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/Transforms/Utils/UnrollLoop.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LoopIterator.h"
25 #include "llvm/Analysis/LoopPass.h"
26 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/IR/BasicBlock.h"
29 #include "llvm/IR/DataLayout.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"
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."));
54 /// Convert the instruction operands from referencing the current values into
55 /// those specified by VMap.
56 static inline void remapInstruction(Instruction *I,
57 ValueToValueMapTy &VMap) {
58 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
59 Value *Op = I->getOperand(op);
60 ValueToValueMapTy::iterator It = VMap.find(Op);
62 I->setOperand(op, It->second);
65 if (PHINode *PN = dyn_cast<PHINode>(I)) {
66 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
67 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
69 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
74 /// Folds a basic block into its predecessor if it only has one predecessor, and
75 /// that predecessor only has one successor.
76 /// The LoopInfo Analysis that is passed will be kept consistent. If folding is
77 /// successful references to the containing loop must be removed from
78 /// ScalarEvolution by calling ScalarEvolution::forgetLoop because SE may have
79 /// references to the eliminated BB. The argument ForgottenLoops contains a set
80 /// of loops that have already been forgotten to prevent redundant, expensive
81 /// calls to ScalarEvolution::forgetLoop. Returns the new combined block.
83 foldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, ScalarEvolution *SE,
84 SmallPtrSetImpl<Loop *> &ForgottenLoops,
86 // Merge basic blocks into their predecessor if there is only one distinct
87 // pred, and if there is only one distinct successor of the predecessor, and
88 // if there are no PHI nodes.
89 BasicBlock *OnlyPred = BB->getSinglePredecessor();
90 if (!OnlyPred) return nullptr;
92 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
95 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
97 // Resolve any PHI nodes at the start of the block. They are all
98 // guaranteed to have exactly one entry if they exist, unless there are
99 // multiple duplicate (but guaranteed to be equal) entries for the
100 // incoming edges. This occurs when there are multiple edges from
101 // OnlyPred to OnlySucc.
102 FoldSingleEntryPHINodes(BB);
104 // Delete the unconditional branch from the predecessor...
105 OnlyPred->getInstList().pop_back();
107 // Make all PHI nodes that referred to BB now refer to Pred as their
109 BB->replaceAllUsesWith(OnlyPred);
111 // Move all definitions in the successor to the predecessor...
112 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
114 // OldName will be valid until erased.
115 StringRef OldName = BB->getName();
117 // Erase the old block and update dominator info.
119 if (DomTreeNode *DTN = DT->getNode(BB)) {
120 DomTreeNode *PredDTN = DT->getNode(OnlyPred);
121 SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end());
122 for (auto *DI : Children)
123 DT->changeImmediateDominator(DI, PredDTN);
128 // ScalarEvolution holds references to loop exit blocks.
130 if (Loop *L = LI->getLoopFor(BB)) {
131 if (ForgottenLoops.insert(L).second)
137 // Inherit predecessor's name if it exists...
138 if (!OldName.empty() && !OnlyPred->hasName())
139 OnlyPred->setName(OldName);
141 BB->eraseFromParent();
146 /// Check if unrolling created a situation where we need to insert phi nodes to
147 /// preserve LCSSA form.
148 /// \param Blocks is a vector of basic blocks representing unrolled loop.
149 /// \param L is the outer loop.
150 /// It's possible that some of the blocks are in L, and some are not. In this
151 /// case, if there is a use is outside L, and definition is inside L, we need to
152 /// insert a phi-node, otherwise LCSSA will be broken.
153 /// The function is just a helper function for llvm::UnrollLoop that returns
154 /// true if this situation occurs, indicating that LCSSA needs to be fixed.
155 static bool needToInsertPhisForLCSSA(Loop *L, std::vector<BasicBlock *> Blocks,
157 for (BasicBlock *BB : Blocks) {
158 if (LI->getLoopFor(BB) == L)
160 for (Instruction &I : *BB) {
161 for (Use &U : I.operands()) {
162 if (auto Def = dyn_cast<Instruction>(U)) {
163 Loop *DefLoop = LI->getLoopFor(Def->getParent());
166 if (DefLoop->contains(L))
175 /// Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary
176 /// and adds a mapping from the original loop to the new loop to NewLoops.
177 /// Returns nullptr if no new loop was created and a pointer to the
178 /// original loop OriginalBB was part of otherwise.
179 const Loop* llvm::addClonedBlockToLoopInfo(BasicBlock *OriginalBB,
180 BasicBlock *ClonedBB, LoopInfo *LI,
181 NewLoopsMap &NewLoops) {
182 // Figure out which loop New is in.
183 const Loop *OldLoop = LI->getLoopFor(OriginalBB);
184 assert(OldLoop && "Should (at least) be in the loop being unrolled!");
186 Loop *&NewLoop = NewLoops[OldLoop];
188 // Found a new sub-loop.
189 assert(OriginalBB == OldLoop->getHeader() &&
190 "Header should be first in RPO");
192 Loop *NewLoopParent = NewLoops.lookup(OldLoop->getParentLoop());
193 assert(NewLoopParent &&
194 "Expected parent loop before sub-loop in RPO");
196 NewLoopParent->addChildLoop(NewLoop);
197 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
200 NewLoop->addBasicBlockToLoop(ClonedBB, *LI);
205 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
206 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
207 /// can only fail when the loop's latch block is not terminated by a conditional
208 /// branch instruction. However, if the trip count (and multiple) are not known,
209 /// loop unrolling will mostly produce more code that is no faster.
211 /// TripCount is the upper bound of the iteration on which control exits
212 /// LatchBlock. Control may exit the loop prior to TripCount iterations either
213 /// via an early branch in other loop block or via LatchBlock terminator. This
214 /// is relaxed from the general definition of trip count which is the number of
215 /// times the loop header executes. Note that UnrollLoop assumes that the loop
216 /// counter test is in LatchBlock in order to remove unnecesssary instances of
217 /// the test. If control can exit the loop from the LatchBlock's terminator
218 /// prior to TripCount iterations, flag PreserveCondBr needs to be set.
220 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock
221 /// needs to be preserved. It is needed when we use trip count upper bound to
222 /// fully unroll the loop. If PreserveOnlyFirst is also set then only the first
223 /// conditional branch needs to be preserved.
225 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
226 /// execute without exiting the loop.
228 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
229 /// have a runtime (i.e. not compile time constant) trip count. Unrolling these
230 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
231 /// iterations before branching into the unrolled loop. UnrollLoop will not
232 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
233 /// AllowExpensiveTripCount is false.
235 /// If we want to perform PGO-based loop peeling, PeelCount is set to the
236 /// number of iterations we want to peel off.
238 /// The LoopInfo Analysis that is passed will be kept consistent.
240 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
241 /// DominatorTree if they are non-null.
242 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool Force,
243 bool AllowRuntime, bool AllowExpensiveTripCount,
244 bool PreserveCondBr, bool PreserveOnlyFirst,
245 unsigned TripMultiple, unsigned PeelCount, LoopInfo *LI,
246 ScalarEvolution *SE, DominatorTree *DT,
247 AssumptionCache *AC, OptimizationRemarkEmitter *ORE,
248 bool PreserveLCSSA) {
250 BasicBlock *Preheader = L->getLoopPreheader();
252 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
256 BasicBlock *LatchBlock = L->getLoopLatch();
258 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
262 // Loops with indirectbr cannot be cloned.
263 if (!L->isSafeToClone()) {
264 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
268 BasicBlock *Header = L->getHeader();
269 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
271 if (!BI || BI->isUnconditional()) {
272 // The loop-rotate pass can be helpful to avoid this in many cases.
274 " Can't unroll; loop not terminated by a conditional branch.\n");
278 if (Header->hasAddressTaken()) {
279 // The loop-rotate pass can be helpful to avoid this in many cases.
281 " Won't unroll loop: address of header block is taken.\n");
286 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
287 if (TripMultiple != 1)
288 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
290 // Effectively "DCE" unrolled iterations that are beyond the tripcount
291 // and will never be executed.
292 if (TripCount != 0 && Count > TripCount)
295 // Don't enter the unroll code if there is nothing to do.
296 if (TripCount == 0 && Count < 2 && PeelCount == 0)
300 assert(TripMultiple > 0);
301 assert(TripCount == 0 || TripCount % TripMultiple == 0);
303 // Are we eliminating the loop control altogether?
304 bool CompletelyUnroll = Count == TripCount;
305 SmallVector<BasicBlock *, 4> ExitBlocks;
306 L->getExitBlocks(ExitBlocks);
307 std::vector<BasicBlock*> OriginalLoopBlocks = L->getBlocks();
309 // Go through all exits of L and see if there are any phi-nodes there. We just
310 // conservatively assume that they're inserted to preserve LCSSA form, which
311 // means that complete unrolling might break this form. We need to either fix
312 // it in-place after the transformation, or entirely rebuild LCSSA. TODO: For
313 // now we just recompute LCSSA for the outer loop, but it should be possible
314 // to fix it in-place.
315 bool NeedToFixLCSSA = PreserveLCSSA && CompletelyUnroll &&
316 any_of(ExitBlocks, [](const BasicBlock *BB) {
317 return isa<PHINode>(BB->begin());
320 // We assume a run-time trip count if the compiler cannot
321 // figure out the loop trip count and the unroll-runtime
322 // flag is specified.
323 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
325 assert((!RuntimeTripCount || !PeelCount) &&
326 "Did not expect runtime trip-count unrolling "
327 "and peeling for the same loop");
330 peelLoop(L, PeelCount, LI, SE, DT, PreserveLCSSA);
332 // Loops containing convergent instructions must have a count that divides
333 // their TripMultiple.
336 bool HasConvergent = false;
337 for (auto &BB : L->blocks())
339 if (auto CS = CallSite(&I))
340 HasConvergent |= CS.isConvergent();
341 assert((!HasConvergent || TripMultiple % Count == 0) &&
342 "Unroll count must divide trip multiple if loop contains a "
343 "convergent operation.");
346 if (RuntimeTripCount && TripMultiple % Count != 0 &&
347 !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount,
348 UnrollRuntimeEpilog, LI, SE, DT,
351 RuntimeTripCount = false;
356 // Notify ScalarEvolution that the loop will be substantially changed,
357 // if not outright eliminated.
361 // If we know the trip count, we know the multiple...
362 unsigned BreakoutTrip = 0;
363 if (TripCount != 0) {
364 BreakoutTrip = TripCount % Count;
367 // Figure out what multiple to use.
368 BreakoutTrip = TripMultiple =
369 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
373 // Report the unrolling decision.
374 if (CompletelyUnroll) {
375 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
376 << " with trip count " << TripCount << "!\n");
377 ORE->emit(OptimizationRemark(DEBUG_TYPE, "FullyUnrolled", L->getStartLoc(),
379 << "completely unrolled loop with "
380 << NV("UnrollCount", TripCount) << " iterations");
381 } else if (PeelCount) {
382 DEBUG(dbgs() << "PEELING loop %" << Header->getName()
383 << " with iteration count " << PeelCount << "!\n");
384 ORE->emit(OptimizationRemark(DEBUG_TYPE, "Peeled", L->getStartLoc(),
386 << " peeled loop by " << NV("PeelCount", PeelCount)
389 OptimizationRemark Diag(DEBUG_TYPE, "PartialUnrolled", L->getStartLoc(),
391 Diag << "unrolled loop by a factor of " << NV("UnrollCount", Count);
393 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
395 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
396 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
397 ORE->emit(Diag << " with a breakout at trip "
398 << NV("BreakoutTrip", BreakoutTrip));
399 } else if (TripMultiple != 1) {
400 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
401 ORE->emit(Diag << " with " << NV("TripMultiple", TripMultiple)
402 << " trips per branch");
403 } else if (RuntimeTripCount) {
404 DEBUG(dbgs() << " with run-time trip count");
405 ORE->emit(Diag << " with run-time trip count");
407 DEBUG(dbgs() << "!\n");
410 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
411 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
413 // For the first iteration of the loop, we should use the precloned values for
414 // PHI nodes. Insert associations now.
415 ValueToValueMapTy LastValueMap;
416 std::vector<PHINode*> OrigPHINode;
417 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
418 OrigPHINode.push_back(cast<PHINode>(I));
421 std::vector<BasicBlock*> Headers;
422 std::vector<BasicBlock*> Latches;
423 Headers.push_back(Header);
424 Latches.push_back(LatchBlock);
426 // The current on-the-fly SSA update requires blocks to be processed in
427 // reverse postorder so that LastValueMap contains the correct value at each
429 LoopBlocksDFS DFS(L);
432 // Stash the DFS iterators before adding blocks to the loop.
433 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
434 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
436 std::vector<BasicBlock*> UnrolledLoopBlocks = L->getBlocks();
438 // Loop Unrolling might create new loops. While we do preserve LoopInfo, we
439 // might break loop-simplified form for these loops (as they, e.g., would
440 // share the same exit blocks). We'll keep track of loops for which we can
441 // break this so that later we can re-simplify them.
442 SmallSetVector<Loop *, 4> LoopsToSimplify;
443 for (Loop *SubLoop : *L)
444 LoopsToSimplify.insert(SubLoop);
446 for (unsigned It = 1; It != Count; ++It) {
447 std::vector<BasicBlock*> NewBlocks;
448 SmallDenseMap<const Loop *, Loop *, 4> NewLoops;
451 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
452 ValueToValueMapTy VMap;
453 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
454 Header->getParent()->getBasicBlockList().push_back(New);
456 // Tell LI about New.
458 assert(LI->getLoopFor(*BB) == L && "Header should not be in a sub-loop");
459 L->addBasicBlockToLoop(New, *LI);
461 const Loop *OldLoop = addClonedBlockToLoopInfo(*BB, New, LI, NewLoops);
463 LoopsToSimplify.insert(NewLoops[OldLoop]);
465 // Forget the old loop, since its inputs may have changed.
467 SE->forgetLoop(OldLoop);
472 // Loop over all of the PHI nodes in the block, changing them to use
473 // the incoming values from the previous block.
474 for (PHINode *OrigPHI : OrigPHINode) {
475 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHI]);
476 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
477 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
478 if (It > 1 && L->contains(InValI))
479 InVal = LastValueMap[InValI];
480 VMap[OrigPHI] = InVal;
481 New->getInstList().erase(NewPHI);
484 // Update our running map of newest clones
485 LastValueMap[*BB] = New;
486 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
488 LastValueMap[VI->first] = VI->second;
490 // Add phi entries for newly created values to all exit blocks.
491 for (BasicBlock *Succ : successors(*BB)) {
492 if (L->contains(Succ))
494 for (BasicBlock::iterator BBI = Succ->begin();
495 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
496 Value *Incoming = phi->getIncomingValueForBlock(*BB);
497 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
498 if (It != LastValueMap.end())
499 Incoming = It->second;
500 phi->addIncoming(Incoming, New);
503 // Keep track of new headers and latches as we create them, so that
504 // we can insert the proper branches later.
506 Headers.push_back(New);
507 if (*BB == LatchBlock)
508 Latches.push_back(New);
510 NewBlocks.push_back(New);
511 UnrolledLoopBlocks.push_back(New);
513 // Update DomTree: since we just copy the loop body, and each copy has a
514 // dedicated entry block (copy of the header block), this header's copy
515 // dominates all copied blocks. That means, dominance relations in the
516 // copied body are the same as in the original body.
519 DT->addNewBlock(New, Latches[It - 1]);
521 auto BBDomNode = DT->getNode(*BB);
522 auto BBIDom = BBDomNode->getIDom();
523 BasicBlock *OriginalBBIDom = BBIDom->getBlock();
525 New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
530 // Remap all instructions in the most recent iteration
531 for (BasicBlock *NewBlock : NewBlocks) {
532 for (Instruction &I : *NewBlock) {
533 ::remapInstruction(&I, LastValueMap);
534 if (auto *II = dyn_cast<IntrinsicInst>(&I))
535 if (II->getIntrinsicID() == Intrinsic::assume)
536 AC->registerAssumption(II);
541 // Loop over the PHI nodes in the original block, setting incoming values.
542 for (PHINode *PN : OrigPHINode) {
543 if (CompletelyUnroll) {
544 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
545 Header->getInstList().erase(PN);
547 else if (Count > 1) {
548 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
549 // If this value was defined in the loop, take the value defined by the
550 // last iteration of the loop.
551 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
552 if (L->contains(InValI))
553 InVal = LastValueMap[InVal];
555 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
556 PN->addIncoming(InVal, Latches.back());
560 // Now that all the basic blocks for the unrolled iterations are in place,
561 // set up the branches to connect them.
562 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
563 // The original branch was replicated in each unrolled iteration.
564 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
566 // The branch destination.
567 unsigned j = (i + 1) % e;
568 BasicBlock *Dest = Headers[j];
569 bool NeedConditional = true;
571 if (RuntimeTripCount && j != 0) {
572 NeedConditional = false;
575 // For a complete unroll, make the last iteration end with a branch
576 // to the exit block.
577 if (CompletelyUnroll) {
580 // If using trip count upper bound to completely unroll, we need to keep
581 // the conditional branch except the last one because the loop may exit
582 // after any iteration.
583 assert(NeedConditional &&
584 "NeedCondition cannot be modified by both complete "
585 "unrolling and runtime unrolling");
586 NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0));
587 } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
588 // If we know the trip count or a multiple of it, we can safely use an
589 // unconditional branch for some iterations.
590 NeedConditional = false;
593 if (NeedConditional) {
594 // Update the conditional branch's successor for the following
596 Term->setSuccessor(!ContinueOnTrue, Dest);
598 // Remove phi operands at this loop exit
599 if (Dest != LoopExit) {
600 BasicBlock *BB = Latches[i];
601 for (BasicBlock *Succ: successors(BB)) {
602 if (Succ == Headers[i])
604 for (BasicBlock::iterator BBI = Succ->begin();
605 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
606 Phi->removeIncomingValue(BB, false);
610 // Replace the conditional branch with an unconditional one.
611 BranchInst::Create(Dest, Term);
612 Term->eraseFromParent();
615 // Update dominators of blocks we might reach through exits.
616 // Immediate dominator of such block might change, because we add more
617 // routes which can lead to the exit: we can now reach it from the copied
618 // iterations too. Thus, the new idom of the block will be the nearest
619 // common dominator of the previous idom and common dominator of all copies of
620 // the previous idom. This is equivalent to the nearest common dominator of
621 // the previous idom and the first latch, which dominates all copies of the
623 if (DT && Count > 1) {
624 for (auto *BB : OriginalLoopBlocks) {
625 auto *BBDomNode = DT->getNode(BB);
626 SmallVector<BasicBlock *, 16> ChildrenToUpdate;
627 for (auto *ChildDomNode : BBDomNode->getChildren()) {
628 auto *ChildBB = ChildDomNode->getBlock();
629 if (!L->contains(ChildBB))
630 ChildrenToUpdate.push_back(ChildBB);
632 BasicBlock *NewIDom = DT->findNearestCommonDominator(BB, Latches[0]);
633 for (auto *ChildBB : ChildrenToUpdate)
634 DT->changeImmediateDominator(ChildBB, NewIDom);
638 // Merge adjacent basic blocks, if possible.
639 SmallPtrSet<Loop *, 4> ForgottenLoops;
640 for (BasicBlock *Latch : Latches) {
641 BranchInst *Term = cast<BranchInst>(Latch->getTerminator());
642 if (Term->isUnconditional()) {
643 BasicBlock *Dest = Term->getSuccessor(0);
644 if (BasicBlock *Fold =
645 foldBlockIntoPredecessor(Dest, LI, SE, ForgottenLoops, DT)) {
646 // Dest has been folded into Fold. Update our worklists accordingly.
647 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
648 UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(),
649 UnrolledLoopBlocks.end(), Dest),
650 UnrolledLoopBlocks.end());
655 // FIXME: We only preserve DT info for complete unrolling now. Incrementally
656 // updating domtree after partial loop unrolling should also be easy.
657 if (DT && !CompletelyUnroll)
658 DT->recalculate(*L->getHeader()->getParent());
660 DEBUG(DT->verifyDomTree());
662 // Simplify any new induction variables in the partially unrolled loop.
663 if (SE && !CompletelyUnroll && Count > 1) {
664 SmallVector<WeakVH, 16> DeadInsts;
665 simplifyLoopIVs(L, SE, DT, LI, DeadInsts);
667 // Aggressively clean up dead instructions that simplifyLoopIVs already
668 // identified. Any remaining should be cleaned up below.
669 while (!DeadInsts.empty())
670 if (Instruction *Inst =
671 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
672 RecursivelyDeleteTriviallyDeadInstructions(Inst);
675 // At this point, the code is well formed. We now do a quick sweep over the
676 // inserted code, doing constant propagation and dead code elimination as we
678 const DataLayout &DL = Header->getModule()->getDataLayout();
679 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
680 for (BasicBlock *BB : NewLoopBlocks) {
681 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
682 Instruction *Inst = &*I++;
684 if (Value *V = SimplifyInstruction(Inst, DL))
685 if (LI->replacementPreservesLCSSAForm(Inst, V))
686 Inst->replaceAllUsesWith(V);
687 if (isInstructionTriviallyDead(Inst))
688 BB->getInstList().erase(Inst);
692 // TODO: after peeling or unrolling, previously loop variant conditions are
693 // likely to fold to constants, eagerly propagating those here will require
694 // fewer cleanup passes to be run. Alternatively, a LoopEarlyCSE might be
697 NumCompletelyUnrolled += CompletelyUnroll;
700 Loop *OuterL = L->getParentLoop();
701 // Update LoopInfo if the loop is completely removed.
702 if (CompletelyUnroll)
703 LI->markAsRemoved(L);
705 // After complete unrolling most of the blocks should be contained in OuterL.
706 // However, some of them might happen to be out of OuterL (e.g. if they
707 // precede a loop exit). In this case we might need to insert PHI nodes in
708 // order to preserve LCSSA form.
709 // We don't need to check this if we already know that we need to fix LCSSA
711 // TODO: For now we just recompute LCSSA for the outer loop in this case, but
712 // it should be possible to fix it in-place.
713 if (PreserveLCSSA && OuterL && CompletelyUnroll && !NeedToFixLCSSA)
714 NeedToFixLCSSA |= ::needToInsertPhisForLCSSA(OuterL, UnrolledLoopBlocks, LI);
716 // If we have a pass and a DominatorTree we should re-simplify impacted loops
717 // to ensure subsequent analyses can rely on this form. We want to simplify
718 // at least one layer outside of the loop that was unrolled so that any
719 // changes to the parent loop exposed by the unrolling are considered.
721 if (!OuterL && !CompletelyUnroll)
724 // OuterL includes all loops for which we can break loop-simplify, so
725 // it's sufficient to simplify only it (it'll recursively simplify inner
727 // TODO: That potentially might be compile-time expensive. We should try
728 // to fix the loop-simplified form incrementally.
729 simplifyLoop(OuterL, DT, LI, SE, AC, PreserveLCSSA);
731 // LCSSA must be performed on the outermost affected loop. The unrolled
732 // loop's last loop latch is guaranteed to be in the outermost loop after
733 // LoopInfo's been updated by markAsRemoved.
734 Loop *LatchLoop = LI->getLoopFor(Latches.back());
735 if (!OuterL->contains(LatchLoop))
736 while (OuterL->getParentLoop() != LatchLoop)
737 OuterL = OuterL->getParentLoop();
740 formLCSSARecursively(*OuterL, *DT, LI, SE);
742 assert(OuterL->isLCSSAForm(*DT) &&
743 "Loops should be in LCSSA form after loop-unroll.");
745 // Simplify loops for which we might've broken loop-simplify form.
746 for (Loop *SubLoop : LoopsToSimplify)
747 simplifyLoop(SubLoop, DT, LI, SE, AC, PreserveLCSSA);
754 /// Given an llvm.loop loop id metadata node, returns the loop hint metadata
755 /// node with the given name (for example, "llvm.loop.unroll.count"). If no
756 /// such metadata node exists, then nullptr is returned.
757 MDNode *llvm::GetUnrollMetadata(MDNode *LoopID, StringRef Name) {
758 // First operand should refer to the loop id itself.
759 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
760 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
762 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
763 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
767 MDString *S = dyn_cast<MDString>(MD->getOperand(0));
771 if (Name.equals(S->getString()))