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/OptimizationRemarkEmitter.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/IR/BasicBlock.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DebugInfoMetadata.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/Cloning.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Utils/LoopSimplify.h"
38 #include "llvm/Transforms/Utils/LoopUtils.h"
39 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
40 #include "llvm/Transforms/Utils/UnrollLoop.h"
43 #define DEBUG_TYPE "loop-unroll"
45 // TODO: Should these be here or in LoopUnroll?
46 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
47 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
50 UnrollRuntimeEpilog("unroll-runtime-epilog", cl::init(false), cl::Hidden,
51 cl::desc("Allow runtime unrolled loops to be unrolled "
52 "with epilog instead of prolog."));
55 UnrollVerifyDomtree("unroll-verify-domtree", cl::Hidden,
56 cl::desc("Verify domtree after unrolling"),
64 /// Convert the instruction operands from referencing the current values into
65 /// those specified by VMap.
66 static inline void remapInstruction(Instruction *I,
67 ValueToValueMapTy &VMap) {
68 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
69 Value *Op = I->getOperand(op);
71 // Unwrap arguments of dbg.value intrinsics.
73 if (auto *V = dyn_cast<MetadataAsValue>(Op))
74 if (auto *Unwrapped = dyn_cast<ValueAsMetadata>(V->getMetadata())) {
75 Op = Unwrapped->getValue();
79 auto wrap = [&](Value *V) {
80 auto &C = I->getContext();
81 return Wrapped ? MetadataAsValue::get(C, ValueAsMetadata::get(V)) : V;
84 ValueToValueMapTy::iterator It = VMap.find(Op);
86 I->setOperand(op, wrap(It->second));
89 if (PHINode *PN = dyn_cast<PHINode>(I)) {
90 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
91 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
93 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
98 /// Folds a basic block into its predecessor if it only has one predecessor, and
99 /// that predecessor only has one successor.
100 /// The LoopInfo Analysis that is passed will be kept consistent. If folding is
101 /// successful references to the containing loop must be removed from
102 /// ScalarEvolution by calling ScalarEvolution::forgetLoop because SE may have
103 /// references to the eliminated BB. The argument ForgottenLoops contains a set
104 /// of loops that have already been forgotten to prevent redundant, expensive
105 /// calls to ScalarEvolution::forgetLoop. Returns the new combined block.
107 foldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, ScalarEvolution *SE,
108 SmallPtrSetImpl<Loop *> &ForgottenLoops,
110 // Merge basic blocks into their predecessor if there is only one distinct
111 // pred, and if there is only one distinct successor of the predecessor, and
112 // if there are no PHI nodes.
113 BasicBlock *OnlyPred = BB->getSinglePredecessor();
114 if (!OnlyPred) return nullptr;
116 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
119 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
121 // Resolve any PHI nodes at the start of the block. They are all
122 // guaranteed to have exactly one entry if they exist, unless there are
123 // multiple duplicate (but guaranteed to be equal) entries for the
124 // incoming edges. This occurs when there are multiple edges from
125 // OnlyPred to OnlySucc.
126 FoldSingleEntryPHINodes(BB);
128 // Delete the unconditional branch from the predecessor...
129 OnlyPred->getInstList().pop_back();
131 // Make all PHI nodes that referred to BB now refer to Pred as their
133 BB->replaceAllUsesWith(OnlyPred);
135 // Move all definitions in the successor to the predecessor...
136 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
138 // OldName will be valid until erased.
139 StringRef OldName = BB->getName();
141 // Erase the old block and update dominator info.
143 if (DomTreeNode *DTN = DT->getNode(BB)) {
144 DomTreeNode *PredDTN = DT->getNode(OnlyPred);
145 SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end());
146 for (auto *DI : Children)
147 DT->changeImmediateDominator(DI, PredDTN);
152 // ScalarEvolution holds references to loop exit blocks.
154 if (Loop *L = LI->getLoopFor(BB)) {
155 if (ForgottenLoops.insert(L).second)
161 // Inherit predecessor's name if it exists...
162 if (!OldName.empty() && !OnlyPred->hasName())
163 OnlyPred->setName(OldName);
165 BB->eraseFromParent();
170 /// Check if unrolling created a situation where we need to insert phi nodes to
171 /// preserve LCSSA form.
172 /// \param Blocks is a vector of basic blocks representing unrolled loop.
173 /// \param L is the outer loop.
174 /// It's possible that some of the blocks are in L, and some are not. In this
175 /// case, if there is a use is outside L, and definition is inside L, we need to
176 /// insert a phi-node, otherwise LCSSA will be broken.
177 /// The function is just a helper function for llvm::UnrollLoop that returns
178 /// true if this situation occurs, indicating that LCSSA needs to be fixed.
179 static bool needToInsertPhisForLCSSA(Loop *L, std::vector<BasicBlock *> Blocks,
181 for (BasicBlock *BB : Blocks) {
182 if (LI->getLoopFor(BB) == L)
184 for (Instruction &I : *BB) {
185 for (Use &U : I.operands()) {
186 if (auto Def = dyn_cast<Instruction>(U)) {
187 Loop *DefLoop = LI->getLoopFor(Def->getParent());
190 if (DefLoop->contains(L))
199 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
200 /// and adds a mapping from the original loop to the new loop to NewLoops.
201 /// Returns nullptr if no new loop was created and a pointer to the
202 /// original loop OriginalBB was part of otherwise.
203 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
204 BasicBlock *ClonedBB, LoopInfo *LI,
205 NewLoopsMap &NewLoops) {
206 // Figure out which loop New is in.
207 const Loop *OldLoop = LI->getLoopFor(OriginalBB);
208 assert(OldLoop && "Should (at least) be in the loop being unrolled!");
210 Loop *&NewLoop = NewLoops[OldLoop];
212 // Found a new sub-loop.
213 assert(OriginalBB == OldLoop->getHeader() &&
214 "Header should be first in RPO");
216 NewLoop = LI->AllocateLoop();
217 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
220 NewLoopParent->addChildLoop(NewLoop);
222 LI->addTopLevelLoop(NewLoop);
224 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
227 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
232 /// The function chooses which type of unroll (epilog or prolog) is more
234 /// Epilog unroll is more profitable when there is PHI that starts from
235 /// constant. In this case epilog will leave PHI start from constant,
236 /// but prolog will convert it to non-constant.
239 /// PN = PHI [I, Latch], [CI, PreHeader]
243 /// Epilog unroll case.
245 /// PN = PHI [I2, Latch], [CI, PreHeader]
249 /// Prolog unroll case.
250 /// NewPN = PHI [PrologI, Prolog], [CI, PreHeader]
252 /// PN = PHI [I2, Latch], [NewPN, PreHeader]
257 static bool isEpilogProfitable(Loop *L) {
258 BasicBlock *PreHeader = L->getLoopPreheader();
259 BasicBlock *Header = L->getHeader();
260 assert(PreHeader && Header);
261 for (Instruction &BBI : *Header) {
262 PHINode *PN = dyn_cast<PHINode>(&BBI);
265 if (isa<ConstantInt>(PN->getIncomingValueForBlock(PreHeader)))
271 /// Unroll the given loop by Count. The loop must be in LCSSA form. Unrolling
272 /// can only fail when the loop's latch block is not terminated by a conditional
273 /// branch instruction. However, if the trip count (and multiple) are not known,
274 /// loop unrolling will mostly produce more code that is no faster.
276 /// TripCount is the upper bound of the iteration on which control exits
277 /// LatchBlock. Control may exit the loop prior to TripCount iterations either
278 /// via an early branch in other loop block or via LatchBlock terminator. This
279 /// is relaxed from the general definition of trip count which is the number of
280 /// times the loop header executes. Note that UnrollLoop assumes that the loop
281 /// counter test is in LatchBlock in order to remove unnecesssary instances of
282 /// the test. If control can exit the loop from the LatchBlock's terminator
283 /// prior to TripCount iterations, flag PreserveCondBr needs to be set.
285 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock
286 /// needs to be preserved. It is needed when we use trip count upper bound to
287 /// fully unroll the loop. If PreserveOnlyFirst is also set then only the first
288 /// conditional branch needs to be preserved.
290 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
291 /// execute without exiting the loop.
293 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
294 /// have a runtime (i.e. not compile time constant) trip count. Unrolling these
295 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
296 /// iterations before branching into the unrolled loop. UnrollLoop will not
297 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
298 /// AllowExpensiveTripCount is false.
300 /// If we want to perform PGO-based loop peeling, PeelCount is set to the
301 /// number of iterations we want to peel off.
303 /// The LoopInfo Analysis that is passed will be kept consistent.
305 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
306 /// DominatorTree if they are non-null.
307 LoopUnrollResult llvm::UnrollLoop(
308 Loop *L, unsigned Count, unsigned TripCount, bool Force, bool AllowRuntime,
309 bool AllowExpensiveTripCount, bool PreserveCondBr, bool PreserveOnlyFirst,
310 unsigned TripMultiple, unsigned PeelCount, bool UnrollRemainder,
311 LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
312 OptimizationRemarkEmitter *ORE, bool PreserveLCSSA) {
314 BasicBlock *Preheader = L->getLoopPreheader();
316 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
317 return LoopUnrollResult::Unmodified;
320 BasicBlock *LatchBlock = L->getLoopLatch();
322 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
323 return LoopUnrollResult::Unmodified;
326 // Loops with indirectbr cannot be cloned.
327 if (!L->isSafeToClone()) {
328 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
329 return LoopUnrollResult::Unmodified;
332 // The current loop unroll pass can only unroll loops with a single latch
333 // that's a conditional branch exiting the loop.
334 // FIXME: The implementation can be extended to work with more complicated
335 // cases, e.g. loops with multiple latches.
336 BasicBlock *Header = L->getHeader();
337 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
339 if (!BI || BI->isUnconditional()) {
340 // The loop-rotate pass can be helpful to avoid this in many cases.
342 " Can't unroll; loop not terminated by a conditional branch.\n");
343 return LoopUnrollResult::Unmodified;
346 auto CheckSuccessors = [&](unsigned S1, unsigned S2) {
347 return BI->getSuccessor(S1) == Header && !L->contains(BI->getSuccessor(S2));
350 if (!CheckSuccessors(0, 1) && !CheckSuccessors(1, 0)) {
351 DEBUG(dbgs() << "Can't unroll; only loops with one conditional latch"
352 " exiting the loop can be unrolled\n");
353 return LoopUnrollResult::Unmodified;
356 if (Header->hasAddressTaken()) {
357 // The loop-rotate pass can be helpful to avoid this in many cases.
359 " Won't unroll loop: address of header block is taken.\n");
360 return LoopUnrollResult::Unmodified;
364 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
365 if (TripMultiple != 1)
366 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
368 // Effectively "DCE" unrolled iterations that are beyond the tripcount
369 // and will never be executed.
370 if (TripCount != 0 && Count > TripCount)
373 // Don't enter the unroll code if there is nothing to do.
374 if (TripCount == 0 && Count < 2 && PeelCount == 0) {
375 DEBUG(dbgs() << "Won't unroll; almost nothing to do\n");
376 return LoopUnrollResult::Unmodified;
380 assert(TripMultiple > 0);
381 assert(TripCount == 0 || TripCount % TripMultiple == 0);
383 // Are we eliminating the loop control altogether?
384 bool CompletelyUnroll = Count == TripCount;
385 SmallVector<BasicBlock *, 4> ExitBlocks;
386 L->getExitBlocks(ExitBlocks);
387 std::vector<BasicBlock*> OriginalLoopBlocks = L->getBlocks();
389 // Go through all exits of L and see if there are any phi-nodes there. We just
390 // conservatively assume that they're inserted to preserve LCSSA form, which
391 // means that complete unrolling might break this form. We need to either fix
392 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
393 // now we just recompute LCSSA for the outer loop, but it should be possible
394 // to fix it in-place.
395 bool NeedToFixLCSSA = PreserveLCSSA && CompletelyUnroll &&
396 any_of(ExitBlocks, [](const BasicBlock *BB) {
397 return isa<PHINode>(BB->begin());
400 // We assume a run-time trip count if the compiler cannot
401 // figure out the loop trip count and the unroll-runtime
402 // flag is specified.
403 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
405 assert((!RuntimeTripCount || !PeelCount) &&
406 "Did not expect runtime trip-count unrolling "
407 "and peeling for the same loop");
410 bool Peeled = peelLoop(L, PeelCount, LI, SE, DT, AC, PreserveLCSSA);
412 // Successful peeling may result in a change in the loop preheader/trip
413 // counts. If we later unroll the loop, we want these to be updated.
415 BasicBlock *ExitingBlock = L->getExitingBlock();
416 assert(ExitingBlock && "Loop without exiting block?");
417 Preheader = L->getLoopPreheader();
418 TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
419 TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
423 // Loops containing convergent instructions must have a count that divides
424 // their TripMultiple.
427 bool HasConvergent = false;
428 for (auto &BB : L->blocks())
430 if (auto CS = CallSite(&I))
431 HasConvergent |= CS.isConvergent();
432 assert((!HasConvergent || TripMultiple % Count == 0) &&
433 "Unroll count must divide trip multiple if loop contains a "
434 "convergent operation.");
437 bool EpilogProfitability =
438 UnrollRuntimeEpilog.getNumOccurrences() ? UnrollRuntimeEpilog
439 : isEpilogProfitable(L);
441 if (RuntimeTripCount && TripMultiple % Count != 0 &&
442 !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount,
443 EpilogProfitability, UnrollRemainder, LI, SE,
444 DT, AC, PreserveLCSSA)) {
446 RuntimeTripCount = false;
449 dbgs() << "Wont unroll; remainder loop could not be generated"
450 "when assuming runtime trip count\n");
451 return LoopUnrollResult::Unmodified;
455 // Notify ScalarEvolution that the loop will be substantially changed,
456 // if not outright eliminated.
460 // If we know the trip count, we know the multiple...
461 unsigned BreakoutTrip = 0;
462 if (TripCount != 0) {
463 BreakoutTrip = TripCount % Count;
466 // Figure out what multiple to use.
467 BreakoutTrip = TripMultiple =
468 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
472 // Report the unrolling decision.
473 if (CompletelyUnroll) {
474 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
475 << " with trip count " << TripCount << "!\n");
478 return OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
480 << "completely unrolled loop with "
481 << NV("UnrollCount", TripCount) << " iterations";
483 } else if (PeelCount) {
484 DEBUG(dbgs() << "PEELING loop %" << Header->getName()
485 << " with iteration count " << PeelCount << "!\n");
488 return OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
490 << " peeled loop by " << NV("PeelCount", PeelCount)
494 auto DiagBuilder = [&]() {
495 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
497 return Diag << "unrolled loop by a factor of "
498 << NV("UnrollCount", Count);
501 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
503 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
504 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
507 return DiagBuilder() << " with a breakout at trip "
508 << NV("BreakoutTrip", BreakoutTrip);
510 } else if (TripMultiple != 1) {
511 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
514 return DiagBuilder() << " with " << NV("TripMultiple", TripMultiple)
515 << " trips per branch";
517 } else if (RuntimeTripCount) {
518 DEBUG(dbgs() << " with run-time trip count");
521 [&]() { return DiagBuilder() << " with run-time trip count"; });
523 DEBUG(dbgs() << "!\n");
526 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
527 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
529 // For the first iteration of the loop, we should use the precloned values for
530 // PHI nodes. Insert associations now.
531 ValueToValueMapTy LastValueMap;
532 std::vector<PHINode*> OrigPHINode;
533 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
534 OrigPHINode.push_back(cast<PHINode>(I));
537 std::vector<BasicBlock*> Headers;
538 std::vector<BasicBlock*> Latches;
539 Headers.push_back(Header);
540 Latches.push_back(LatchBlock);
542 // The current on-the-fly SSA update requires blocks to be processed in
543 // reverse postorder so that LastValueMap contains the correct value at each
545 LoopBlocksDFS DFS(L);
548 // Stash the DFS iterators before adding blocks to the loop.
549 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
550 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
552 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
554 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we
555 // might break loop-simplified form for these loops (as they, e.g., would
556 // share the same exit blocks). We'll keep track of loops for which we can
557 // break this so that later we can re-simplify them.
558 SmallSetVector<Loop *, 4> LoopsToSimplify;
559 for (Loop *SubLoop : *L)
560 LoopsToSimplify.insert(SubLoop);
562 if (Header->getParent()->isDebugInfoForProfiling())
563 for (BasicBlock *BB : L->getBlocks())
564 for (Instruction &I : *BB)
565 if (!isa<DbgInfoIntrinsic>(&I))
566 if (const DILocation *DIL = I.getDebugLoc())
567 I.setDebugLoc(DIL->cloneWithDuplicationFactor(Count));
569 for (unsigned It = 1; It != Count; ++It) {
570 std::vector<BasicBlock*> NewBlocks;
571 SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
574 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
575 ValueToValueMapTy VMap;
576 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
577 Header->getParent()->getBasicBlockList().push_back(New);
579 assert((*BB != Header || LI->getLoopFor(*BB) == L) &&
580 "Header should not be in a sub-loop");
581 // Tell LI about New.
582 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
584 LoopsToSimplify.insert(NewLoops[OldLoop]);
586 // Forget the old loop, since its inputs may have changed.
588 SE->forgetLoop(OldLoop);
592 // Loop over all of the PHI nodes in the block, changing them to use
593 // the incoming values from the previous block.
594 for (PHINode *OrigPHI : OrigPHINode) {
595 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
596 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
597 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
598 if (It > 1 && L->contains(InValI))
599 InVal = LastValueMap[InValI];
600 VMap[OrigPHI] = InVal;
601 New->getInstList().erase(NewPHI);
604 // Update our running map of newest clones
605 LastValueMap[*BB] = New;
606 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
608 LastValueMap[VI->first] = VI->second;
610 // Add phi entries for newly created values to all exit blocks.
611 for (BasicBlock *Succ : successors(*BB)) {
612 if (L->contains(Succ))
614 for (BasicBlock::iterator BBI = Succ->begin();
615 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
616 Value *Incoming = phi->getIncomingValueForBlock(*BB);
617 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
618 if (It != LastValueMap.end())
619 Incoming = It->second;
620 phi->addIncoming(Incoming, New);
623 // Keep track of new headers and latches as we create them, so that
624 // we can insert the proper branches later.
626 Headers.push_back(New);
627 if (*BB == LatchBlock)
628 Latches.push_back(New);
630 NewBlocks.push_back(New);
631 UnrolledLoopBlocks.push_back(New);
633 // Update DomTree: since we just copy the loop body, and each copy has a
634 // dedicated entry block (copy of the header block), this header's copy
635 // dominates all copied blocks. That means, dominance relations in the
636 // copied body are the same as in the original body.
639 DT->addNewBlock(New, Latches[It - 1]);
641 auto BBDomNode = DT->getNode(*BB);
642 auto BBIDom = BBDomNode->getIDom();
643 BasicBlock *OriginalBBIDom = BBIDom->getBlock();
645 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
650 // Remap all instructions in the most recent iteration
651 for (BasicBlock *NewBlock : NewBlocks) {
652 for (Instruction &I : *NewBlock) {
653 ::remapInstruction(&I, LastValueMap);
654 if (auto *II = dyn_cast<IntrinsicInst>(&I))
655 if (II->getIntrinsicID() == Intrinsic::assume)
656 AC->registerAssumption(II);
661 // Loop over the PHI nodes in the original block, setting incoming values.
662 for (PHINode *PN : OrigPHINode) {
663 if (CompletelyUnroll) {
664 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
665 Header->getInstList().erase(PN);
667 else if (Count > 1) {
668 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
669 // If this value was defined in the loop, take the value defined by the
670 // last iteration of the loop.
671 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
672 if (L->contains(InValI))
673 InVal = LastValueMap[InVal];
675 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
676 PN->addIncoming(InVal, Latches.back());
680 // Now that all the basic blocks for the unrolled iterations are in place,
681 // set up the branches to connect them.
682 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
683 // The original branch was replicated in each unrolled iteration.
684 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
686 // The branch destination.
687 unsigned j = (i + 1) % e;
688 BasicBlock *Dest = Headers[j];
689 bool NeedConditional = true;
691 if (RuntimeTripCount && j != 0) {
692 NeedConditional = false;
695 // For a complete unroll, make the last iteration end with a branch
696 // to the exit block.
697 if (CompletelyUnroll) {
700 // If using trip count upper bound to completely unroll, we need to keep
701 // the conditional branch except the last one because the loop may exit
702 // after any iteration.
703 assert(NeedConditional &&
704 "NeedCondition cannot be modified by both complete "
705 "unrolling and runtime unrolling");
706 NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0));
707 } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
708 // If we know the trip count or a multiple of it, we can safely use an
709 // unconditional branch for some iterations.
710 NeedConditional = false;
713 if (NeedConditional) {
714 // Update the conditional branch's successor for the following
716 Term->setSuccessor(!ContinueOnTrue, Dest);
718 // Remove phi operands at this loop exit
719 if (Dest != LoopExit) {
720 BasicBlock *BB = Latches[i];
721 for (BasicBlock *Succ: successors(BB)) {
722 if (Succ == Headers[i])
724 for (BasicBlock::iterator BBI = Succ->begin();
725 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
726 Phi->removeIncomingValue(BB, false);
730 // Replace the conditional branch with an unconditional one.
731 BranchInst::Create(Dest, Term);
732 Term->eraseFromParent();
736 // Update dominators of blocks we might reach through exits.
737 // Immediate dominator of such block might change, because we add more
738 // routes which can lead to the exit: we can now reach it from the copied
740 if (DT && Count > 1) {
741 for (auto *BB : OriginalLoopBlocks) {
742 auto *BBDomNode = DT->getNode(BB);
743 SmallVector<BasicBlock *, 16> ChildrenToUpdate;
744 for (auto *ChildDomNode : BBDomNode->getChildren()) {
745 auto *ChildBB = ChildDomNode->getBlock();
746 if (!L->contains(ChildBB))
747 ChildrenToUpdate.push_back(ChildBB);
750 if (BB == LatchBlock) {
751 // The latch is special because we emit unconditional branches in
752 // some cases where the original loop contained a conditional branch.
753 // Since the latch is always at the bottom of the loop, if the latch
754 // dominated an exit before unrolling, the new dominator of that exit
755 // must also be a latch. Specifically, the dominator is the first
756 // latch which ends in a conditional branch, or the last latch if
757 // there is no such latch.
758 NewIDom = Latches.back();
759 for (BasicBlock *IterLatch : Latches) {
760 TerminatorInst *Term = IterLatch->getTerminator();
761 if (isa<BranchInst>(Term) && cast<BranchInst>(Term)->isConditional()) {
767 // The new idom of the block will be the nearest common dominator
768 // of all copies of the previous idom. This is equivalent to the
769 // nearest common dominator of the previous idom and the first latch,
770 // which dominates all copies of the previous idom.
771 NewIDom = DT->findNearestCommonDominator(BB, LatchBlock);
773 for (auto *ChildBB : ChildrenToUpdate)
774 DT->changeImmediateDominator(ChildBB, NewIDom);
778 if (DT && UnrollVerifyDomtree)
781 // Merge adjacent basic blocks, if possible.
782 SmallPtrSet<Loop *, 4> ForgottenLoops;
783 for (BasicBlock *Latch : Latches) {
784 BranchInst *Term = cast<BranchInst>(Latch->getTerminator());
785 if (Term->isUnconditional()) {
786 BasicBlock *Dest = Term->getSuccessor(0);
787 if (BasicBlock *Fold =
788 foldBlockIntoPredecessor(Dest, LI, SE, ForgottenLoops, DT)) {
789 // Dest has been folded into Fold. Update our worklists accordingly.
790 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
791 UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(),
792 UnrolledLoopBlocks.end(), Dest),
793 UnrolledLoopBlocks.end());
798 // Simplify any new induction variables in the partially unrolled loop.
799 if (SE && !CompletelyUnroll && Count > 1) {
800 SmallVector<WeakTrackingVH, 16> DeadInsts;
801 simplifyLoopIVs(L, SE, DT, LI, DeadInsts);
803 // Aggressively clean up dead instructions that simplifyLoopIVs already
804 // identified. Any remaining should be cleaned up below.
805 while (!DeadInsts.empty())
806 if (Instruction *Inst =
807 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
808 RecursivelyDeleteTriviallyDeadInstructions(Inst);
811 // At this point, the code is well formed. We now do a quick sweep over the
812 // inserted code, doing constant propagation and dead code elimination as we
814 const DataLayout &DL = Header->getModule()->getDataLayout();
815 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
816 for (BasicBlock *BB : NewLoopBlocks) {
817 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
818 Instruction *Inst = &*I++;
820 if (Value *V = SimplifyInstruction(Inst, {DL, nullptr, DT, AC}))
821 if (LI->replacementPreservesLCSSAForm(Inst, V))
822 Inst->replaceAllUsesWith(V);
823 if (isInstructionTriviallyDead(Inst))
824 BB->getInstList().erase(Inst);
828 // TODO: after peeling or unrolling, previously loop variant conditions are
829 // likely to fold to constants, eagerly propagating those here will require
830 // fewer cleanup passes to be run. Alternatively, a LoopEarlyCSE might be
833 NumCompletelyUnrolled += CompletelyUnroll;
836 Loop *OuterL = L->getParentLoop();
837 // Update LoopInfo if the loop is completely removed.
838 if (CompletelyUnroll)
841 // After complete unrolling most of the blocks should be contained in OuterL.
842 // However, some of them might happen to be out of OuterL (e.g. if they
843 // precede a loop exit). In this case we might need to insert PHI nodes in
844 // order to preserve LCSSA form.
845 // We don't need to check this if we already know that we need to fix LCSSA
847 // TODO: For now we just recompute LCSSA for the outer loop in this case, but
848 // it should be possible to fix it in-place.
849 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
850 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
852 // If we have a pass and a DominatorTree we should re-simplify impacted loops
853 // to ensure subsequent analyses can rely on this form. We want to simplify
854 // at least one layer outside of the loop that was unrolled so that any
855 // changes to the parent loop exposed by the unrolling are considered.
858 // OuterL includes all loops for which we can break loop-simplify, so
859 // it's sufficient to simplify only it (it'll recursively simplify inner
861 if (NeedToFixLCSSA) {
862 // LCSSA must be performed on the outermost affected loop. The unrolled
863 // loop's last loop latch is guaranteed to be in the outermost loop
864 // after LoopInfo's been updated by LoopInfo::erase.
865 Loop *LatchLoop = LI->getLoopFor(Latches.back());
866 Loop *FixLCSSALoop = OuterL;
867 if (!FixLCSSALoop->contains(LatchLoop))
868 while (FixLCSSALoop->getParentLoop() != LatchLoop)
869 FixLCSSALoop = FixLCSSALoop->getParentLoop();
871 formLCSSARecursively(*FixLCSSALoop, *DT, LI, SE);
872 } else if (PreserveLCSSA) {
873 assert(OuterL->isLCSSAForm(*DT) &&
874 "Loops should be in LCSSA form after loop-unroll.");
877 // TODO: That potentially might be compile-time expensive. We should try
878 // to fix the loop-simplified form incrementally.
879 simplifyLoop(OuterL, DT, LI, SE, AC, PreserveLCSSA);
881 // Simplify loops for which we might've broken loop-simplify form.
882 for (Loop *SubLoop : LoopsToSimplify)
883 simplifyLoop(SubLoop, DT, LI, SE, AC, PreserveLCSSA);
887 return CompletelyUnroll ? LoopUnrollResult::FullyUnrolled
888 : LoopUnrollResult::PartiallyUnrolled;
891 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata
892 /// node with the given name (for example, "llvm.loop.unroll.count"). If no
893 /// such metadata node exists, then nullptr is returned.
894 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
895 // First operand should refer to the loop id itself.
896 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
897 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
899 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
900 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
904 MDString *S = dyn_cast<MDString>(MD->getOperand(0));
908 if (Name.equals(S->getString()))