1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
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 file implements some loop unrolling utilities. It does not define any
11 // actual pass or policy, but provides a single function to perform loop
14 // The process of unrolling can produce extraneous basic blocks linked with
15 // unconditional branches. This will be corrected in the future.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopIterator.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
26 #include "llvm/Analysis/ScalarEvolution.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DebugInfoMetadata.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Cloning.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Transforms/Utils/LoopSimplify.h"
39 #include "llvm/Transforms/Utils/LoopUtils.h"
40 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
41 #include "llvm/Transforms/Utils/UnrollLoop.h"
44 #define DEBUG_TYPE "loop-unroll"
46 // TODO: Should these be here or in LoopUnroll?
47 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
48 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
51 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden,
52 cl::desc("Allow runtime unrolled loops to be unrolled "
53 "with epilog instead of prolog."));
56 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden,
57 cl::desc("Verify domtree after unrolling"),
65 /// Convert the instruction operands from referencing the current values into
66 /// those specified by VMap.
67 static inline void remapInstruction(Instruction *I,
68 ValueToValueMapTy &VMap) {
69 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
70 Value *Op = I->getOperand(op);
71 ValueToValueMapTy::iterator It = VMap.find(Op);
73 I->setOperand(op, It->second);
76 if (PHINode *PN = dyn_cast<PHINode>(I)) {
77 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
78 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
80 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
85 /// Folds a basic block into its predecessor if it only has one predecessor, and
86 /// that predecessor only has one successor.
87 /// The LoopInfo Analysis that is passed will be kept consistent. If folding is
88 /// successful references to the containing loop must be removed from
89 /// ScalarEvolution by calling ScalarEvolution::forgetLoop because SE may have
90 /// references to the eliminated BB. The argument ForgottenLoops contains a set
91 /// of loops that have already been forgotten to prevent redundant, expensive
92 /// calls to ScalarEvolution::forgetLoop. Returns the new combined block.
94 foldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, ScalarEvolution *SE,
95 SmallPtrSetImpl<Loop *> &ForgottenLoops,
97 // Merge basic blocks into their predecessor if there is only one distinct
98 // pred, and if there is only one distinct successor of the predecessor, and
99 // if there are no PHI nodes.
100 BasicBlock *OnlyPred = BB->getSinglePredecessor();
101 if (!OnlyPred) return nullptr;
103 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
106 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
108 // Resolve any PHI nodes at the start of the block. They are all
109 // guaranteed to have exactly one entry if they exist, unless there are
110 // multiple duplicate (but guaranteed to be equal) entries for the
111 // incoming edges. This occurs when there are multiple edges from
112 // OnlyPred to OnlySucc.
113 FoldSingleEntryPHINodes(BB);
115 // Delete the unconditional branch from the predecessor...
116 OnlyPred->getInstList().pop_back();
118 // Make all PHI nodes that referred to BB now refer to Pred as their
120 BB->replaceAllUsesWith(OnlyPred);
122 // Move all definitions in the successor to the predecessor...
123 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
125 // OldName will be valid until erased.
126 StringRef OldName = BB->getName();
128 // Erase the old block and update dominator info.
130 if (DomTreeNode *DTN = DT->getNode(BB)) {
131 DomTreeNode *PredDTN = DT->getNode(OnlyPred);
132 SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end());
133 for (auto *DI : Children)
134 DT->changeImmediateDominator(DI, PredDTN);
139 // ScalarEvolution holds references to loop exit blocks.
141 if (Loop *L = LI->getLoopFor(BB)) {
142 if (ForgottenLoops.insert(L).second)
148 // Inherit predecessor's name if it exists...
149 if (!OldName.empty() && !OnlyPred->hasName())
150 OnlyPred->setName(OldName);
152 BB->eraseFromParent();
157 /// Check if unrolling created a situation where we need to insert phi nodes to
158 /// preserve LCSSA form.
159 /// \param Blocks is a vector of basic blocks representing unrolled loop.
160 /// \param L is the outer loop.
161 /// It's possible that some of the blocks are in L, and some are not. In this
162 /// case, if there is a use is outside L, and definition is inside L, we need to
163 /// insert a phi-node, otherwise LCSSA will be broken.
164 /// The function is just a helper function for llvm::UnrollLoop that returns
165 /// true if this situation occurs, indicating that LCSSA needs to be fixed.
166 static bool needToInsertPhisForLCSSA(Loop *L, std::vector<BasicBlock *> Blocks,
168 for (BasicBlock *BB : Blocks) {
169 if (LI->getLoopFor(BB) == L)
171 for (Instruction &I : *BB) {
172 for (Use &U : I.operands()) {
173 if (auto Def = dyn_cast<Instruction>(U)) {
174 Loop *DefLoop = LI->getLoopFor(Def->getParent());
177 if (DefLoop->contains(L))
186 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
187 /// and adds a mapping from the original loop to the new loop to NewLoops.
188 /// Returns nullptr if no new loop was created and a pointer to the
189 /// original loop OriginalBB was part of otherwise.
190 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
191 BasicBlock *ClonedBB, LoopInfo *LI,
192 NewLoopsMap &NewLoops) {
193 // Figure out which loop New is in.
194 const Loop *OldLoop = LI->getLoopFor(OriginalBB);
195 assert(OldLoop && "Should (at least) be in the loop being unrolled!");
197 Loop *&NewLoop = NewLoops[OldLoop];
199 // Found a new sub-loop.
200 assert(OriginalBB == OldLoop->getHeader() &&
201 "Header should be first in RPO");
203 NewLoop = new Loop();
204 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
207 NewLoopParent->addChildLoop(NewLoop);
209 LI->addTopLevelLoop(NewLoop);
211 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
214 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
219 /// The function chooses which type of unroll (epilog or prolog) is more
221 /// Epilog unroll is more profitable when there is PHI that starts from
222 /// constant. In this case epilog will leave PHI start from constant,
223 /// but prolog will convert it to non-constant.
226 /// PN = PHI [I, Latch], [CI, PreHeader]
230 /// Epilog unroll case.
232 /// PN = PHI [I2, Latch], [CI, PreHeader]
236 /// Prolog unroll case.
237 /// NewPN = PHI [PrologI, Prolog], [CI, PreHeader]
239 /// PN = PHI [I2, Latch], [NewPN, PreHeader]
244 static bool isEpilogProfitable(Loop *L) {
245 BasicBlock *PreHeader = L->getLoopPreheader();
246 BasicBlock *Header = L->getHeader();
247 assert(PreHeader && Header);
248 for (Instruction &BBI : *Header) {
249 PHINode *PN = dyn_cast<PHINode>(&BBI);
252 if (isa<ConstantInt>(PN->getIncomingValueForBlock(PreHeader)))
258 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
259 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
260 /// can only fail when the loop's latch block is not terminated by a conditional
261 /// branch instruction. However, if the trip count (and multiple) are not known,
262 /// loop unrolling will mostly produce more code that is no faster.
264 /// TripCount is the upper bound of the iteration on which control exits
265 /// LatchBlock. Control may exit the loop prior to TripCount iterations either
266 /// via an early branch in other loop block or via LatchBlock terminator. This
267 /// is relaxed from the general definition of trip count which is the number of
268 /// times the loop header executes. Note that UnrollLoop assumes that the loop
269 /// counter test is in LatchBlock in order to remove unnecesssary instances of
270 /// the test. If control can exit the loop from the LatchBlock's terminator
271 /// prior to TripCount iterations, flag PreserveCondBr needs to be set.
273 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock
274 /// needs to be preserved. It is needed when we use trip count upper bound to
275 /// fully unroll the loop. If PreserveOnlyFirst is also set then only the first
276 /// conditional branch needs to be preserved.
278 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
279 /// execute without exiting the loop.
281 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
282 /// have a runtime (i.e. not compile time constant) trip count. Unrolling these
283 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
284 /// iterations before branching into the unrolled loop. UnrollLoop will not
285 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
286 /// AllowExpensiveTripCount is false.
288 /// If we want to perform PGO-based loop peeling, PeelCount is set to the
289 /// number of iterations we want to peel off.
291 /// The LoopInfo Analysis that is passed will be kept consistent.
293 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
294 /// DominatorTree if they are non-null.
295 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool Force,
296 bool AllowRuntime, bool AllowExpensiveTripCount,
297 bool PreserveCondBr, bool PreserveOnlyFirst,
298 unsigned TripMultiple, unsigned PeelCount, LoopInfo *LI,
299 ScalarEvolution *SE, DominatorTree *DT,
300 AssumptionCache *AC, OptimizationRemarkEmitter *ORE,
301 bool PreserveLCSSA) {
303 BasicBlock *Preheader = L->getLoopPreheader();
305 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
309 BasicBlock *LatchBlock = L->getLoopLatch();
311 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
315 // Loops with indirectbr cannot be cloned.
316 if (!L->isSafeToClone()) {
317 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
321 // The current loop unroll pass can only unroll loops with a single latch
322 // that's a conditional branch exiting the loop.
323 // FIXME: The implementation can be extended to work with more complicated
324 // cases, e.g. loops with multiple latches.
325 BasicBlock *Header = L->getHeader();
326 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
328 if (!BI || BI->isUnconditional()) {
329 // The loop-rotate pass can be helpful to avoid this in many cases.
331 " Can't unroll; loop not terminated by a conditional branch.\n");
335 auto CheckSuccessors = [&](unsigned S1, unsigned S2) {
336 return BI->getSuccessor(S1) == Header && !L->contains(BI->getSuccessor(S2));
339 if (!CheckSuccessors(0, 1) && !CheckSuccessors(1, 0)) {
340 DEBUG(dbgs() << "Can't unroll; only loops with one conditional latch"
341 " exiting the loop can be unrolled\n");
345 if (Header->hasAddressTaken()) {
346 // The loop-rotate pass can be helpful to avoid this in many cases.
348 " Won't unroll loop: address of header block is taken.\n");
353 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
354 if (TripMultiple != 1)
355 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
357 // Effectively "DCE" unrolled iterations that are beyond the tripcount
358 // and will never be executed.
359 if (TripCount != 0 && Count > TripCount)
362 // Don't enter the unroll code if there is nothing to do.
363 if (TripCount == 0 && Count < 2 && PeelCount == 0) {
364 DEBUG(dbgs() << "Won't unroll; almost nothing to do\n");
369 assert(TripMultiple > 0);
370 assert(TripCount == 0 || TripCount % TripMultiple == 0);
372 // Are we eliminating the loop control altogether?
373 bool CompletelyUnroll = Count == TripCount;
374 SmallVector<BasicBlock *, 4> ExitBlocks;
375 L->getExitBlocks(ExitBlocks);
376 std::vector<BasicBlock*> OriginalLoopBlocks = L->getBlocks();
378 // Go through all exits of L and see if there are any phi-nodes there. We just
379 // conservatively assume that they're inserted to preserve LCSSA form, which
380 // means that complete unrolling might break this form. We need to either fix
381 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
382 // now we just recompute LCSSA for the outer loop, but it should be possible
383 // to fix it in-place.
384 bool NeedToFixLCSSA = PreserveLCSSA && CompletelyUnroll &&
385 any_of(ExitBlocks, [](const BasicBlock *BB) {
386 return isa<PHINode>(BB->begin());
389 // We assume a run-time trip count if the compiler cannot
390 // figure out the loop trip count and the unroll-runtime
391 // flag is specified.
392 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
394 assert((!RuntimeTripCount || !PeelCount) &&
395 "Did not expect runtime trip-count unrolling "
396 "and peeling for the same loop");
399 peelLoop(L, PeelCount, LI, SE, DT, AC, PreserveLCSSA);
401 // Loops containing convergent instructions must have a count that divides
402 // their TripMultiple.
405 bool HasConvergent = false;
406 for (auto &BB : L->blocks())
408 if (auto CS = CallSite(&I))
409 HasConvergent |= CS.isConvergent();
410 assert((!HasConvergent || TripMultiple % Count == 0) &&
411 "Unroll count must divide trip multiple if loop contains a "
412 "convergent operation.");
415 bool EpilogProfitability =
416 UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog
417 : isEpilogProfitable(L);
419 if (RuntimeTripCount && TripMultiple % Count != 0 &&
420 !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount,
421 EpilogProfitability, LI, SE, DT,
424 RuntimeTripCount = false;
427 dbgs() << "Wont unroll; remainder loop could not be generated"
428 "when assuming runtime trip count\n");
433 // Notify ScalarEvolution that the loop will be substantially changed,
434 // if not outright eliminated.
438 // If we know the trip count, we know the multiple...
439 unsigned BreakoutTrip = 0;
440 if (TripCount != 0) {
441 BreakoutTrip = TripCount % Count;
444 // Figure out what multiple to use.
445 BreakoutTrip = TripMultiple =
446 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
450 // Report the unrolling decision.
451 if (CompletelyUnroll) {
452 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
453 << " with trip count " << TripCount << "!\n");
454 ORE->emit(OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
456 << "completely unrolled loop with "
457 << NV("UnrollCount", TripCount) << " iterations");
458 } else if (PeelCount) {
459 DEBUG(dbgs() << "PEELING loop %" << Header->getName()
460 << " with iteration count " << PeelCount << "!\n");
461 ORE->emit(OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
463 << " peeled loop by " << NV("PeelCount", PeelCount)
466 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
468 Diag << "unrolled loop by a factor of " << NV("UnrollCount", Count);
470 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
472 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
473 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
474 ORE->emit(Diag << " with a breakout at trip "
475 << NV("BreakoutTrip", BreakoutTrip));
476 } else if (TripMultiple != 1) {
477 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
478 ORE->emit(Diag << " with " << NV("TripMultiple", TripMultiple)
479 << " trips per branch");
480 } else if (RuntimeTripCount) {
481 DEBUG(dbgs() << " with run-time trip count");
482 ORE->emit(Diag << " with run-time trip count");
484 DEBUG(dbgs() << "!\n");
487 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
488 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
490 // For the first iteration of the loop, we should use the precloned values for
491 // PHI nodes. Insert associations now.
492 ValueToValueMapTy LastValueMap;
493 std::vector<PHINode*> OrigPHINode;
494 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
495 OrigPHINode.push_back(cast<PHINode>(I));
498 std::vector<BasicBlock*> Headers;
499 std::vector<BasicBlock*> Latches;
500 Headers.push_back(Header);
501 Latches.push_back(LatchBlock);
503 // The current on-the-fly SSA update requires blocks to be processed in
504 // reverse postorder so that LastValueMap contains the correct value at each
506 LoopBlocksDFS DFS(L);
509 // Stash the DFS iterators before adding blocks to the loop.
510 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
511 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
513 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
515 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we
516 // might break loop-simplified form for these loops (as they, e.g., would
517 // share the same exit blocks). We'll keep track of loops for which we can
518 // break this so that later we can re-simplify them.
519 SmallSetVector<Loop *, 4> LoopsToSimplify;
520 for (Loop *SubLoop : *L)
521 LoopsToSimplify.insert(SubLoop);
523 if (Header->getParent()->isDebugInfoForProfiling())
524 for (BasicBlock *BB : L->getBlocks())
525 for (Instruction &I : *BB)
526 if (const DILocation *DIL = I.getDebugLoc())
527 I.setDebugLoc(DIL->cloneWithDuplicationFactor(Count));
529 for (unsigned It = 1; It != Count; ++It) {
530 std::vector<BasicBlock*> NewBlocks;
531 SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
534 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
535 ValueToValueMapTy VMap;
536 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
537 Header->getParent()->getBasicBlockList().push_back(New);
539 assert((*BB != Header || LI->getLoopFor(*BB) == L) &&
540 "Header should not be in a sub-loop");
541 // Tell LI about New.
542 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
544 LoopsToSimplify.insert(NewLoops[OldLoop]);
546 // Forget the old loop, since its inputs may have changed.
548 SE->forgetLoop(OldLoop);
552 // Loop over all of the PHI nodes in the block, changing them to use
553 // the incoming values from the previous block.
554 for (PHINode *OrigPHI : OrigPHINode) {
555 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
556 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
557 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
558 if (It > 1 && L->contains(InValI))
559 InVal = LastValueMap[InValI];
560 VMap[OrigPHI] = InVal;
561 New->getInstList().erase(NewPHI);
564 // Update our running map of newest clones
565 LastValueMap[*BB] = New;
566 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
568 LastValueMap[VI->first] = VI->second;
570 // Add phi entries for newly created values to all exit blocks.
571 for (BasicBlock *Succ : successors(*BB)) {
572 if (L->contains(Succ))
574 for (BasicBlock::iterator BBI = Succ->begin();
575 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
576 Value *Incoming = phi->getIncomingValueForBlock(*BB);
577 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
578 if (It != LastValueMap.end())
579 Incoming = It->second;
580 phi->addIncoming(Incoming, New);
583 // Keep track of new headers and latches as we create them, so that
584 // we can insert the proper branches later.
586 Headers.push_back(New);
587 if (*BB == LatchBlock)
588 Latches.push_back(New);
590 NewBlocks.push_back(New);
591 UnrolledLoopBlocks.push_back(New);
593 // Update DomTree: since we just copy the loop body, and each copy has a
594 // dedicated entry block (copy of the header block), this header's copy
595 // dominates all copied blocks. That means, dominance relations in the
596 // copied body are the same as in the original body.
599 DT->addNewBlock(New, Latches[It - 1]);
601 auto BBDomNode = DT->getNode(*BB);
602 auto BBIDom = BBDomNode->getIDom();
603 BasicBlock *OriginalBBIDom = BBIDom->getBlock();
605 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
610 // Remap all instructions in the most recent iteration
611 for (BasicBlock *NewBlock : NewBlocks) {
612 for (Instruction &I : *NewBlock) {
613 ::remapInstruction(&I, LastValueMap);
614 if (auto *II = dyn_cast<IntrinsicInst>(&I))
615 if (II->getIntrinsicID() == Intrinsic::assume)
616 AC->registerAssumption(II);
621 // Loop over the PHI nodes in the original block, setting incoming values.
622 for (PHINode *PN : OrigPHINode) {
623 if (CompletelyUnroll) {
624 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
625 Header->getInstList().erase(PN);
627 else if (Count > 1) {
628 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
629 // If this value was defined in the loop, take the value defined by the
630 // last iteration of the loop.
631 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
632 if (L->contains(InValI))
633 InVal = LastValueMap[InVal];
635 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
636 PN->addIncoming(InVal, Latches.back());
640 // Now that all the basic blocks for the unrolled iterations are in place,
641 // set up the branches to connect them.
642 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
643 // The original branch was replicated in each unrolled iteration.
644 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
646 // The branch destination.
647 unsigned j = (i + 1) % e;
648 BasicBlock *Dest = Headers[j];
649 bool NeedConditional = true;
651 if (RuntimeTripCount && j != 0) {
652 NeedConditional = false;
655 // For a complete unroll, make the last iteration end with a branch
656 // to the exit block.
657 if (CompletelyUnroll) {
660 // If using trip count upper bound to completely unroll, we need to keep
661 // the conditional branch except the last one because the loop may exit
662 // after any iteration.
663 assert(NeedConditional &&
664 "NeedCondition cannot be modified by both complete "
665 "unrolling and runtime unrolling");
666 NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0));
667 } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
668 // If we know the trip count or a multiple of it, we can safely use an
669 // unconditional branch for some iterations.
670 NeedConditional = false;
673 if (NeedConditional) {
674 // Update the conditional branch's successor for the following
676 Term->setSuccessor(!ContinueOnTrue, Dest);
678 // Remove phi operands at this loop exit
679 if (Dest != LoopExit) {
680 BasicBlock *BB = Latches[i];
681 for (BasicBlock *Succ: successors(BB)) {
682 if (Succ == Headers[i])
684 for (BasicBlock::iterator BBI = Succ->begin();
685 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
686 Phi->removeIncomingValue(BB, false);
690 // Replace the conditional branch with an unconditional one.
691 BranchInst::Create(Dest, Term);
692 Term->eraseFromParent();
696 // Update dominators of blocks we might reach through exits.
697 // Immediate dominator of such block might change, because we add more
698 // routes which can lead to the exit: we can now reach it from the copied
700 if (DT && Count > 1) {
701 for (auto *BB : OriginalLoopBlocks) {
702 auto *BBDomNode = DT->getNode(BB);
703 SmallVector<BasicBlock *, 16> ChildrenToUpdate;
704 for (auto *ChildDomNode : BBDomNode->getChildren()) {
705 auto *ChildBB = ChildDomNode->getBlock();
706 if (!L->contains(ChildBB))
707 ChildrenToUpdate.push_back(ChildBB);
710 if (BB == LatchBlock) {
711 // The latch is special because we emit unconditional branches in
712 // some cases where the original loop contained a conditional branch.
713 // Since the latch is always at the bottom of the loop, if the latch
714 // dominated an exit before unrolling, the new dominator of that exit
715 // must also be a latch. Specifically, the dominator is the first
716 // latch which ends in a conditional branch, or the last latch if
717 // there is no such latch.
718 NewIDom = Latches.back();
719 for (BasicBlock *IterLatch : Latches) {
720 TerminatorInst *Term = IterLatch->getTerminator();
721 if (isa<BranchInst>(Term) && cast<BranchInst>(Term)->isConditional()) {
727 // The new idom of the block will be the nearest common dominator
728 // of all copies of the previous idom. This is equivalent to the
729 // nearest common dominator of the previous idom and the first latch,
730 // which dominates all copies of the previous idom.
731 NewIDom = DT->findNearestCommonDominator(BB, LatchBlock);
733 for (auto *ChildBB : ChildrenToUpdate)
734 DT->changeImmediateDominator(ChildBB, NewIDom);
738 if (DT && UnrollVerifyDomtree)
741 // Merge adjacent basic blocks, if possible.
742 SmallPtrSet<Loop *, 4> ForgottenLoops;
743 for (BasicBlock *Latch : Latches) {
744 BranchInst *Term = cast<BranchInst>(Latch->getTerminator());
745 if (Term->isUnconditional()) {
746 BasicBlock *Dest = Term->getSuccessor(0);
747 if (BasicBlock *Fold =
748 foldBlockIntoPredecessor(Dest, LI, SE, ForgottenLoops, DT)) {
749 // Dest has been folded into Fold. Update our worklists accordingly.
750 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
751 UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(),
752 UnrolledLoopBlocks.end(), Dest),
753 UnrolledLoopBlocks.end());
758 // Simplify any new induction variables in the partially unrolled loop.
759 if (SE && !CompletelyUnroll && Count > 1) {
760 SmallVector<WeakTrackingVH, 16> DeadInsts;
761 simplifyLoopIVs(L, SE, DT, LI, DeadInsts);
763 // Aggressively clean up dead instructions that simplifyLoopIVs already
764 // identified. Any remaining should be cleaned up below.
765 while (!DeadInsts.empty())
766 if (Instruction *Inst =
767 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
768 RecursivelyDeleteTriviallyDeadInstructions(Inst);
771 // At this point, the code is well formed. We now do a quick sweep over the
772 // inserted code, doing constant propagation and dead code elimination as we
774 const DataLayout &DL = Header->getModule()->getDataLayout();
775 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
776 for (BasicBlock *BB : NewLoopBlocks) {
777 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
778 Instruction *Inst = &*I++;
780 if (Value *V = SimplifyInstruction(Inst, {DL, nullptr, DT, AC}))
781 if (LI->replacementPreservesLCSSAForm(Inst, V))
782 Inst->replaceAllUsesWith(V);
783 if (isInstructionTriviallyDead(Inst))
784 BB->getInstList().erase(Inst);
788 // TODO: after peeling or unrolling, previously loop variant conditions are
789 // likely to fold to constants, eagerly propagating those here will require
790 // fewer cleanup passes to be run. Alternatively, a LoopEarlyCSE might be
793 NumCompletelyUnrolled += CompletelyUnroll;
796 Loop *OuterL = L->getParentLoop();
797 // Update LoopInfo if the loop is completely removed.
798 if (CompletelyUnroll)
799 LI->markAsRemoved(L);
801 // After complete unrolling most of the blocks should be contained in OuterL.
802 // However, some of them might happen to be out of OuterL (e.g. if they
803 // precede a loop exit). In this case we might need to insert PHI nodes in
804 // order to preserve LCSSA form.
805 // We don't need to check this if we already know that we need to fix LCSSA
807 // TODO: For now we just recompute LCSSA for the outer loop in this case, but
808 // it should be possible to fix it in-place.
809 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
810 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
812 // If we have a pass and a DominatorTree we should re-simplify impacted loops
813 // to ensure subsequent analyses can rely on this form. We want to simplify
814 // at least one layer outside of the loop that was unrolled so that any
815 // changes to the parent loop exposed by the unrolling are considered.
818 // OuterL includes all loops for which we can break loop-simplify, so
819 // it's sufficient to simplify only it (it'll recursively simplify inner
821 if (NeedToFixLCSSA) {
822 // LCSSA must be performed on the outermost affected loop. The unrolled
823 // loop's last loop latch is guaranteed to be in the outermost loop
824 // after LoopInfo's been updated by markAsRemoved.
825 Loop *LatchLoop = LI->getLoopFor(Latches.back());
826 Loop *FixLCSSALoop = OuterL;
827 if (!FixLCSSALoop->contains(LatchLoop))
828 while (FixLCSSALoop->getParentLoop() != LatchLoop)
829 FixLCSSALoop = FixLCSSALoop->getParentLoop();
831 formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE);
832 } else if (PreserveLCSSA) {
833 assert(OuterL->isLCSSAForm(*DT) &&
834 "Loops should be in LCSSA form after loop-unroll.");
837 // TODO: That potentially might be compile-time expensive. We should try
838 // to fix the loop-simplified form incrementally.
839 simplifyLoop(OuterL, DT, LI, SE, AC, PreserveLCSSA);
841 // Simplify loops for which we might've broken loop-simplify form.
842 for (Loop *SubLoop : LoopsToSimplify)
843 simplifyLoop(SubLoop, DT, LI, SE, AC, PreserveLCSSA);
850 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata
851 /// node with the given name (for example, "llvm.loop.unroll.count"). If no
852 /// such metadata node exists, then nullptr is returned.
853 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
854 // First operand should refer to the loop id itself.
855 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
856 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
858 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
859 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
863 MDString *S = dyn_cast<MDString>(MD->getOperand(0));
867 if (Name.equals(S->getString()))