each exit block, get the DT node and walk up the DT until the
608 // instruction's basic block is found or we exit the loop.
609 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
610 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
617 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
618 /// stores out of the loop and moving loads to before the loop. We do this by
619 /// looping over the stores in the loop, looking for stores to Must pointers
620 /// which are loop invariant. We promote these memory locations to use allocas
621 /// instead. These allocas can easily be raised to register values by the
622 /// PromoteMem2Reg functionality.
624 void LICM::PromoteValuesInLoop() {
625 // PromotedValues - List of values that are promoted out of the loop. Each
626 // value has an alloca instruction for it, and a canonical version of the
628 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
629 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
631 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
632 if (ValueToAllocaMap.empty()) return; // If there are values to promote.
635 NumPromoted += PromotedValues.size();
637 std::vector<Value*> PointerValueNumbers;
639 // Emit a copy from the value into the alloca'd value in the loop preheader
640 TerminatorInst *LoopPredInst = Preheader->getTerminator();
641 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
642 Value *Ptr = PromotedValues[i].second;
644 // If we are promoting a pointer value, update alias information for the
646 Value *LoadValue = 0;
647 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) {
648 // Locate a load or store through the pointer, and assign the same value
649 // to LI as we are loading or storing. Since we know that the value is
650 // stored in this loop, this will always succeed.
651 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end();
653 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
656 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
657 if (SI->getOperand(1) == Ptr) {
658 LoadValue = SI->getOperand(0);
662 assert(LoadValue && "No store through the pointer found!");
663 PointerValueNumbers.push_back(LoadValue); // Remember this for later.
666 // Load from the memory we are promoting.
667 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst);
669 if (LoadValue) CurAST->copyValue(LoadValue, LI);
671 // Store into the temporary alloca.
672 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
675 // Scan the basic blocks in the loop, replacing uses of our pointers with
676 // uses of the allocas in question.
678 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
679 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
680 E = LoopBBs.end(); I != E; ++I) {
681 // Rewrite all loads and stores in the block of the pointer...
682 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
684 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
685 std::map<Value*, AllocaInst*>::iterator
686 I = ValueToAllocaMap.find(L->getOperand(0));
687 if (I != ValueToAllocaMap.end())
688 L->setOperand(0, I->second); // Rewrite load instruction...
689 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
690 std::map<Value*, AllocaInst*>::iterator
691 I = ValueToAllocaMap.find(S->getOperand(1));
692 if (I != ValueToAllocaMap.end())
693 S->setOperand(1, I->second); // Rewrite store instruction...
698 // Now that the body of the loop uses the allocas instead of the original
699 // memory locations, insert code to copy the alloca value back into the
700 // original memory location on all exits from the loop. Note that we only
701 // want to insert one copy of the code in each exit block, though the loop may
702 // exit to the same block more than once.
704 std::set<BasicBlock*> ProcessedBlocks;
706 std::vector<BasicBlock*> ExitBlocks;
707 CurLoop->getExitBlocks(ExitBlocks);
708 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
709 if (ProcessedBlocks.insert(ExitBlocks[i]).second) {
710 // Copy all of the allocas into their memory locations.
711 BasicBlock::iterator BI = ExitBlocks[i]->begin();
712 while (isa<PHINode>(*BI))
713 ++BI; // Skip over all of the phi nodes in the block.
714 Instruction *InsertPos = BI;
716 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
717 // Load from the alloca.
718 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
720 // If this is a pointer type, update alias info appropriately.
721 if (isa<PointerType>(LI->getType()))
722 CurAST->copyValue(PointerValueNumbers[PVN++], LI);
724 // Store into the memory we promoted.
725 new StoreInst(LI, PromotedValues[i].second, InsertPos);
729 // Now that we have done the deed, use the mem2reg functionality to promote
730 // all of the new allocas we just created into real SSA registers.
732 std::vector<AllocaInst*> PromotedAllocas;
733 PromotedAllocas.reserve(PromotedValues.size());
734 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
735 PromotedAllocas.push_back(PromotedValues[i].first);
736 PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData(), CurAST);
739 /// FindPromotableValuesInLoop - Check the current loop for stores to definite
740 /// pointers, which are not loaded and stored through may aliases. If these are
741 /// found, create an alloca for the value, add it to the PromotedValues list,
742 /// and keep track of the mapping from value to alloca.
744 void LICM::FindPromotableValuesInLoop(
745 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
746 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
747 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
749 // Loop over all of the alias sets in the tracker object.
750 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
753 // We can promote this alias set if it has a store, if it is a "Must" alias
754 // set, if the pointer is loop invariant, and if we are not eliminating any
755 // volatile loads or stores.
756 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
757 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) {
758 assert(AS.begin() != AS.end() &&
759 "Must alias set should have at least one pointer element in it!");
760 Value *V = AS.begin()->first;
762 // Check that all of the pointers in the alias set have the same type. We
763 // cannot (yet) promote a memory location that is loaded and stored in
765 bool PointerOk = true;
766 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
767 if (V->getType() != I->first->getType()) {
773 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
774 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
775 PromotedValues.push_back(std::make_pair(AI, V));
777 // Update the AST and alias analysis.
778 CurAST->copyValue(V, AI);
780 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
781 ValueToAllocaMap.insert(std::make_pair(I->first, AI));
783 DOUT << "LICM: Promoting value: " << *V << "\n";