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 static char ID; // Pass identification, replacement for typeid
67 LICM() : LoopPass((intptr_t)&ID) {}
69 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
71 /// This transformation requires natural loop information & requires that
72 /// loop preheaders be inserted into the CFG...
74 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
76 AU.addRequiredID(LoopSimplifyID);
77 AU.addRequired<LoopInfo>();
78 AU.addRequired<DominatorTree>();
79 AU.addRequired<ETForest>();
80 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
81 AU.addRequired<AliasAnalysis>();
84 bool doFinalization() {
85 LoopToAliasMap.clear();
90 // Various analyses that we use...
91 AliasAnalysis *AA; // Current AliasAnalysis information
92 LoopInfo *LI; // Current LoopInfo
93 ETForest *ET; // ETForest for the current loop..
94 DominatorTree *DT; // Dominator Tree for the current Loop...
95 DominanceFrontier *DF; // Current Dominance Frontier
97 // State that is updated as we process loops
98 bool Changed; // Set to true when we change anything.
99 BasicBlock *Preheader; // The preheader block of the current loop...
100 Loop *CurLoop; // The current loop we are working on...
101 AliasSetTracker *CurAST; // AliasSet information for the current loop...
102 std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
104 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
105 /// dominated by the specified block, and that are in the current loop) in
106 /// reverse depth first order w.r.t the DominatorTree. This allows us to
107 /// visit uses before definitions, allowing us to sink a loop body in one
108 /// pass without iteration.
110 void SinkRegion(DominatorTree::Node *N);
112 /// HoistRegion - Walk the specified region of the CFG (defined by all
113 /// blocks dominated by the specified block, and that are in the current
114 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
115 /// visit definitions before uses, allowing us to hoist a loop body in one
116 /// pass without iteration.
118 void HoistRegion(DominatorTree::Node *N);
120 /// inSubLoop - Little predicate that returns true if the specified basic
121 /// block is in a subloop of the current one, not the current one itself.
123 bool inSubLoop(BasicBlock *BB) {
124 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
125 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
126 if ((*I)->contains(BB))
127 return true; // A subloop actually contains this block!
131 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
132 /// specified exit block of the loop is dominated by the specified block
133 /// that is in the body of the loop. We use these constraints to
134 /// dramatically limit the amount of the dominator tree that needs to be
136 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
137 BasicBlock *BlockInLoop) const {
138 // If the block in the loop is the loop header, it must be dominated!
139 BasicBlock *LoopHeader = CurLoop->getHeader();
140 if (BlockInLoop == LoopHeader)
143 DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop);
144 DominatorTree::Node *IDom = DT->getNode(ExitBlock);
146 // Because the exit block is not in the loop, we know we have to get _at
147 // least_ its immediate dominator.
149 // Get next Immediate Dominator.
150 IDom = IDom->getIDom();
152 // If we have got to the header of the loop, then the instructions block
153 // did not dominate the exit node, so we can't hoist it.
154 if (IDom->getBlock() == LoopHeader)
157 } while (IDom != BlockInLoopNode);
162 /// sink - When an instruction is found to only be used outside of the loop,
163 /// this function moves it to the exit blocks and patches up SSA form as
166 void sink(Instruction &I);
168 /// hoist - When an instruction is found to only use loop invariant operands
169 /// that is safe to hoist, this instruction is called to do the dirty work.
171 void hoist(Instruction &I);
173 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
174 /// is not a trapping instruction or if it is a trapping instruction and is
175 /// guaranteed to execute.
177 bool isSafeToExecuteUnconditionally(Instruction &I);
179 /// pointerInvalidatedByLoop - Return true if the body of this loop may
180 /// store into the memory location pointed to by V.
182 bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
183 // Check to see if any of the basic blocks in CurLoop invalidate *V.
184 return CurAST->getAliasSetForPointer(V, Size).isMod();
187 bool canSinkOrHoistInst(Instruction &I);
188 bool isLoopInvariantInst(Instruction &I);
189 bool isNotUsedInLoop(Instruction &I);
191 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
192 /// to scalars as we can.
194 void PromoteValuesInLoop();
196 /// FindPromotableValuesInLoop - Check the current loop for stores to
197 /// definite pointers, which are not loaded and stored through may aliases.
198 /// If these are found, create an alloca for the value, add it to the
199 /// PromotedValues list, and keep track of the mapping from value to
202 void FindPromotableValuesInLoop(
203 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
204 std::map<Value*, AllocaInst*> &Val2AlMap);
208 RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
211 LoopPass *llvm::createLICMPass() { return new LICM(); }
213 /// Hoist expressions out of the specified loop...
215 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
218 // Get our Loop and Alias Analysis information...
219 LI = &getAnalysis<LoopInfo>();
220 AA = &getAnalysis<AliasAnalysis>();
221 DF = &getAnalysis<DominanceFrontier>();
222 DT = &getAnalysis<DominatorTree>();
223 ET = &getAnalysis<ETForest>();
225 CurAST = new AliasSetTracker(*AA);
226 // Collect Alias info from subloops
227 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
228 LoopItr != LoopItrE; ++LoopItr) {
229 Loop *InnerL = *LoopItr;
230 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
231 assert (InnerAST && "Where is my AST?");
233 // What if InnerLoop was modified by other passes ?
234 CurAST->add(*InnerAST);
239 // Get the preheader block to move instructions into...
240 Preheader = L->getLoopPreheader();
241 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
243 // Loop over the body of this loop, looking for calls, invokes, and stores.
244 // Because subloops have already been incorporated into AST, we skip blocks in
247 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
248 E = L->getBlocks().end(); I != E; ++I)
249 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
250 CurAST->add(**I); // Incorporate the specified basic block
252 // We want to visit all of the instructions in this loop... that are not parts
253 // of our subloops (they have already had their invariants hoisted out of
254 // their loop, into this loop, so there is no need to process the BODIES of
257 // Traverse the body of the loop in depth first order on the dominator tree so
258 // that we are guaranteed to see definitions before we see uses. This allows
259 // us to sink instructions in one pass, without iteration. AFter sinking
260 // instructions, we perform another pass to hoist them out of the loop.
262 SinkRegion(DT->getNode(L->getHeader()));
263 HoistRegion(DT->getNode(L->getHeader()));
265 // Now that all loop invariants have been removed from the loop, promote any
266 // memory references to scalars that we can...
267 if (!DisablePromotion)
268 PromoteValuesInLoop();
270 // Clear out loops state information for the next iteration
274 LoopToAliasMap[L] = CurAST;
278 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
279 /// dominated by the specified block, and that are in the current loop) in
280 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
281 /// uses before definitions, allowing us to sink a loop body in one pass without
284 void LICM::SinkRegion(DominatorTree::Node *N) {
285 assert(N != 0 && "Null dominator tree node?");
286 BasicBlock *BB = N->getBlock();
288 // If this subregion is not in the top level loop at all, exit.
289 if (!CurLoop->contains(BB)) return;
291 // We are processing blocks in reverse dfo, so process children first...
292 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
293 for (unsigned i = 0, e = Children.size(); i != e; ++i)
294 SinkRegion(Children[i]);
296 // Only need to process the contents of this block if it is not part of a
297 // subloop (which would already have been processed).
298 if (inSubLoop(BB)) return;
300 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
301 Instruction &I = *--II;
303 // Check to see if we can sink this instruction to the exit blocks
304 // of the loop. We can do this if the all users of the instruction are
305 // outside of the loop. In this case, it doesn't even matter if the
306 // operands of the instruction are loop invariant.
308 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
316 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
317 /// dominated by the specified block, and that are in the current loop) in depth
318 /// first order w.r.t the DominatorTree. This allows us to visit definitions
319 /// before uses, allowing us to hoist a loop body in one pass without iteration.
321 void LICM::HoistRegion(DominatorTree::Node *N) {
322 assert(N != 0 && "Null dominator tree node?");
323 BasicBlock *BB = N->getBlock();
325 // If this subregion is not in the top level loop at all, exit.
326 if (!CurLoop->contains(BB)) return;
328 // Only need to process the contents of this block if it is not part of a
329 // subloop (which would already have been processed).
331 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
332 Instruction &I = *II++;
334 // Try hoisting the instruction out to the preheader. We can only do this
335 // if all of the operands of the instruction are loop invariant and if it
336 // is safe to hoist the instruction.
338 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
339 isSafeToExecuteUnconditionally(I))
343 const std::vector<DominatorTree::Node*> &Children = N->getChildren();
344 for (unsigned i = 0, e = Children.size(); i != e; ++i)
345 HoistRegion(Children[i]);
348 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
351 bool LICM::canSinkOrHoistInst(Instruction &I) {
352 // Loads have extra constraints we have to verify before we can hoist them.
353 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
354 if (LI->isVolatile())
355 return false; // Don't hoist volatile loads!
357 // Don't hoist loads which have may-aliased stores in loop.
359 if (LI->getType()->isSized())
360 Size = AA->getTargetData().getTypeSize(LI->getType());
361 return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
362 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
363 // Handle obvious cases efficiently.
364 if (Function *Callee = CI->getCalledFunction()) {
365 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI);
366 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
368 else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
369 // If this call only reads from memory and there are no writes to memory
370 // in the loop, we can hoist or sink the call as appropriate.
371 bool FoundMod = false;
372 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
375 if (!AS.isForwardingAliasSet() && AS.isMod()) {
380 if (!FoundMod) return true;
384 // FIXME: This should use mod/ref information to see if we can hoist or sink
390 // Otherwise these instructions are hoistable/sinkable
391 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
392 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I);
395 /// isNotUsedInLoop - Return true if the only users of this instruction are
396 /// outside of the loop. If this is true, we can sink the instruction to the
397 /// exit blocks of the loop.
399 bool LICM::isNotUsedInLoop(Instruction &I) {
400 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
401 Instruction *User = cast<Instruction>(*UI);
402 if (PHINode *PN = dyn_cast<PHINode>(User)) {
403 // PHI node uses occur in predecessor blocks!
404 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
405 if (PN->getIncomingValue(i) == &I)
406 if (CurLoop->contains(PN->getIncomingBlock(i)))
408 } else if (CurLoop->contains(User->getParent())) {
416 /// isLoopInvariantInst - Return true if all operands of this instruction are
417 /// loop invariant. We also filter out non-hoistable instructions here just for
420 bool LICM::isLoopInvariantInst(Instruction &I) {
421 // The instruction is loop invariant if all of its operands are loop-invariant
422 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
423 if (!CurLoop->isLoopInvariant(I.getOperand(i)))
426 // If we got this far, the instruction is loop invariant!
430 /// sink - When an instruction is found to only be used outside of the loop,
431 /// this function moves it to the exit blocks and patches up SSA form as needed.
432 /// This method is guaranteed to remove the original instruction from its
433 /// position, and may either delete it or move it to outside of the loop.
435 void LICM::sink(Instruction &I) {
436 DOUT << "LICM sinking instruction: " << I;
438 std::vector<BasicBlock*> ExitBlocks;
439 CurLoop->getExitBlocks(ExitBlocks);
441 if (isa<LoadInst>(I)) ++NumMovedLoads;
442 else if (isa<CallInst>(I)) ++NumMovedCalls;
446 // The case where there is only a single exit node of this loop is common
447 // enough that we handle it as a special (more efficient) case. It is more
448 // efficient to handle because there are no PHI nodes that need to be placed.
449 if (ExitBlocks.size() == 1) {
450 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
451 // Instruction is not used, just delete it.
452 CurAST->deleteValue(&I);
453 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
454 I.replaceAllUsesWith(UndefValue::get(I.getType()));
457 // Move the instruction to the start of the exit block, after any PHI
459 I.removeFromParent();
461 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
462 while (isa<PHINode>(InsertPt)) ++InsertPt;
463 ExitBlocks[0]->getInstList().insert(InsertPt, &I);
465 } else if (ExitBlocks.size() == 0) {
466 // The instruction is actually dead if there ARE NO exit blocks.
467 CurAST->deleteValue(&I);
468 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
469 I.replaceAllUsesWith(UndefValue::get(I.getType()));
472 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
473 // do all of the hard work of inserting PHI nodes as necessary. We convert
474 // the value into a stack object to get it to do this.
476 // Firstly, we create a stack object to hold the value...
479 if (I.getType() != Type::VoidTy)
480 AI = new AllocaInst(I.getType(), 0, I.getName(),
481 I.getParent()->getParent()->getEntryBlock().begin());
483 // Secondly, insert load instructions for each use of the instruction
484 // outside of the loop.
485 while (!I.use_empty()) {
486 Instruction *U = cast<Instruction>(I.use_back());
488 // If the user is a PHI Node, we actually have to insert load instructions
489 // in all predecessor blocks, not in the PHI block itself!
490 if (PHINode *UPN = dyn_cast<PHINode>(U)) {
491 // Only insert into each predecessor once, so that we don't have
492 // different incoming values from the same block!
493 std::map<BasicBlock*, Value*> InsertedBlocks;
494 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i)
495 if (UPN->getIncomingValue(i) == &I) {
496 BasicBlock *Pred = UPN->getIncomingBlock(i);
497 Value *&PredVal = InsertedBlocks[Pred];
499 // Insert a new load instruction right before the terminator in
500 // the predecessor block.
501 PredVal = new LoadInst(AI, "", Pred->getTerminator());
504 UPN->setIncomingValue(i, PredVal);
508 LoadInst *L = new LoadInst(AI, "", U);
509 U->replaceUsesOfWith(&I, L);
513 // Thirdly, insert a copy of the instruction in each exit block of the loop
514 // that is dominated by the instruction, storing the result into the memory
515 // location. Be careful not to insert the instruction into any particular
516 // basic block more than once.
517 std::set<BasicBlock*> InsertedBlocks;
518 BasicBlock *InstOrigBB = I.getParent();
520 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
521 BasicBlock *ExitBlock = ExitBlocks[i];
523 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) {
524 // If we haven't already processed this exit block, do so now.
525 if (InsertedBlocks.insert(ExitBlock).second) {
526 // Insert the code after the last PHI node...
527 BasicBlock::iterator InsertPt = ExitBlock->begin();
528 while (isa<PHINode>(InsertPt)) ++InsertPt;
530 // If this is the first exit block processed, just move the original
531 // instruction, otherwise clone the original instruction and insert
534 if (InsertedBlocks.size() == 1) {
535 I.removeFromParent();
536 ExitBlock->getInstList().insert(InsertPt, &I);
540 CurAST->copyValue(&I, New);
541 if (!I.getName().empty())
542 New->setName(I.getName()+".le");
543 ExitBlock->getInstList().insert(InsertPt, New);
546 // Now that we have inserted the instruction, store it into the alloca
547 if (AI) new StoreInst(New, AI, InsertPt);
552 // If the instruction doesn't dominate any exit blocks, it must be dead.
553 if (InsertedBlocks.empty()) {
554 CurAST->deleteValue(&I);
558 // Finally, promote the fine value to SSA form.
560 std::vector<AllocaInst*> Allocas;
561 Allocas.push_back(AI);
562 PromoteMemToReg(Allocas, *ET, *DF, CurAST);
567 /// hoist - When an instruction is found to only use loop invariant operands
568 /// that is safe to hoist, this instruction is called to do the dirty work.
570 void LICM::hoist(Instruction &I) {
571 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I;
573 // Remove the instruction from its current basic block... but don't delete the
575 I.removeFromParent();
577 // Insert the new node in Preheader, before the terminator.
578 Preheader->getInstList().insert(Preheader->getTerminator(), &I);
580 if (isa<LoadInst>(I)) ++NumMovedLoads;
581 else if (isa<CallInst>(I)) ++NumMovedCalls;
586 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
587 /// not a trapping instruction or if it is a trapping instruction and is
588 /// guaranteed to execute.
590 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
591 // If it is not a trapping instruction, it is always safe to hoist.
592 if (!Inst.isTrapping()) return true;
594 // Otherwise we have to check to make sure that the instruction dominates all
595 // of the exit blocks. If it doesn't, then there is a path out of the loop
596 // which does not execute this instruction, so we can't hoist it.
598 // If the instruction is in the header block for the loop (which is very
599 // common), it is always guaranteed to dominate the exit blocks. Since this
600 // is a common case, and can save some work, check it now.
601 if (Inst.getParent() == CurLoop->getHeader())
604 // It's always safe to load from a global or alloca.
605 if (isa<LoadInst>(Inst))
606 if (isa<AllocationInst>(Inst.getOperand(0)) ||
607 isa<GlobalVariable>(Inst.getOperand(0)))
610 // Get the exit blocks for the current loop.
611 std::vector<BasicBlock*> ExitBlocks;
612 CurLoop->getExitBlocks(ExitBlocks);
614 // For each exit block, get the DT node and walk up the DT until the
615 // instruction's basic block is found or we exit the loop.
616 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
617 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
624 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
625 /// stores out of the loop and moving loads to before the loop. We do this by
626 /// looping over the stores in the loop, looking for stores to Must pointers
627 /// which are loop invariant. We promote these memory locations to use allocas
628 /// instead. These allocas can easily be raised to register values by the
629 /// PromoteMem2Reg functionality.
631 void LICM::PromoteValuesInLoop() {
632 // PromotedValues - List of values that are promoted out of the loop. Each
633 // value has an alloca instruction for it, and a canonical version of the
635 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
636 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
638 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
639 if (ValueToAllocaMap.empty()) return; // If there are values to promote.
642 NumPromoted += PromotedValues.size();
644 std::vector<Value*> PointerValueNumbers;
646 // Emit a copy from the value into the alloca'd value in the loop preheader
647 TerminatorInst *LoopPredInst = Preheader->getTerminator();
648 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
649 Value *Ptr = PromotedValues[i].second;
651 // If we are promoting a pointer value, update alias information for the
653 Value *LoadValue = 0;
654 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) {
655 // Locate a load or store through the pointer, and assign the same value
656 // to LI as we are loading or storing. Since we know that the value is
657 // stored in this loop, this will always succeed.
658 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end();
660 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
663 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
664 if (SI->getOperand(1) == Ptr) {
665 LoadValue = SI->getOperand(0);
669 assert(LoadValue && "No store through the pointer found!");
670 PointerValueNumbers.push_back(LoadValue); // Remember this for later.
673 // Load from the memory we are promoting.
674 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst);
676 if (LoadValue) CurAST->copyValue(LoadValue, LI);
678 // Store into the temporary alloca.
679 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
682 // Scan the basic blocks in the loop, replacing uses of our pointers with
683 // uses of the allocas in question.
685 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
686 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
687 E = LoopBBs.end(); I != E; ++I) {
688 // Rewrite all loads and stores in the block of the pointer...
689 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
691 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
692 std::map<Value*, AllocaInst*>::iterator
693 I = ValueToAllocaMap.find(L->getOperand(0));
694 if (I != ValueToAllocaMap.end())
695 L->setOperand(0, I->second); // Rewrite load instruction...
696 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
697 std::map<Value*, AllocaInst*>::iterator
698 I = ValueToAllocaMap.find(S->getOperand(1));
699 if (I != ValueToAllocaMap.end())
700 S->setOperand(1, I->second); // Rewrite store instruction...
705 // Now that the body of the loop uses the allocas instead of the original
706 // memory locations, insert code to copy the alloca value back into the
707 // original memory location on all exits from the loop. Note that we only
708 // want to insert one copy of the code in each exit block, though the loop may
709 // exit to the same block more than once.
711 std::set<BasicBlock*> ProcessedBlocks;
713 std::vector<BasicBlock*> ExitBlocks;
714 CurLoop->getExitBlocks(ExitBlocks);
715 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
716 if (ProcessedBlocks.insert(ExitBlocks[i]).second) {
717 // Copy all of the allocas into their memory locations.
718 BasicBlock::iterator BI = ExitBlocks[i]->begin();
719 while (isa<PHINode>(*BI))
720 ++BI; // Skip over all of the phi nodes in the block.
721 Instruction *InsertPos = BI;
723 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
724 // Load from the alloca.
725 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
727 // If this is a pointer type, update alias info appropriately.
728 if (isa<PointerType>(LI->getType()))
729 CurAST->copyValue(PointerValueNumbers[PVN++], LI);
731 // Store into the memory we promoted.
732 new StoreInst(LI, PromotedValues[i].second, InsertPos);
736 // Now that we have done the deed, use the mem2reg functionality to promote
737 // all of the new allocas we just created into real SSA registers.
739 std::vector<AllocaInst*> PromotedAllocas;
740 PromotedAllocas.reserve(PromotedValues.size());
741 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
742 PromotedAllocas.push_back(PromotedValues[i].first);
743 PromoteMemToReg(PromotedAllocas, *ET, *DF, CurAST);
746 /// FindPromotableValuesInLoop - Check the current loop for stores to definite
747 /// pointers, which are not loaded and stored through may aliases. If these are
748 /// found, create an alloca for the value, add it to the PromotedValues list,
749 /// and keep track of the mapping from value to alloca.
751 void LICM::FindPromotableValuesInLoop(
752 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
753 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
754 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
756 // Loop over all of the alias sets in the tracker object.
757 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
760 // We can promote this alias set if it has a store, if it is a "Must" alias
761 // set, if the pointer is loop invariant, and if we are not eliminating any
762 // volatile loads or stores.
763 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
764 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) {
765 assert(AS.begin() != AS.end() &&
766 "Must alias set should have at least one pointer element in it!");
767 Value *V = AS.begin()->first;
769 // Check that all of the pointers in the alias set have the same type. We
770 // cannot (yet) promote a memory location that is loaded and stored in
772 bool PointerOk = true;
773 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
774 if (V->getType() != I->first->getType()) {
780 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
781 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
782 PromotedValues.push_back(std::make_pair(AI, V));
784 // Update the AST and alias analysis.
785 CurAST->copyValue(V, AI);
787 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
788 ValueToAllocaMap.insert(std::make_pair(I->first, AI));
790 DOUT << "LICM: Promoting value: " << *V << "\n";