1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs loop invariant code motion, attempting to remove as much
11 // code from the body of a loop as possible. It does this by either hoisting
12 // code into the preheader block, or by sinking code to the exit blocks if it is
13 // safe. This pass also promotes must-aliased memory locations in the loop to
14 // live in registers, thus hoisting and sinking "invariant" loads and stores.
16 // This pass uses alias analysis for two purposes:
18 // 1. Moving loop invariant loads and calls out of loops. If we can determine
19 // that a load or call inside of a loop never aliases anything stored to,
20 // we can hoist it or sink it like any other instruction.
21 // 2. Scalar Promotion of Memory - If there is a store instruction inside of
22 // the loop, we try to move the store to happen AFTER the loop instead of
23 // inside of the loop. This can only happen if a few conditions are true:
24 // A. The pointer stored through is loop invariant
25 // B. There are no stores or loads in the loop which _may_ alias the
26 // pointer. There are no calls in the loop which mod/ref the pointer.
27 // If these conditions are true, we can promote the loads and stores in the
28 // loop of the pointer to use a temporary alloca'd variable. We then use
29 // the SSAUpdater to construct the appropriate SSA form for the value.
31 //===----------------------------------------------------------------------===//
33 #define DEBUG_TYPE "licm"
34 #include "llvm/Transforms/Scalar.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/AliasSetTracker.h"
38 #include "llvm/Analysis/ConstantFolding.h"
39 #include "llvm/Analysis/LoopInfo.h"
40 #include "llvm/Analysis/LoopPass.h"
41 #include "llvm/Analysis/ScalarEvolution.h"
42 #include "llvm/Analysis/ValueTracking.h"
43 #include "llvm/IR/CFG.h"
44 #include "llvm/IR/Constants.h"
45 #include "llvm/IR/DataLayout.h"
46 #include "llvm/IR/DerivedTypes.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/Metadata.h"
52 #include "llvm/IR/PredIteratorCache.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Target/TargetLibraryInfo.h"
57 #include "llvm/Transforms/Utils/Local.h"
58 #include "llvm/Transforms/Utils/LoopUtils.h"
59 #include "llvm/Transforms/Utils/SSAUpdater.h"
63 STATISTIC(NumSunk , "Number of instructions sunk out of loop");
64 STATISTIC(NumHoisted , "Number of instructions hoisted out of loop");
65 STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk");
66 STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk");
67 STATISTIC(NumPromoted , "Number of memory locations promoted to registers");
70 DisablePromotion("disable-licm-promotion", cl::Hidden,
71 cl::desc("Disable memory promotion in LICM pass"));
74 struct LICM : public LoopPass {
75 static char ID; // Pass identification, replacement for typeid
76 LICM() : LoopPass(ID) {
77 initializeLICMPass(*PassRegistry::getPassRegistry());
80 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
82 /// This transformation requires natural loop information & requires that
83 /// loop preheaders be inserted into the CFG...
85 void getAnalysisUsage(AnalysisUsage &AU) const override {
87 AU.addRequired<DominatorTreeWrapperPass>();
88 AU.addRequired<LoopInfo>();
89 AU.addRequiredID(LoopSimplifyID);
90 AU.addPreservedID(LoopSimplifyID);
91 AU.addRequiredID(LCSSAID);
92 AU.addPreservedID(LCSSAID);
93 AU.addRequired<AliasAnalysis>();
94 AU.addPreserved<AliasAnalysis>();
95 AU.addPreserved<ScalarEvolution>();
96 AU.addRequired<TargetLibraryInfo>();
99 using llvm::Pass::doFinalization;
101 bool doFinalization() override {
102 assert(LoopToAliasSetMap.empty() && "Didn't free loop alias sets");
107 AliasAnalysis *AA; // Current AliasAnalysis information
108 LoopInfo *LI; // Current LoopInfo
109 DominatorTree *DT; // Dominator Tree for the current Loop.
111 const DataLayout *DL; // DataLayout for constant folding.
112 TargetLibraryInfo *TLI; // TargetLibraryInfo for constant folding.
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 bool MayThrow; // The current loop contains an instruction which
120 // may throw, thus preventing code motion of
121 // instructions with side effects.
122 DenseMap<Loop*, AliasSetTracker*> LoopToAliasSetMap;
124 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
125 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
128 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
130 void deleteAnalysisValue(Value *V, Loop *L) override;
132 /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
133 /// dominated by the specified block, and that are in the current loop) in
134 /// reverse depth first order w.r.t the DominatorTree. This allows us to
135 /// visit uses before definitions, allowing us to sink a loop body in one
136 /// pass without iteration.
138 void SinkRegion(DomTreeNode *N);
140 /// HoistRegion - Walk the specified region of the CFG (defined by all
141 /// blocks dominated by the specified block, and that are in the current
142 /// loop) in depth first order w.r.t the DominatorTree. This allows us to
143 /// visit definitions before uses, allowing us to hoist a loop body in one
144 /// pass without iteration.
146 void HoistRegion(DomTreeNode *N);
148 /// inSubLoop - Little predicate that returns true if the specified basic
149 /// block is in a subloop of the current one, not the current one itself.
151 bool inSubLoop(BasicBlock *BB) {
152 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
153 return LI->getLoopFor(BB) != CurLoop;
156 /// sink - When an instruction is found to only be used outside of the loop,
157 /// this function moves it to the exit blocks and patches up SSA form as
160 void sink(Instruction &I);
162 /// hoist - When an instruction is found to only use loop invariant operands
163 /// that is safe to hoist, this instruction is called to do the dirty work.
165 void hoist(Instruction &I);
167 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
168 /// is not a trapping instruction or if it is a trapping instruction and is
169 /// guaranteed to execute.
171 bool isSafeToExecuteUnconditionally(Instruction &I);
173 /// isGuaranteedToExecute - Check that the instruction is guaranteed to
176 bool isGuaranteedToExecute(Instruction &I);
178 /// pointerInvalidatedByLoop - Return true if the body of this loop may
179 /// store into the memory location pointed to by V.
181 bool pointerInvalidatedByLoop(Value *V, uint64_t Size,
182 const MDNode *TBAAInfo) {
183 // Check to see if any of the basic blocks in CurLoop invalidate *V.
184 return CurAST->getAliasSetForPointer(V, Size, TBAAInfo).isMod();
187 bool canSinkOrHoistInst(Instruction &I);
188 bool isNotUsedInLoop(Instruction &I);
190 void PromoteAliasSet(AliasSet &AS,
191 SmallVectorImpl<BasicBlock*> &ExitBlocks,
192 SmallVectorImpl<Instruction*> &InsertPts,
193 PredIteratorCache &PIC);
198 INITIALIZE_PASS_BEGIN(LICM, "licm", "Loop Invariant Code Motion", false, false)
199 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
200 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
201 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
202 INITIALIZE_PASS_DEPENDENCY(LCSSA)
203 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
204 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
205 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
206 INITIALIZE_PASS_END(LICM, "licm", "Loop Invariant Code Motion", false, false)
208 Pass *llvm::createLICMPass() { return new LICM(); }
210 /// Hoist expressions out of the specified loop. Note, alias info for inner
211 /// loop is not preserved so it is not a good idea to run LICM multiple
212 /// times on one loop.
214 bool LICM::runOnLoop(Loop *L, LPPassManager &LPM) {
215 if (skipOptnoneFunction(L))
220 // Get our Loop and Alias Analysis information...
221 LI = &getAnalysis<LoopInfo>();
222 AA = &getAnalysis<AliasAnalysis>();
223 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
225 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
226 DL = DLP ? &DLP->getDataLayout() : 0;
227 TLI = &getAnalysis<TargetLibraryInfo>();
229 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.");
231 CurAST = new AliasSetTracker(*AA);
232 // Collect Alias info from subloops.
233 for (Loop::iterator LoopItr = L->begin(), LoopItrE = L->end();
234 LoopItr != LoopItrE; ++LoopItr) {
235 Loop *InnerL = *LoopItr;
236 AliasSetTracker *InnerAST = LoopToAliasSetMap[InnerL];
237 assert(InnerAST && "Where is my AST?");
239 // What if InnerLoop was modified by other passes ?
240 CurAST->add(*InnerAST);
242 // Once we've incorporated the inner loop's AST into ours, we don't need the
243 // subloop's anymore.
245 LoopToAliasSetMap.erase(InnerL);
250 // Get the preheader block to move instructions into...
251 Preheader = L->getLoopPreheader();
253 // Loop over the body of this loop, looking for calls, invokes, and stores.
254 // Because subloops have already been incorporated into AST, we skip blocks in
257 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
260 if (LI->getLoopFor(BB) == L) // Ignore blocks in subloops.
261 CurAST->add(*BB); // Incorporate the specified basic block
265 // TODO: We've already searched for instructions which may throw in subloops.
266 // We may want to reuse this information.
267 for (Loop::block_iterator BB = L->block_begin(), BBE = L->block_end();
268 (BB != BBE) && !MayThrow ; ++BB)
269 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end();
270 (I != E) && !MayThrow; ++I)
271 MayThrow |= I->mayThrow();
273 // We want to visit all of the instructions in this loop... that are not parts
274 // of our subloops (they have already had their invariants hoisted out of
275 // their loop, into this loop, so there is no need to process the BODIES of
278 // Traverse the body of the loop in depth first order on the dominator tree so
279 // that we are guaranteed to see definitions before we see uses. This allows
280 // us to sink instructions in one pass, without iteration. After sinking
281 // instructions, we perform another pass to hoist them out of the loop.
283 if (L->hasDedicatedExits())
284 SinkRegion(DT->getNode(L->getHeader()));
286 HoistRegion(DT->getNode(L->getHeader()));
288 // Now that all loop invariants have been removed from the loop, promote any
289 // memory references to scalars that we can.
290 if (!DisablePromotion && (Preheader || L->hasDedicatedExits())) {
291 SmallVector<BasicBlock *, 8> ExitBlocks;
292 SmallVector<Instruction *, 8> InsertPts;
293 PredIteratorCache PIC;
295 // Loop over all of the alias sets in the tracker object.
296 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
298 PromoteAliasSet(*I, ExitBlocks, InsertPts, PIC);
300 // Once we have promoted values across the loop body we have to recursively
301 // reform LCSSA as any nested loop may now have values defined within the
302 // loop used in the outer loop.
303 // FIXME: This is really heavy handed. It would be a bit better to use an
304 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
307 formLCSSARecursively(*L, *DT, getAnalysisIfAvailable<ScalarEvolution>());
310 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
311 // specifically moving instructions across the loop boundary and so it is
312 // especially in need of sanity checking here.
313 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!");
314 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&
315 "Parent loop not left in LCSSA form after LICM!");
317 // Clear out loops state information for the next iteration
321 // If this loop is nested inside of another one, save the alias information
322 // for when we process the outer loop.
323 if (L->getParentLoop())
324 LoopToAliasSetMap[L] = CurAST;
330 /// SinkRegion - 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
332 /// reverse depth first order w.r.t the DominatorTree. This allows us to visit
333 /// uses before definitions, allowing us to sink a loop body in one pass without
336 void LICM::SinkRegion(DomTreeNode *N) {
337 assert(N != 0 && "Null dominator tree node?");
338 BasicBlock *BB = N->getBlock();
340 // If this subregion is not in the top level loop at all, exit.
341 if (!CurLoop->contains(BB)) return;
343 // We are processing blocks in reverse dfo, so process children first.
344 const std::vector<DomTreeNode*> &Children = N->getChildren();
345 for (unsigned i = 0, e = Children.size(); i != e; ++i)
346 SinkRegion(Children[i]);
348 // Only need to process the contents of this block if it is not part of a
349 // subloop (which would already have been processed).
350 if (inSubLoop(BB)) return;
352 for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
353 Instruction &I = *--II;
355 // If the instruction is dead, we would try to sink it because it isn't used
356 // in the loop, instead, just delete it.
357 if (isInstructionTriviallyDead(&I, TLI)) {
358 DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
360 CurAST->deleteValue(&I);
366 // Check to see if we can sink this instruction to the exit blocks
367 // of the loop. We can do this if the all users of the instruction are
368 // outside of the loop. In this case, it doesn't even matter if the
369 // operands of the instruction are loop invariant.
371 if (isNotUsedInLoop(I) && canSinkOrHoistInst(I)) {
378 /// HoistRegion - Walk the specified region of the CFG (defined by all blocks
379 /// dominated by the specified block, and that are in the current loop) in depth
380 /// first order w.r.t the DominatorTree. This allows us to visit definitions
381 /// before uses, allowing us to hoist a loop body in one pass without iteration.
383 void LICM::HoistRegion(DomTreeNode *N) {
384 assert(N != 0 && "Null dominator tree node?");
385 BasicBlock *BB = N->getBlock();
387 // If this subregion is not in the top level loop at all, exit.
388 if (!CurLoop->contains(BB)) return;
390 // Only need to process the contents of this block if it is not part of a
391 // subloop (which would already have been processed).
393 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
394 Instruction &I = *II++;
396 // Try constant folding this instruction. If all the operands are
397 // constants, it is technically hoistable, but it would be better to just
399 if (Constant *C = ConstantFoldInstruction(&I, DL, TLI)) {
400 DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *C << '\n');
401 CurAST->copyValue(&I, C);
402 CurAST->deleteValue(&I);
403 I.replaceAllUsesWith(C);
408 // Try hoisting the instruction out to the preheader. We can only do this
409 // if all of the operands of the instruction are loop invariant and if it
410 // is safe to hoist the instruction.
412 if (CurLoop->hasLoopInvariantOperands(&I) && canSinkOrHoistInst(I) &&
413 isSafeToExecuteUnconditionally(I))
417 const std::vector<DomTreeNode*> &Children = N->getChildren();
418 for (unsigned i = 0, e = Children.size(); i != e; ++i)
419 HoistRegion(Children[i]);
422 /// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
425 bool LICM::canSinkOrHoistInst(Instruction &I) {
426 // Loads have extra constraints we have to verify before we can hoist them.
427 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
428 if (!LI->isUnordered())
429 return false; // Don't hoist volatile/atomic loads!
431 // Loads from constant memory are always safe to move, even if they end up
432 // in the same alias set as something that ends up being modified.
433 if (AA->pointsToConstantMemory(LI->getOperand(0)))
435 if (LI->getMetadata("invariant.load"))
438 // Don't hoist loads which have may-aliased stores in loop.
440 if (LI->getType()->isSized())
441 Size = AA->getTypeStoreSize(LI->getType());
442 return !pointerInvalidatedByLoop(LI->getOperand(0), Size,
443 LI->getMetadata(LLVMContext::MD_tbaa));
444 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
445 // Don't sink or hoist dbg info; it's legal, but not useful.
446 if (isa<DbgInfoIntrinsic>(I))
449 // Handle simple cases by querying alias analysis.
450 AliasAnalysis::ModRefBehavior Behavior = AA->getModRefBehavior(CI);
451 if (Behavior == AliasAnalysis::DoesNotAccessMemory)
453 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
454 // If this call only reads from memory and there are no writes to memory
455 // in the loop, we can hoist or sink the call as appropriate.
456 bool FoundMod = false;
457 for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
460 if (!AS.isForwardingAliasSet() && AS.isMod()) {
465 if (!FoundMod) return true;
468 // FIXME: This should use mod/ref information to see if we can hoist or
474 // Only these instructions are hoistable/sinkable.
475 if (!isa<BinaryOperator>(I) && !isa<CastInst>(I) && !isa<SelectInst>(I) &&
476 !isa<GetElementPtrInst>(I) && !isa<CmpInst>(I) &&
477 !isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
478 !isa<ShuffleVectorInst>(I) && !isa<ExtractValueInst>(I) &&
479 !isa<InsertValueInst>(I))
482 return isSafeToExecuteUnconditionally(I);
485 /// \brief Returns true if a PHINode is a trivially replaceable with an
488 /// This is true when all incoming values are that instruction. This pattern
489 /// occurs most often with LCSSA PHI nodes.
490 static bool isTriviallyReplacablePHI(PHINode &PN, Instruction &I) {
491 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
492 if (PN.getIncomingValue(i) != &I)
498 /// isNotUsedInLoop - Return true if the only users of this instruction are
499 /// outside of the loop. If this is true, we can sink the instruction to the
500 /// exit blocks of the loop.
502 bool LICM::isNotUsedInLoop(Instruction &I) {
503 for (User *U : I.users()) {
504 Instruction *UI = cast<Instruction>(U);
505 if (PHINode *PN = dyn_cast<PHINode>(UI)) {
506 // A PHI node where all of the incoming values are this instruction are
507 // special -- they can just be RAUW'ed with the instruction and thus
508 // don't require a use in the predecessor. This is a particular important
509 // special case because it is the pattern found in LCSSA form.
510 if (isTriviallyReplacablePHI(*PN, I)) {
511 if (CurLoop->contains(PN))
517 // Otherwise, PHI node uses occur in predecessor blocks if the incoming
518 // values. Check for such a use being inside the loop.
519 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
520 if (PN->getIncomingValue(i) == &I)
521 if (CurLoop->contains(PN->getIncomingBlock(i)))
527 if (CurLoop->contains(UI))
533 /// sink - When an instruction is found to only be used outside of the loop,
534 /// this function moves it to the exit blocks and patches up SSA form as needed.
535 /// This method is guaranteed to remove the original instruction from its
536 /// position, and may either delete it or move it to outside of the loop.
538 void LICM::sink(Instruction &I) {
539 DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n");
541 if (isa<LoadInst>(I)) ++NumMovedLoads;
542 else if (isa<CallInst>(I)) ++NumMovedCalls;
547 SmallVector<BasicBlock *, 32> ExitBlocks;
548 CurLoop->getUniqueExitBlocks(ExitBlocks);
549 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
552 // If this instruction is only used outside of the loop, then all users are
553 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
555 while (!I.use_empty()) {
556 // The user must be a PHI node.
557 PHINode *PN = cast<PHINode>(I.user_back());
559 BasicBlock *ExitBlock = PN->getParent();
560 assert(ExitBlockSet.count(ExitBlock) &&
561 "The LCSSA PHI is not in an exit block!");
563 Instruction *New = I.clone();
564 ExitBlock->getInstList().insert(ExitBlock->getFirstInsertionPt(), New);
565 if (!I.getName().empty())
566 New->setName(I.getName() + ".le");
568 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
569 // particularly cheap because we can rip off the PHI node that we're
570 // replacing for the number and blocks of the predecessors.
571 // OPT: If this shows up in a profile, we can instead finish sinking all
572 // invariant instructions, and then walk their operands to re-establish
573 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
574 // sinking bottom-up.
575 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
577 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
578 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
579 if (!OLoop->contains(PN)) {
580 PHINode *OpPN = PHINode::Create(
581 OInst->getType(), PN->getNumIncomingValues(),
582 OInst->getName() + ".lcssa", ExitBlock->begin());
583 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
584 OpPN->addIncoming(OInst, PN->getIncomingBlock(i));
588 PN->replaceAllUsesWith(New);
589 PN->eraseFromParent();
592 CurAST->deleteValue(&I);
596 /// hoist - When an instruction is found to only use loop invariant operands
597 /// that is safe to hoist, this instruction is called to do the dirty work.
599 void LICM::hoist(Instruction &I) {
600 DEBUG(dbgs() << "LICM hoisting to " << Preheader->getName() << ": "
603 // Move the new node to the Preheader, before its terminator.
604 I.moveBefore(Preheader->getTerminator());
606 if (isa<LoadInst>(I)) ++NumMovedLoads;
607 else if (isa<CallInst>(I)) ++NumMovedCalls;
612 /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
613 /// not a trapping instruction or if it is a trapping instruction and is
614 /// guaranteed to execute.
616 bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
617 // If it is not a trapping instruction, it is always safe to hoist.
618 if (isSafeToSpeculativelyExecute(&Inst))
621 return isGuaranteedToExecute(Inst);
624 bool LICM::isGuaranteedToExecute(Instruction &Inst) {
626 // Somewhere in this loop there is an instruction which may throw and make us
631 // Otherwise we have to check to make sure that the instruction dominates all
632 // of the exit blocks. If it doesn't, then there is a path out of the loop
633 // which does not execute this instruction, so we can't hoist it.
635 // If the instruction is in the header block for the loop (which is very
636 // common), it is always guaranteed to dominate the exit blocks. Since this
637 // is a common case, and can save some work, check it now.
638 if (Inst.getParent() == CurLoop->getHeader())
641 // Get the exit blocks for the current loop.
642 SmallVector<BasicBlock*, 8> ExitBlocks;
643 CurLoop->getExitBlocks(ExitBlocks);
645 // Verify that the block dominates each of the exit blocks of the loop.
646 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
647 if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
650 // As a degenerate case, if the loop is statically infinite then we haven't
651 // proven anything since there are no exit blocks.
652 if (ExitBlocks.empty())
659 class LoopPromoter : public LoadAndStorePromoter {
660 Value *SomePtr; // Designated pointer to store to.
661 SmallPtrSet<Value*, 4> &PointerMustAliases;
662 SmallVectorImpl<BasicBlock*> &LoopExitBlocks;
663 SmallVectorImpl<Instruction*> &LoopInsertPts;
664 PredIteratorCache &PredCache;
665 AliasSetTracker &AST;
671 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
672 if (Instruction *I = dyn_cast<Instruction>(V))
673 if (Loop *L = LI.getLoopFor(I->getParent()))
674 if (!L->contains(BB)) {
675 // We need to create an LCSSA PHI node for the incoming value and
677 PHINode *PN = PHINode::Create(
678 I->getType(), PredCache.GetNumPreds(BB),
679 I->getName() + ".lcssa", BB->begin());
680 for (BasicBlock **PI = PredCache.GetPreds(BB); *PI; ++PI)
681 PN->addIncoming(I, *PI);
688 LoopPromoter(Value *SP, const SmallVectorImpl<Instruction *> &Insts,
689 SSAUpdater &S, SmallPtrSet<Value *, 4> &PMA,
690 SmallVectorImpl<BasicBlock *> &LEB,
691 SmallVectorImpl<Instruction *> &LIP, PredIteratorCache &PIC,
692 AliasSetTracker &ast, LoopInfo &li, DebugLoc dl, int alignment,
694 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
695 LoopExitBlocks(LEB), LoopInsertPts(LIP), PredCache(PIC), AST(ast),
696 LI(li), DL(dl), Alignment(alignment), TBAATag(TBAATag) {}
698 bool isInstInList(Instruction *I,
699 const SmallVectorImpl<Instruction*> &) const override {
701 if (LoadInst *LI = dyn_cast<LoadInst>(I))
702 Ptr = LI->getOperand(0);
704 Ptr = cast<StoreInst>(I)->getPointerOperand();
705 return PointerMustAliases.count(Ptr);
708 void doExtraRewritesBeforeFinalDeletion() const override {
709 // Insert stores after in the loop exit blocks. Each exit block gets a
710 // store of the live-out values that feed them. Since we've already told
711 // the SSA updater about the defs in the loop and the preheader
712 // definition, it is all set and we can start using it.
713 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
714 BasicBlock *ExitBlock = LoopExitBlocks[i];
715 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
716 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
717 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
718 Instruction *InsertPos = LoopInsertPts[i];
719 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
720 NewSI->setAlignment(Alignment);
721 NewSI->setDebugLoc(DL);
722 if (TBAATag) NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag);
726 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
727 // Update alias analysis.
728 AST.copyValue(LI, V);
730 void instructionDeleted(Instruction *I) const override {
734 } // end anon namespace
736 /// PromoteAliasSet - Try to promote memory values to scalars by sinking
737 /// stores out of the loop and moving loads to before the loop. We do this by
738 /// looping over the stores in the loop, looking for stores to Must pointers
739 /// which are loop invariant.
741 void LICM::PromoteAliasSet(AliasSet &AS,
742 SmallVectorImpl<BasicBlock*> &ExitBlocks,
743 SmallVectorImpl<Instruction*> &InsertPts,
744 PredIteratorCache &PIC) {
745 // We can promote this alias set if it has a store, if it is a "Must" alias
746 // set, if the pointer is loop invariant, and if we are not eliminating any
747 // volatile loads or stores.
748 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
749 AS.isVolatile() || !CurLoop->isLoopInvariant(AS.begin()->getValue()))
752 assert(!AS.empty() &&
753 "Must alias set should have at least one pointer element in it!");
754 Value *SomePtr = AS.begin()->getValue();
756 // It isn't safe to promote a load/store from the loop if the load/store is
757 // conditional. For example, turning:
759 // for () { if (c) *P += 1; }
763 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
765 // is not safe, because *P may only be valid to access if 'c' is true.
767 // It is safe to promote P if all uses are direct load/stores and if at
768 // least one is guaranteed to be executed.
769 bool GuaranteedToExecute = false;
771 SmallVector<Instruction*, 64> LoopUses;
772 SmallPtrSet<Value*, 4> PointerMustAliases;
774 // We start with an alignment of one and try to find instructions that allow
775 // us to prove better alignment.
776 unsigned Alignment = 1;
779 // Check that all of the pointers in the alias set have the same type. We
780 // cannot (yet) promote a memory location that is loaded and stored in
781 // different sizes. While we are at it, collect alignment and TBAA info.
782 for (AliasSet::iterator ASI = AS.begin(), E = AS.end(); ASI != E; ++ASI) {
783 Value *ASIV = ASI->getValue();
784 PointerMustAliases.insert(ASIV);
786 // Check that all of the pointers in the alias set have the same type. We
787 // cannot (yet) promote a memory location that is loaded and stored in
789 if (SomePtr->getType() != ASIV->getType())
792 for (User *U : ASIV->users()) {
793 // Ignore instructions that are outside the loop.
794 Instruction *UI = dyn_cast<Instruction>(U);
795 if (!UI || !CurLoop->contains(UI))
798 // If there is an non-load/store instruction in the loop, we can't promote
800 if (LoadInst *load = dyn_cast<LoadInst>(UI)) {
801 assert(!load->isVolatile() && "AST broken");
802 if (!load->isSimple())
804 } else if (StoreInst *store = dyn_cast<StoreInst>(UI)) {
805 // Stores *of* the pointer are not interesting, only stores *to* the
807 if (UI->getOperand(1) != ASIV)
809 assert(!store->isVolatile() && "AST broken");
810 if (!store->isSimple())
813 // Note that we only check GuaranteedToExecute inside the store case
814 // so that we do not introduce stores where they did not exist before
815 // (which would break the LLVM concurrency model).
817 // If the alignment of this instruction allows us to specify a more
818 // restrictive (and performant) alignment and if we are sure this
819 // instruction will be executed, update the alignment.
820 // Larger is better, with the exception of 0 being the best alignment.
821 unsigned InstAlignment = store->getAlignment();
822 if ((InstAlignment > Alignment || InstAlignment == 0) && Alignment != 0)
823 if (isGuaranteedToExecute(*UI)) {
824 GuaranteedToExecute = true;
825 Alignment = InstAlignment;
828 if (!GuaranteedToExecute)
829 GuaranteedToExecute = isGuaranteedToExecute(*UI);
832 return; // Not a load or store.
834 // Merge the TBAA tags.
835 if (LoopUses.empty()) {
836 // On the first load/store, just take its TBAA tag.
837 TBAATag = UI->getMetadata(LLVMContext::MD_tbaa);
838 } else if (TBAATag) {
839 TBAATag = MDNode::getMostGenericTBAA(TBAATag,
840 UI->getMetadata(LLVMContext::MD_tbaa));
843 LoopUses.push_back(UI);
847 // If there isn't a guaranteed-to-execute instruction, we can't promote.
848 if (!GuaranteedToExecute)
851 // Otherwise, this is safe to promote, lets do it!
852 DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " <<*SomePtr<<'\n');
856 // Grab a debug location for the inserted loads/stores; given that the
857 // inserted loads/stores have little relation to the original loads/stores,
858 // this code just arbitrarily picks a location from one, since any debug
859 // location is better than none.
860 DebugLoc DL = LoopUses[0]->getDebugLoc();
862 // Figure out the loop exits and their insertion points, if this is the
864 if (ExitBlocks.empty()) {
865 CurLoop->getUniqueExitBlocks(ExitBlocks);
866 InsertPts.resize(ExitBlocks.size());
867 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
868 InsertPts[i] = ExitBlocks[i]->getFirstInsertionPt();
871 // We use the SSAUpdater interface to insert phi nodes as required.
872 SmallVector<PHINode*, 16> NewPHIs;
873 SSAUpdater SSA(&NewPHIs);
874 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
875 InsertPts, PIC, *CurAST, *LI, DL, Alignment, TBAATag);
877 // Set up the preheader to have a definition of the value. It is the live-out
878 // value from the preheader that uses in the loop will use.
879 LoadInst *PreheaderLoad =
880 new LoadInst(SomePtr, SomePtr->getName()+".promoted",
881 Preheader->getTerminator());
882 PreheaderLoad->setAlignment(Alignment);
883 PreheaderLoad->setDebugLoc(DL);
884 if (TBAATag) PreheaderLoad->setMetadata(LLVMContext::MD_tbaa, TBAATag);
885 SSA.AddAvailableValue(Preheader, PreheaderLoad);
887 // Rewrite all the loads in the loop and remember all the definitions from
888 // stores in the loop.
889 Promoter.run(LoopUses);
891 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
892 if (PreheaderLoad->use_empty())
893 PreheaderLoad->eraseFromParent();
897 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
898 void LICM::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To, Loop *L) {
899 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);
903 AST->copyValue(From, To);
906 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
908 void LICM::deleteAnalysisValue(Value *V, Loop *L) {
909 AliasSetTracker *AST = LoopToAliasSetMap.lookup(L);