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/Analysis/ScalarEvolution.h"
46 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
47 #include "llvm/Support/CFG.h"
48 #include "llvm/Support/Compiler.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/ADT/Statistic.h"
55 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
56 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
57 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
58 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
59 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 struct VISIBILITY_HIDDEN LICM : public LoopPass {
67 static char ID; // Pass identification, replacement for typeid
68 LICM() : LoopPass((intptr_t)&ID) {}
70 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
72 /// This transformation requires natural loop information & requires that
73 /// loop preheaders be inserted into the CFG...
75 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 AU.addRequiredID(LoopSimplifyID);
78 AU.addRequired<LoopInfo>();
79 AU.addRequired<DominatorTree>();
80 AU.addRequired<DominanceFrontier>(); // For scalar promotion (mem2reg)
81 AU.addRequired<AliasAnalysis>();
82 AU.addPreserved<ScalarEvolution>();
83 AU.addPreserved<DominanceFrontier>();
86 bool doFinalization() {
87 // Free the values stored in the map
88 for (std::map<Loop *, AliasSetTracker *>::iterator
89 I = LoopToAliasMap.begin(), E = LoopToAliasMap.end(); I != E; ++I)
92 LoopToAliasMap.clear();
97 // Various analyses that we use...
98 AliasAnalysis *AA; // Current AliasAnalysis information
99 LoopInfo *LI; // Current LoopInfo
100 DominatorTree *DT; // Dominator Tree for the current Loop...
101 DominanceFrontier *DF; // Current Dominance Frontier
103 // State that is updated as we process loops
104 bool Changed; // Set to true when we change anything.
105 BasicBlock *Preheader; // The preheader block of the current loop...
106 Loop *CurLoop; // The current loop we are working on...
107 AliasSetTracker *CurAST; // AliasSet information for the current loop...
108 std::map<Loop *, AliasSetTracker *> LoopToAliasMap;
110 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
111 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L);
113 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
115 void deleteAnalysisValue(Value *V, Loop *L);
117 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
118 /// dominated by the specified block, and that are in the current loop) in
119 /// reverse depth first order w.r.t the DominatorTree. This allows us to
120 /// visit uses before definitions, allowing us to sink a loop body in one
121 /// pass without iteration.
123 void SinkRegion(DomTreeNode *N);
125 /// HoistRegion - Walk the specified region of the CFG (defined by all
126 /// blocks dominated by the specified block, and that are in the current
127 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
128 /// visit definitions before uses, allowing us to hoist a loop body in one
129 /// pass without iteration.
131 void HoistRegion(DomTreeNode *N);
133 /// inSubLoop - Little predicate that returns true if the specified basic
134 /// block is in a subloop of the current one, not the current one itself.
136 bool inSubLoop(BasicBlock *BB) {
137 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
138 for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
139 if ((*I)->contains(BB))
140 return true; // A subloop actually contains this block!
144 /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
145 /// specified exit block of the loop is dominated by the specified block
146 /// that is in the body of the loop. We use these constraints to
147 /// dramatically limit the amount of the dominator tree that needs to be
149 bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
150 BasicBlock *BlockInLoop) const {
151 // If the block in the loop is the loop header, it must be dominated!
152 BasicBlock *LoopHeader = CurLoop->getHeader();
153 if (BlockInLoop == LoopHeader)
156 DomTreeNode *BlockInLoopNode = DT->getNode(BlockInLoop);
157 DomTreeNode *IDom = DT->getNode(ExitBlock);
159 // Because the exit block is not in the loop, we know we have to get _at
160 // least_ its immediate dominator.
162 // Get next Immediate Dominator.
163 IDom = IDom->getIDom();
165 // If we have got to the header of the loop, then the instructions block
166 // did not dominate the exit node, so we can't hoist it.
167 if (IDom->getBlock() == LoopHeader)
170 } while (IDom != BlockInLoopNode);
175 /// sink - When an instruction is found to only be used outside of the loop,
176 /// this function moves it to the exit blocks and patches up SSA form as
179 void sink(Instruction &I);
181 /// hoist - When an instruction is found to only use loop invariant operands
182 /// that is safe to hoist, this instruction is called to do the dirty work.
184 void hoist(Instruction &I);
186 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
187 /// is not a trapping instruction or if it is a trapping instruction and is
188 /// guaranteed to execute.
190 bool isSafeToExecuteUnconditionally(Instruction &I);
192 /// pointerInvalidatedByLoop - Return true if the body of this loop may
193 /// store into the memory location pointed to by V.
195 bool pointerInvalidatedByLoop(Value *V, unsigned Size) {
196 // Check to see if any of the basic blocks in CurLoop invalidate *V.
197 return CurAST->getAliasSetForPointer(V, Size).isMod();
200 bool canSinkOrHoistInst(Instruction &I);
201 bool isLoopInvariantInst(Instruction &I);
202 bool isNotUsedInLoop(Instruction &I);
204 /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
205 /// to scalars as we can.
207 void PromoteValuesInLoop();
209 /// FindPromotableValuesInLoop - Check the current loop for stores to
210 /// definite pointers, which are not loaded and stored through may aliases.
211 /// If these are found, create an alloca for the value, add it to the
212 /// PromotedValues list, and keep track of the mapping from value to
215 void FindPromotableValuesInLoop(
216 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
217 std::map<Value*, AllocaInst*> &Val2AlMap);
221 RegisterPass<LICM> X("licm", "Loop Invariant Code Motion");
224 LoopPass *llvm::createLICMPass() { return new LICM(); }
226 /// Hoist expressions out of the specified loop. Note, alias info for inner
227 /// loop is not preserved so it is not a good idea to run LICM multiple
228 /// times on one loop.
230 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
233 // Get our Loop and Alias Analysis information...
234 LI = &getAnalysis<LoopInfo>();
235 AA = &getAnalysis<AliasAnalysis>();
236 DF = &getAnalysis<DominanceFrontier>();
237 DT = &getAnalysis<DominatorTree>();
239 CurAST = new AliasSetTracker(*AA);
240 // Collect Alias info from subloops
241 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
242 LoopItr != LoopItrE; ++LoopItr) {
243 Loop *InnerL = *LoopItr;
244 AliasSetTracker *InnerAST = LoopToAliasMap[InnerL];
245 assert (InnerAST && "Where is my AST?");
247 // What if InnerLoop was modified by other passes ?
248 CurAST->add(*InnerAST);
253 // Get the preheader block to move instructions into...
254 Preheader = L->getLoopPreheader();
255 assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
257 // Loop over the body of this loop, looking for calls, invokes, and stores.
258 // Because subloops have already been incorporated into AST, we skip blocks in
261 for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
262 E = L->getBlocks().end(); I != E; ++I)
263 if (LI->getLoopFor(*I) == L) // Ignore blocks in subloops...
264 CurAST->add(**I); // Incorporate the specified basic block
266 // We want to visit all of the instructions in this loop... that are not parts
267 // of our subloops (they have already had their invariants hoisted out of
268 // their loop, into this loop, so there is no need to process the BODIES of
271 // Traverse the body of the loop in depth first order on the dominator tree so
272 // that we are guaranteed to see definitions before we see uses. This allows
273 // us to sink instructions in one pass, without iteration. After sinking
274 // instructions, we perform another pass to hoist them out of the loop.
276 SinkRegion(DT->getNode(L->getHeader()));
277 HoistRegion(DT->getNode(L->getHeader()));
279 // Now that all loop invariants have been removed from the loop, promote any
280 // memory references to scalars that we can...
281 if (!DisablePromotion)
282 PromoteValuesInLoop();
284 // Clear out loops state information for the next iteration
288 LoopToAliasMap[L] = CurAST;
292 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
293 /// dominated by the specified block, and that are in the current loop) in
294 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
295 /// uses before definitions, allowing us to sink a loop body in one pass without
298 void LICM::SinkRegion(DomTreeNode *N) {
299 assert(N != 0 && "Null dominator tree node?");
300 BasicBlock *BB = N->getBlock();
302 // If this subregion is not in the top level loop at all, exit.
303 if (!CurLoop->contains(BB)) return;
305 // We are processing blocks in reverse dfo, so process children first...
306 const std::vector<DomTreeNode*> &Children = N->getChildren();
307 for (unsigned i = 0, e = Children.size(); i != e; ++i)
308 SinkRegion(Children[i]);
310 // Only need to process the contents of this block if it is not part of a
311 // subloop (which would already have been processed).
312 if (inSubLoop(BB)) return;
314 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
315 Instruction &I = *--II;
317 // Check to see if we can sink this instruction to the exit blocks
318 // of the loop. We can do this if the all users of the instruction are
319 // outside of the loop. In this case, it doesn't even matter if the
320 // operands of the instruction are loop invariant.
322 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
330 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
331 /// dominated by the specified block, and that are in the current loop) in depth
332 /// first order w.r.t the DominatorTree. This allows us to visit definitions
333 /// before uses, allowing us to hoist a loop body in one pass without iteration.
335 void LICM::HoistRegion(DomTreeNode *N) {
336 assert(N != 0 && "Null dominator tree node?");
337 BasicBlock *BB = N->getBlock();
339 // If this subregion is not in the top level loop at all, exit.
340 if (!CurLoop->contains(BB)) return;
342 // Only need to process the contents of this block if it is not part of a
343 // subloop (which would already have been processed).
345 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
346 Instruction &I = *II++;
348 // Try hoisting the instruction out to the preheader. We can only do this
349 // if all of the operands of the instruction are loop invariant and if it
350 // is safe to hoist the instruction.
352 if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
353 isSafeToExecuteUnconditionally(I))
357 const std::vector<DomTreeNode*> &Children = N->getChildren();
358 for (unsigned i = 0, e = Children.size(); i != e; ++i)
359 HoistRegion(Children[i]);
362 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
365 bool LICM::canSinkOrHoistInst(Instruction &I) {
366 // Loads have extra constraints we have to verify before we can hoist them.
367 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
368 if (LI->isVolatile())
369 return false; // Don't hoist volatile loads!
371 // Don't hoist loads which have may-aliased stores in loop.
373 if (LI->getType()->isSized())
374 Size = AA->getTargetData().getTypeStoreSize(LI->getType());
375 return !pointerInvalidatedByLoop(LI->getOperand(0), Size);
376 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
377 // Handle obvious cases efficiently.
378 if (Function *Callee = CI->getCalledFunction()) {
379 AliasAnalysis::ModRefBehavior Behavior =AA->getModRefBehavior(Callee, CI);
380 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
382 else if (Behavior == AliasAnalysis::OnlyReadsMemory) {
383 // If this call only reads from memory and there are no writes to memory
384 // in the loop, we can hoist or sink the call as appropriate.
385 bool FoundMod = false;
386 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
389 if (!AS.isForwardingAliasSet() && AS.isMod()) {
394 if (!FoundMod) return true;
398 // FIXME: This should use mod/ref information to see if we can hoist or sink
404 // Otherwise these instructions are hoistable/sinkable
405 return isa<BinaryOperator>(I) || isa<CastInst>(I) ||
406 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
407 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
408 isa<ShuffleVectorInst>(I);
411 /// isNotUsedInLoop - Return true if the only users of this instruction are
412 /// outside of the loop. If this is true, we can sink the instruction to the
413 /// exit blocks of the loop.
415 bool LICM::isNotUsedInLoop(Instruction &I) {
416 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
417 Instruction *User = cast<Instruction>(*UI);
418 if (PHINode *PN = dyn_cast<PHINode>(User)) {
419 // PHI node uses occur in predecessor blocks!
420 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
421 if (PN->getIncomingValue(i) == &I)
422 if (CurLoop->contains(PN->getIncomingBlock(i)))
424 } else if (CurLoop->contains(User->getParent())) {
432 /// isLoopInvariantInst - Return true if all operands of this instruction are
433 /// loop invariant. We also filter out non-hoistable instructions here just for
436 bool LICM::isLoopInvariantInst(Instruction &I) {
437 // The instruction is loop invariant if all of its operands are loop-invariant
438 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
439 if (!CurLoop->isLoopInvariant(I.getOperand(i)))
442 // If we got this far, the instruction is loop invariant!
446 /// sink - When an instruction is found to only be used outside of the loop,
447 /// this function moves it to the exit blocks and patches up SSA form as needed.
448 /// This method is guaranteed to remove the original instruction from its
449 /// position, and may either delete it or move it to outside of the loop.
451 void LICM::sink(Instruction &I) {
452 DOUT << "LICM sinking instruction: " << I;
454 SmallVector<BasicBlock*, 8> ExitBlocks;
455 CurLoop->getExitBlocks(ExitBlocks);
457 if (isa<LoadInst>(I)) ++NumMovedLoads;
458 else if (isa<CallInst>(I)) ++NumMovedCalls;
462 // The case where there is only a single exit node of this loop is common
463 // enough that we handle it as a special (more efficient) case. It is more
464 // efficient to handle because there are no PHI nodes that need to be placed.
465 if (ExitBlocks.size() == 1) {
466 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
467 // Instruction is not used, just delete it.
468 CurAST->deleteValue(&I);
469 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
470 I.replaceAllUsesWith(UndefValue::get(I.getType()));
473 // Move the instruction to the start of the exit block, after any PHI
475 I.removeFromParent();
477 BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
478 while (isa<PHINode>(InsertPt)) ++InsertPt;
479 ExitBlocks[0]->getInstList().insert(InsertPt, &I);
481 } else if (ExitBlocks.size() == 0) {
482 // The instruction is actually dead if there ARE NO exit blocks.
483 CurAST->deleteValue(&I);
484 if (!I.use_empty()) // If I has users in unreachable blocks, eliminate.
485 I.replaceAllUsesWith(UndefValue::get(I.getType()));
488 // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
489 // do all of the hard work of inserting PHI nodes as necessary. We convert
490 // the value into a stack object to get it to do this.
492 // Firstly, we create a stack object to hold the value...
495 if (I.getType() != Type::VoidTy) {
496 AI = new AllocaInst(I.getType(), 0, I.getName(),
497 I.getParent()->getParent()->getEntryBlock().begin());
501 // Secondly, insert load instructions for each use of the instruction
502 // outside of the loop.
503 while (!I.use_empty()) {
504 Instruction *U = cast<Instruction>(I.use_back());
506 // If the user is a PHI Node, we actually have to insert load instructions
507 // in all predecessor blocks, not in the PHI block itself!
508 if (PHINode *UPN = dyn_cast<PHINode>(U)) {
509 // Only insert into each predecessor once, so that we don't have
510 // different incoming values from the same block!
511 std::map<BasicBlock*, Value*> InsertedBlocks;
512 for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i)
513 if (UPN->getIncomingValue(i) == &I) {
514 BasicBlock *Pred = UPN->getIncomingBlock(i);
515 Value *&PredVal = InsertedBlocks[Pred];
517 // Insert a new load instruction right before the terminator in
518 // the predecessor block.
519 PredVal = new LoadInst(AI, "", Pred->getTerminator());
520 CurAST->add(cast<LoadInst>(PredVal));
523 UPN->setIncomingValue(i, PredVal);
527 LoadInst *L = new LoadInst(AI, "", U);
528 U->replaceUsesOfWith(&I, L);
533 // Thirdly, insert a copy of the instruction in each exit block of the loop
534 // that is dominated by the instruction, storing the result into the memory
535 // location. Be careful not to insert the instruction into any particular
536 // basic block more than once.
537 std::set<BasicBlock*> InsertedBlocks;
538 BasicBlock *InstOrigBB = I.getParent();
540 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
541 BasicBlock *ExitBlock = ExitBlocks[i];
543 if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) {
544 // If we haven't already processed this exit block, do so now.
545 if (InsertedBlocks.insert(ExitBlock).second) {
546 // Insert the code after the last PHI node...
547 BasicBlock::iterator InsertPt = ExitBlock->begin();
548 while (isa<PHINode>(InsertPt)) ++InsertPt;
550 // If this is the first exit block processed, just move the original
551 // instruction, otherwise clone the original instruction and insert
554 if (InsertedBlocks.size() == 1) {
555 I.removeFromParent();
556 ExitBlock->getInstList().insert(InsertPt, &I);
560 CurAST->copyValue(&I, New);
561 if (!I.getName().empty())
562 New->setName(I.getName()+".le");
563 ExitBlock->getInstList().insert(InsertPt, New);
566 // Now that we have inserted the instruction, store it into the alloca
567 if (AI) new StoreInst(New, AI, InsertPt);
572 // If the instruction doesn't dominate any exit blocks, it must be dead.
573 if (InsertedBlocks.empty()) {
574 CurAST->deleteValue(&I);
578 // Finally, promote the fine value to SSA form.
580 std::vector<AllocaInst*> Allocas;
581 Allocas.push_back(AI);
582 PromoteMemToReg(Allocas, *DT, *DF, CurAST);
587 /// hoist - When an instruction is found to only use loop invariant operands
588 /// that is safe to hoist, this instruction is called to do the dirty work.
590 void LICM::hoist(Instruction &I) {
591 DOUT << "LICM hoisting to " << Preheader->getName() << ": " << I;
593 // Remove the instruction from its current basic block... but don't delete the
595 I.removeFromParent();
597 // Insert the new node in Preheader, before the terminator.
598 Preheader->getInstList().insert(Preheader->getTerminator(), &I);
600 if (isa<LoadInst>(I)) ++NumMovedLoads;
601 else if (isa<CallInst>(I)) ++NumMovedCalls;
606 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
607 /// not a trapping instruction or if it is a trapping instruction and is
608 /// guaranteed to execute.
610 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
611 // If it is not a trapping instruction, it is always safe to hoist.
612 if (!Inst.isTrapping()) return true;
614 // Otherwise we have to check to make sure that the instruction dominates all
615 // of the exit blocks. If it doesn't, then there is a path out of the loop
616 // which does not execute this instruction, so we can't hoist it.
618 // If the instruction is in the header block for the loop (which is very
619 // common), it is always guaranteed to dominate the exit blocks. Since this
620 // is a common case, and can save some work, check it now.
621 if (Inst.getParent() == CurLoop->getHeader())
624 // It's always safe to load from a global or alloca.
625 if (isa<LoadInst>(Inst))
626 if (isa<AllocationInst>(Inst.getOperand(0)) ||
627 isa<GlobalVariable>(Inst.getOperand(0)))
630 // Get the exit blocks for the current loop.
631 SmallVector<BasicBlock*, 8> ExitBlocks;
632 CurLoop->getExitBlocks(ExitBlocks);
634 // For each exit block, get the DT node and walk up the DT until the
635 // instruction's basic block is found or we exit the loop.
636 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
637 if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
644 /// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
645 /// stores out of the loop and moving loads to before the loop. We do this by
646 /// looping over the stores in the loop, looking for stores to Must pointers
647 /// which are loop invariant. We promote these memory locations to use allocas
648 /// instead. These allocas can easily be raised to register values by the
649 /// PromoteMem2Reg functionality.
651 void LICM::PromoteValuesInLoop() {
652 // PromotedValues - List of values that are promoted out of the loop. Each
653 // value has an alloca instruction for it, and a canonical version of the
655 std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
656 std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
658 FindPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
659 if (ValueToAllocaMap.empty()) return; // If there are values to promote.
662 NumPromoted += PromotedValues.size();
664 std::vector<Value*> PointerValueNumbers;
666 // Emit a copy from the value into the alloca'd value in the loop preheader
667 TerminatorInst *LoopPredInst = Preheader->getTerminator();
668 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
669 Value *Ptr = PromotedValues[i].second;
671 // If we are promoting a pointer value, update alias information for the
673 Value *LoadValue = 0;
674 if (isa<PointerType>(cast<PointerType>(Ptr->getType())->getElementType())) {
675 // Locate a load or store through the pointer, and assign the same value
676 // to LI as we are loading or storing. Since we know that the value is
677 // stored in this loop, this will always succeed.
678 for (Value::use_iterator UI = Ptr->use_begin(), E = Ptr->use_end();
680 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
683 } else if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
684 if (SI->getOperand(1) == Ptr) {
685 LoadValue = SI->getOperand(0);
689 assert(LoadValue && "No store through the pointer found!");
690 PointerValueNumbers.push_back(LoadValue); // Remember this for later.
693 // Load from the memory we are promoting.
694 LoadInst *LI = new LoadInst(Ptr, Ptr->getName()+".promoted", LoopPredInst);
696 if (LoadValue) CurAST->copyValue(LoadValue, LI);
698 // Store into the temporary alloca.
699 new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
702 // Scan the basic blocks in the loop, replacing uses of our pointers with
703 // uses of the allocas in question.
705 const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
706 for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
707 E = LoopBBs.end(); I != E; ++I) {
708 // Rewrite all loads and stores in the block of the pointer...
709 for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
711 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
712 std::map<Value*, AllocaInst*>::iterator
713 I = ValueToAllocaMap.find(L->getOperand(0));
714 if (I != ValueToAllocaMap.end())
715 L->setOperand(0, I->second); // Rewrite load instruction...
716 } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
717 std::map<Value*, AllocaInst*>::iterator
718 I = ValueToAllocaMap.find(S->getOperand(1));
719 if (I != ValueToAllocaMap.end())
720 S->setOperand(1, I->second); // Rewrite store instruction...
725 // Now that the body of the loop uses the allocas instead of the original
726 // memory locations, insert code to copy the alloca value back into the
727 // original memory location on all exits from the loop. Note that we only
728 // want to insert one copy of the code in each exit block, though the loop may
729 // exit to the same block more than once.
731 std::set<BasicBlock*> ProcessedBlocks;
733 SmallVector<BasicBlock*, 8> ExitBlocks;
734 CurLoop->getExitBlocks(ExitBlocks);
735 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
736 if (ProcessedBlocks.insert(ExitBlocks[i]).second) {
737 // Copy all of the allocas into their memory locations.
738 BasicBlock::iterator BI = ExitBlocks[i]->begin();
739 while (isa<PHINode>(*BI))
740 ++BI; // Skip over all of the phi nodes in the block.
741 Instruction *InsertPos = BI;
743 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
744 // Load from the alloca.
745 LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
747 // If this is a pointer type, update alias info appropriately.
748 if (isa<PointerType>(LI->getType()))
749 CurAST->copyValue(PointerValueNumbers[PVN++], LI);
751 // Store into the memory we promoted.
752 new StoreInst(LI, PromotedValues[i].second, InsertPos);
756 // Now that we have done the deed, use the mem2reg functionality to promote
757 // all of the new allocas we just created into real SSA registers.
759 std::vector<AllocaInst*> PromotedAllocas;
760 PromotedAllocas.reserve(PromotedValues.size());
761 for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
762 PromotedAllocas.push_back(PromotedValues[i].first);
763 PromoteMemToReg(PromotedAllocas, *DT, *DF, CurAST);
766 /// FindPromotableValuesInLoop - Check the current loop for stores to definite
767 /// pointers, which are not loaded and stored through may aliases and are safe
768 /// for promotion. If these are found, create an alloca for the value, add it
769 /// to the PromotedValues list, and keep track of the mapping from value to
771 void LICM::FindPromotableValuesInLoop(
772 std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
773 std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
774 Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
776 SmallVector<Instruction *, 4> LoopExits;
777 SmallVector<BasicBlock *, 4> Blocks;
778 CurLoop->getExitingBlocks(Blocks);
779 for (SmallVector<BasicBlock *, 4>::iterator BI = Blocks.begin(),
780 BE = Blocks.end(); BI != BE; ++BI) {
781 BasicBlock *BB = *BI;
782 LoopExits.push_back(BB->getTerminator());
785 // Loop over all of the alias sets in the tracker object.
786 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
789 // We can promote this alias set if it has a store, if it is a "Must" alias
790 // set, if the pointer is loop invariant, and if we are not eliminating any
791 // volatile loads or stores.
792 if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
793 !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) {
794 assert(!AS.empty() &&
795 "Must alias set should have at least one pointer element in it!");
796 Value *V = AS.begin()->first;
798 // Check that all of the pointers in the alias set have the same type. We
799 // cannot (yet) promote a memory location that is loaded and stored in
801 bool PointerOk = true;
802 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
803 if (V->getType() != I->first->getType()) {
808 // If one use of value V inside the loop is safe then it is OK to promote
809 // this value. On the otherside if there is not any unsafe use inside the
810 // loop then also it is OK to promote this value. Otherwise it is
811 // unsafe to promote this value.
813 bool oneSafeUse = false;
814 bool oneUnsafeUse = false;
815 for(Value::use_iterator UI = V->use_begin(), UE = V->use_end();
817 Instruction *Use = dyn_cast<Instruction>(*UI);
818 if (!Use || !CurLoop->contains(Use->getParent()))
820 for (SmallVector<Instruction *, 4>::iterator
821 ExitI = LoopExits.begin(), ExitE = LoopExits.end();
822 ExitI != ExitE; ++ExitI) {
823 Instruction *Ex = *ExitI;
824 if (!isa<PHINode>(Use) && DT->dominates(Use, Ex)) {
838 else if (!oneUnsafeUse)
845 const Type *Ty = cast<PointerType>(V->getType())->getElementType();
846 AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
847 PromotedValues.push_back(std::make_pair(AI, V));
849 // Update the AST and alias analysis.
850 CurAST->copyValue(V, AI);
852 for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
853 ValueToAllocaMap.insert(std::make_pair(I->first, AI));
855 DOUT << "LICM: Promoting value: " << *V << "\n";
861 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
862 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
863 AliasSetTracker *AST = LoopToAliasMap[L];
867 AST->copyValue(From, To);
870 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
872 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
873 AliasSetTracker *AST = LoopToAliasMap[L];