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/InstructionSimplify.h"
23 #include "llvm/Analysis/LoopIterator.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/IR/BasicBlock.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/DiagnosticInfo.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Cloning.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Transforms/Utils/LoopUtils.h"
37 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
40 #define DEBUG_TYPE "loop-unroll"
42 // TODO: Should these be here or in LoopUnroll?
43 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
44 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
46 /// RemapInstruction - Convert the instruction operands from referencing the
47 /// current values into those specified by VMap.
48 static inline void RemapInstruction(Instruction *I,
49 ValueToValueMapTy &VMap) {
50 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
51 Value *Op = I->getOperand(op);
52 ValueToValueMapTy::iterator It = VMap.find(Op);
54 I->setOperand(op, It->second);
57 if (PHINode *PN = dyn_cast<PHINode>(I)) {
58 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
59 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
61 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
66 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
67 /// only has one predecessor, and that predecessor only has one successor.
68 /// The LoopInfo Analysis that is passed will be kept consistent. If folding is
69 /// successful references to the containing loop must be removed from
70 /// ScalarEvolution by calling ScalarEvolution::forgetLoop because SE may have
71 /// references to the eliminated BB. The argument ForgottenLoops contains a set
72 /// of loops that have already been forgotten to prevent redundant, expensive
73 /// calls to ScalarEvolution::forgetLoop. Returns the new combined block.
75 FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI, LPPassManager *LPM,
76 SmallPtrSetImpl<Loop *> &ForgottenLoops) {
77 // Merge basic blocks into their predecessor if there is only one distinct
78 // pred, and if there is only one distinct successor of the predecessor, and
79 // if there are no PHI nodes.
80 BasicBlock *OnlyPred = BB->getSinglePredecessor();
81 if (!OnlyPred) return nullptr;
83 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
86 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
88 // Resolve any PHI nodes at the start of the block. They are all
89 // guaranteed to have exactly one entry if they exist, unless there are
90 // multiple duplicate (but guaranteed to be equal) entries for the
91 // incoming edges. This occurs when there are multiple edges from
92 // OnlyPred to OnlySucc.
93 FoldSingleEntryPHINodes(BB);
95 // Delete the unconditional branch from the predecessor...
96 OnlyPred->getInstList().pop_back();
98 // Make all PHI nodes that referred to BB now refer to Pred as their
100 BB->replaceAllUsesWith(OnlyPred);
102 // Move all definitions in the successor to the predecessor...
103 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
105 // OldName will be valid until erased.
106 StringRef OldName = BB->getName();
108 // Erase basic block from the function...
110 // ScalarEvolution holds references to loop exit blocks.
112 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
113 if (Loop *L = LI->getLoopFor(BB)) {
114 if (ForgottenLoops.insert(L))
121 // Inherit predecessor's name if it exists...
122 if (!OldName.empty() && !OnlyPred->hasName())
123 OnlyPred->setName(OldName);
125 BB->eraseFromParent();
130 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
131 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
132 /// can only fail when the loop's latch block is not terminated by a conditional
133 /// branch instruction. However, if the trip count (and multiple) are not known,
134 /// loop unrolling will mostly produce more code that is no faster.
136 /// TripCount is generally defined as the number of times the loop header
137 /// executes. UnrollLoop relaxes the definition to permit early exits: here
138 /// TripCount is the iteration on which control exits LatchBlock if no early
139 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
140 /// terminates LatchBlock in order to remove unnecesssary instances of the
141 /// test. In other words, control may exit the loop prior to TripCount
142 /// iterations via an early branch, but control may not exit the loop from the
143 /// LatchBlock's terminator prior to TripCount iterations.
145 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
146 /// execute without exiting the loop.
148 /// The LoopInfo Analysis that is passed will be kept consistent.
150 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
151 /// removed from the LoopPassManager as well. LPM can also be NULL.
153 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
154 /// available from the Pass it must also preserve those analyses.
155 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
156 bool AllowRuntime, unsigned TripMultiple,
157 LoopInfo *LI, Pass *PP, LPPassManager *LPM) {
158 BasicBlock *Preheader = L->getLoopPreheader();
160 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
164 BasicBlock *LatchBlock = L->getLoopLatch();
166 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
170 // Loops with indirectbr cannot be cloned.
171 if (!L->isSafeToClone()) {
172 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
176 BasicBlock *Header = L->getHeader();
177 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
179 if (!BI || BI->isUnconditional()) {
180 // The loop-rotate pass can be helpful to avoid this in many cases.
182 " Can't unroll; loop not terminated by a conditional branch.\n");
186 if (Header->hasAddressTaken()) {
187 // The loop-rotate pass can be helpful to avoid this in many cases.
189 " Won't unroll loop: address of header block is taken.\n");
194 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
195 if (TripMultiple != 1)
196 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
198 // Effectively "DCE" unrolled iterations that are beyond the tripcount
199 // and will never be executed.
200 if (TripCount != 0 && Count > TripCount)
203 // Don't enter the unroll code if there is nothing to do. This way we don't
204 // need to support "partial unrolling by 1".
205 if (TripCount == 0 && Count < 2)
209 assert(TripMultiple > 0);
210 assert(TripCount == 0 || TripCount % TripMultiple == 0);
212 // Are we eliminating the loop control altogether?
213 bool CompletelyUnroll = Count == TripCount;
215 // We assume a run-time trip count if the compiler cannot
216 // figure out the loop trip count and the unroll-runtime
217 // flag is specified.
218 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
220 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
223 // Notify ScalarEvolution that the loop will be substantially changed,
224 // if not outright eliminated.
226 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
231 // If we know the trip count, we know the multiple...
232 unsigned BreakoutTrip = 0;
233 if (TripCount != 0) {
234 BreakoutTrip = TripCount % Count;
237 // Figure out what multiple to use.
238 BreakoutTrip = TripMultiple =
239 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
242 // Report the unrolling decision.
243 DebugLoc LoopLoc = L->getStartLoc();
244 Function *F = Header->getParent();
245 LLVMContext &Ctx = F->getContext();
247 if (CompletelyUnroll) {
248 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
249 << " with trip count " << TripCount << "!\n");
250 emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
251 Twine("completely unrolled loop with ") +
252 Twine(TripCount) + " iterations");
254 auto EmitDiag = [&](const Twine &T) {
255 emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
256 "unrolled loop by a factor of " + Twine(Count) +
260 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
262 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
263 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
264 EmitDiag(" with a breakout at trip " + Twine(BreakoutTrip));
265 } else if (TripMultiple != 1) {
266 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
267 EmitDiag(" with " + Twine(TripMultiple) + " trips per branch");
268 } else if (RuntimeTripCount) {
269 DEBUG(dbgs() << " with run-time trip count");
270 EmitDiag(" with run-time trip count");
272 DEBUG(dbgs() << "!\n");
275 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
276 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
278 // For the first iteration of the loop, we should use the precloned values for
279 // PHI nodes. Insert associations now.
280 ValueToValueMapTy LastValueMap;
281 std::vector<PHINode*> OrigPHINode;
282 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
283 OrigPHINode.push_back(cast<PHINode>(I));
286 std::vector<BasicBlock*> Headers;
287 std::vector<BasicBlock*> Latches;
288 Headers.push_back(Header);
289 Latches.push_back(LatchBlock);
291 // The current on-the-fly SSA update requires blocks to be processed in
292 // reverse postorder so that LastValueMap contains the correct value at each
294 LoopBlocksDFS DFS(L);
297 // Stash the DFS iterators before adding blocks to the loop.
298 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
299 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
301 for (unsigned It = 1; It != Count; ++It) {
302 std::vector<BasicBlock*> NewBlocks;
304 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
305 ValueToValueMapTy VMap;
306 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
307 Header->getParent()->getBasicBlockList().push_back(New);
309 // Loop over all of the PHI nodes in the block, changing them to use the
310 // incoming values from the previous block.
312 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
313 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
314 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
315 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
316 if (It > 1 && L->contains(InValI))
317 InVal = LastValueMap[InValI];
318 VMap[OrigPHINode[i]] = InVal;
319 New->getInstList().erase(NewPHI);
322 // Update our running map of newest clones
323 LastValueMap[*BB] = New;
324 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
326 LastValueMap[VI->first] = VI->second;
328 L->addBasicBlockToLoop(New, LI->getBase());
330 // Add phi entries for newly created values to all exit blocks.
331 for (BasicBlock *Succ : successors(*BB)) {
332 if (L->contains(Succ))
334 for (BasicBlock::iterator BBI = Succ->begin();
335 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
336 Value *Incoming = phi->getIncomingValueForBlock(*BB);
337 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
338 if (It != LastValueMap.end())
339 Incoming = It->second;
340 phi->addIncoming(Incoming, New);
343 // Keep track of new headers and latches as we create them, so that
344 // we can insert the proper branches later.
346 Headers.push_back(New);
347 if (*BB == LatchBlock)
348 Latches.push_back(New);
350 NewBlocks.push_back(New);
353 // Remap all instructions in the most recent iteration
354 for (unsigned i = 0; i < NewBlocks.size(); ++i)
355 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
356 E = NewBlocks[i]->end(); I != E; ++I)
357 ::RemapInstruction(I, LastValueMap);
360 // Loop over the PHI nodes in the original block, setting incoming values.
361 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
362 PHINode *PN = OrigPHINode[i];
363 if (CompletelyUnroll) {
364 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
365 Header->getInstList().erase(PN);
367 else if (Count > 1) {
368 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
369 // If this value was defined in the loop, take the value defined by the
370 // last iteration of the loop.
371 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
372 if (L->contains(InValI))
373 InVal = LastValueMap[InVal];
375 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
376 PN->addIncoming(InVal, Latches.back());
380 // Now that all the basic blocks for the unrolled iterations are in place,
381 // set up the branches to connect them.
382 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
383 // The original branch was replicated in each unrolled iteration.
384 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
386 // The branch destination.
387 unsigned j = (i + 1) % e;
388 BasicBlock *Dest = Headers[j];
389 bool NeedConditional = true;
391 if (RuntimeTripCount && j != 0) {
392 NeedConditional = false;
395 // For a complete unroll, make the last iteration end with a branch
396 // to the exit block.
397 if (CompletelyUnroll && j == 0) {
399 NeedConditional = false;
402 // If we know the trip count or a multiple of it, we can safely use an
403 // unconditional branch for some iterations.
404 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
405 NeedConditional = false;
408 if (NeedConditional) {
409 // Update the conditional branch's successor for the following
411 Term->setSuccessor(!ContinueOnTrue, Dest);
413 // Remove phi operands at this loop exit
414 if (Dest != LoopExit) {
415 BasicBlock *BB = Latches[i];
416 for (BasicBlock *Succ : successors(BB)) {
417 if (Succ == Headers[i])
419 for (BasicBlock::iterator BBI = Succ->begin();
420 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
421 Phi->removeIncomingValue(BB, false);
425 // Replace the conditional branch with an unconditional one.
426 BranchInst::Create(Dest, Term);
427 Term->eraseFromParent();
431 // Merge adjacent basic blocks, if possible.
432 SmallPtrSet<Loop *, 4> ForgottenLoops;
433 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
434 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
435 if (Term->isUnconditional()) {
436 BasicBlock *Dest = Term->getSuccessor(0);
437 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM,
439 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
443 DominatorTree *DT = nullptr;
445 // FIXME: Reconstruct dom info, because it is not preserved properly.
446 // Incrementally updating domtree after loop unrolling would be easy.
447 if (DominatorTreeWrapperPass *DTWP =
448 PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
449 DT = &DTWP->getDomTree();
450 DT->recalculate(*L->getHeader()->getParent());
453 // Simplify any new induction variables in the partially unrolled loop.
454 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
455 if (SE && !CompletelyUnroll) {
456 SmallVector<WeakVH, 16> DeadInsts;
457 simplifyLoopIVs(L, SE, LPM, DeadInsts);
459 // Aggressively clean up dead instructions that simplifyLoopIVs already
460 // identified. Any remaining should be cleaned up below.
461 while (!DeadInsts.empty())
462 if (Instruction *Inst =
463 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
464 RecursivelyDeleteTriviallyDeadInstructions(Inst);
467 // At this point, the code is well formed. We now do a quick sweep over the
468 // inserted code, doing constant propagation and dead code elimination as we
470 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
471 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
472 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
473 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
474 Instruction *Inst = I++;
476 if (isInstructionTriviallyDead(Inst))
477 (*BB)->getInstList().erase(Inst);
478 else if (Value *V = SimplifyInstruction(Inst))
479 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
480 Inst->replaceAllUsesWith(V);
481 (*BB)->getInstList().erase(Inst);
485 NumCompletelyUnrolled += CompletelyUnroll;
488 Loop *OuterL = L->getParentLoop();
489 // Remove the loop from the LoopPassManager if it's completely removed.
490 if (CompletelyUnroll && LPM != nullptr)
491 LPM->deleteLoopFromQueue(L);
493 // If we have a pass and a DominatorTree we should re-simplify impacted loops
494 // to ensure subsequent analyses can rely on this form. We want to simplify
495 // at least one layer outside of the loop that was unrolled so that any
496 // changes to the parent loop exposed by the unrolling are considered.
498 if (!OuterL && !CompletelyUnroll)
501 DataLayoutPass *DLP = PP->getAnalysisIfAvailable<DataLayoutPass>();
502 const DataLayout *DL = DLP ? &DLP->getDataLayout() : nullptr;
503 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
504 simplifyLoop(OuterL, DT, LI, PP, /*AliasAnalysis*/ nullptr, SE, DL);
506 // LCSSA must be performed on the outermost affected loop. The unrolled
507 // loop's last loop latch is guaranteed to be in the outermost loop after
508 // deleteLoopFromQueue updates LoopInfo.
509 Loop *LatchLoop = LI->getLoopFor(Latches.back());
510 if (!OuterL->contains(LatchLoop))
511 while (OuterL->getParentLoop() != LatchLoop)
512 OuterL = OuterL->getParentLoop();
514 formLCSSARecursively(*OuterL, *DT, SE);