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/Statistic.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/Analysis/LoopIterator.h"
23 #include "llvm/Analysis/LoopPass.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/DiagnosticInfo.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Cloning.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/Transforms/Utils/LoopUtils.h"
35 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
38 #define DEBUG_TYPE "loop-unroll"
40 // TODO: Should these be here or in LoopUnroll?
41 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
42 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
44 /// RemapInstruction - Convert the instruction operands from referencing the
45 /// current values into those specified by VMap.
46 static inline void RemapInstruction(Instruction *I,
47 ValueToValueMapTy &VMap) {
48 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
49 Value *Op = I->getOperand(op);
50 ValueToValueMapTy::iterator It = VMap.find(Op);
52 I->setOperand(op, It->second);
55 if (PHINode *PN = dyn_cast<PHINode>(I)) {
56 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
57 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
59 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
64 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
65 /// only has one predecessor, and that predecessor only has one successor.
66 /// The LoopInfo Analysis that is passed will be kept consistent.
67 /// Returns the new combined block.
68 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI,
70 // Merge basic blocks into their predecessor if there is only one distinct
71 // pred, and if there is only one distinct successor of the predecessor, and
72 // if there are no PHI nodes.
73 BasicBlock *OnlyPred = BB->getSinglePredecessor();
74 if (!OnlyPred) return nullptr;
76 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
79 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
81 // Resolve any PHI nodes at the start of the block. They are all
82 // guaranteed to have exactly one entry if they exist, unless there are
83 // multiple duplicate (but guaranteed to be equal) entries for the
84 // incoming edges. This occurs when there are multiple edges from
85 // OnlyPred to OnlySucc.
86 FoldSingleEntryPHINodes(BB);
88 // Delete the unconditional branch from the predecessor...
89 OnlyPred->getInstList().pop_back();
91 // Make all PHI nodes that referred to BB now refer to Pred as their
93 BB->replaceAllUsesWith(OnlyPred);
95 // Move all definitions in the successor to the predecessor...
96 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
98 // OldName will be valid until erased.
99 StringRef OldName = BB->getName();
101 // Erase basic block from the function...
103 // ScalarEvolution holds references to loop exit blocks.
105 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
106 if (Loop *L = LI->getLoopFor(BB))
112 // Inherit predecessor's name if it exists...
113 if (!OldName.empty() && !OnlyPred->hasName())
114 OnlyPred->setName(OldName);
116 BB->eraseFromParent();
121 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
122 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
123 /// can only fail when the loop's latch block is not terminated by a conditional
124 /// branch instruction. However, if the trip count (and multiple) are not known,
125 /// loop unrolling will mostly produce more code that is no faster.
127 /// TripCount is generally defined as the number of times the loop header
128 /// executes. UnrollLoop relaxes the definition to permit early exits: here
129 /// TripCount is the iteration on which control exits LatchBlock if no early
130 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
131 /// terminates LatchBlock in order to remove unnecesssary instances of the
132 /// test. In other words, control may exit the loop prior to TripCount
133 /// iterations via an early branch, but control may not exit the loop from the
134 /// LatchBlock's terminator prior to TripCount iterations.
136 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
137 /// execute without exiting the loop.
139 /// The LoopInfo Analysis that is passed will be kept consistent.
141 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
142 /// removed from the LoopPassManager as well. LPM can also be NULL.
144 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
145 /// available from the Pass it must also preserve those analyses.
146 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
147 bool AllowRuntime, unsigned TripMultiple,
148 LoopInfo *LI, Pass *PP, LPPassManager *LPM) {
149 BasicBlock *Preheader = L->getLoopPreheader();
151 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
155 BasicBlock *LatchBlock = L->getLoopLatch();
157 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
161 // Loops with indirectbr cannot be cloned.
162 if (!L->isSafeToClone()) {
163 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
167 BasicBlock *Header = L->getHeader();
168 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
170 if (!BI || BI->isUnconditional()) {
171 // The loop-rotate pass can be helpful to avoid this in many cases.
173 " Can't unroll; loop not terminated by a conditional branch.\n");
177 if (Header->hasAddressTaken()) {
178 // The loop-rotate pass can be helpful to avoid this in many cases.
180 " Won't unroll loop: address of header block is taken.\n");
185 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
186 if (TripMultiple != 1)
187 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
189 // Effectively "DCE" unrolled iterations that are beyond the tripcount
190 // and will never be executed.
191 if (TripCount != 0 && Count > TripCount)
194 // Don't enter the unroll code if there is nothing to do. This way we don't
195 // need to support "partial unrolling by 1".
196 if (TripCount == 0 && Count < 2)
200 assert(TripMultiple > 0);
201 assert(TripCount == 0 || TripCount % TripMultiple == 0);
203 // Are we eliminating the loop control altogether?
204 bool CompletelyUnroll = Count == TripCount;
206 // We assume a run-time trip count if the compiler cannot
207 // figure out the loop trip count and the unroll-runtime
208 // flag is specified.
209 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
211 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
214 // Notify ScalarEvolution that the loop will be substantially changed,
215 // if not outright eliminated.
217 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
222 // If we know the trip count, we know the multiple...
223 unsigned BreakoutTrip = 0;
224 if (TripCount != 0) {
225 BreakoutTrip = TripCount % Count;
228 // Figure out what multiple to use.
229 BreakoutTrip = TripMultiple =
230 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
233 // Report the unrolling decision.
234 DebugLoc LoopLoc = L->getStartLoc();
235 Function *F = Header->getParent();
236 LLVMContext &Ctx = F->getContext();
238 if (CompletelyUnroll) {
239 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
240 << " with trip count " << TripCount << "!\n");
241 emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
242 Twine("completely unrolled loop with ") +
243 Twine(TripCount) + " iterations");
245 auto EmitDiag = [&](const Twine &T) {
246 emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, LoopLoc,
247 "unrolled loop by a factor of " + Twine(Count) +
251 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
253 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
254 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
255 EmitDiag(" with a breakout at trip " + Twine(BreakoutTrip));
256 } else if (TripMultiple != 1) {
257 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
258 EmitDiag(" with " + Twine(TripMultiple) + " trips per branch");
259 } else if (RuntimeTripCount) {
260 DEBUG(dbgs() << " with run-time trip count");
261 EmitDiag(" with run-time trip count");
263 DEBUG(dbgs() << "!\n");
266 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
267 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
269 // For the first iteration of the loop, we should use the precloned values for
270 // PHI nodes. Insert associations now.
271 ValueToValueMapTy LastValueMap;
272 std::vector<PHINode*> OrigPHINode;
273 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
274 OrigPHINode.push_back(cast<PHINode>(I));
277 std::vector<BasicBlock*> Headers;
278 std::vector<BasicBlock*> Latches;
279 Headers.push_back(Header);
280 Latches.push_back(LatchBlock);
282 // The current on-the-fly SSA update requires blocks to be processed in
283 // reverse postorder so that LastValueMap contains the correct value at each
285 LoopBlocksDFS DFS(L);
288 // Stash the DFS iterators before adding blocks to the loop.
289 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
290 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
292 for (unsigned It = 1; It != Count; ++It) {
293 std::vector<BasicBlock*> NewBlocks;
295 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
296 ValueToValueMapTy VMap;
297 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
298 Header->getParent()->getBasicBlockList().push_back(New);
300 // Loop over all of the PHI nodes in the block, changing them to use the
301 // incoming values from the previous block.
303 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
304 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
305 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
306 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
307 if (It > 1 && L->contains(InValI))
308 InVal = LastValueMap[InValI];
309 VMap[OrigPHINode[i]] = InVal;
310 New->getInstList().erase(NewPHI);
313 // Update our running map of newest clones
314 LastValueMap[*BB] = New;
315 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
317 LastValueMap[VI->first] = VI->second;
319 L->addBasicBlockToLoop(New, LI->getBase());
321 // Add phi entries for newly created values to all exit blocks.
322 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
324 if (L->contains(*SI))
326 for (BasicBlock::iterator BBI = (*SI)->begin();
327 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
328 Value *Incoming = phi->getIncomingValueForBlock(*BB);
329 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
330 if (It != LastValueMap.end())
331 Incoming = It->second;
332 phi->addIncoming(Incoming, New);
335 // Keep track of new headers and latches as we create them, so that
336 // we can insert the proper branches later.
338 Headers.push_back(New);
339 if (*BB == LatchBlock)
340 Latches.push_back(New);
342 NewBlocks.push_back(New);
345 // Remap all instructions in the most recent iteration
346 for (unsigned i = 0; i < NewBlocks.size(); ++i)
347 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
348 E = NewBlocks[i]->end(); I != E; ++I)
349 ::RemapInstruction(I, LastValueMap);
352 // Loop over the PHI nodes in the original block, setting incoming values.
353 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
354 PHINode *PN = OrigPHINode[i];
355 if (CompletelyUnroll) {
356 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
357 Header->getInstList().erase(PN);
359 else if (Count > 1) {
360 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
361 // If this value was defined in the loop, take the value defined by the
362 // last iteration of the loop.
363 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
364 if (L->contains(InValI))
365 InVal = LastValueMap[InVal];
367 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
368 PN->addIncoming(InVal, Latches.back());
372 // Now that all the basic blocks for the unrolled iterations are in place,
373 // set up the branches to connect them.
374 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
375 // The original branch was replicated in each unrolled iteration.
376 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
378 // The branch destination.
379 unsigned j = (i + 1) % e;
380 BasicBlock *Dest = Headers[j];
381 bool NeedConditional = true;
383 if (RuntimeTripCount && j != 0) {
384 NeedConditional = false;
387 // For a complete unroll, make the last iteration end with a branch
388 // to the exit block.
389 if (CompletelyUnroll && j == 0) {
391 NeedConditional = false;
394 // If we know the trip count or a multiple of it, we can safely use an
395 // unconditional branch for some iterations.
396 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
397 NeedConditional = false;
400 if (NeedConditional) {
401 // Update the conditional branch's successor for the following
403 Term->setSuccessor(!ContinueOnTrue, Dest);
405 // Remove phi operands at this loop exit
406 if (Dest != LoopExit) {
407 BasicBlock *BB = Latches[i];
408 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
410 if (*SI == Headers[i])
412 for (BasicBlock::iterator BBI = (*SI)->begin();
413 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
414 Phi->removeIncomingValue(BB, false);
418 // Replace the conditional branch with an unconditional one.
419 BranchInst::Create(Dest, Term);
420 Term->eraseFromParent();
424 // Merge adjacent basic blocks, if possible.
425 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
426 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
427 if (Term->isUnconditional()) {
428 BasicBlock *Dest = Term->getSuccessor(0);
429 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
430 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
434 DominatorTree *DT = nullptr;
436 // FIXME: Reconstruct dom info, because it is not preserved properly.
437 // Incrementally updating domtree after loop unrolling would be easy.
438 if (DominatorTreeWrapperPass *DTWP =
439 PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
440 DT = &DTWP->getDomTree();
441 DT->recalculate(*L->getHeader()->getParent());
444 // Simplify any new induction variables in the partially unrolled loop.
445 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
446 if (SE && !CompletelyUnroll) {
447 SmallVector<WeakVH, 16> DeadInsts;
448 simplifyLoopIVs(L, SE, LPM, DeadInsts);
450 // Aggressively clean up dead instructions that simplifyLoopIVs already
451 // identified. Any remaining should be cleaned up below.
452 while (!DeadInsts.empty())
453 if (Instruction *Inst =
454 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
455 RecursivelyDeleteTriviallyDeadInstructions(Inst);
458 // At this point, the code is well formed. We now do a quick sweep over the
459 // inserted code, doing constant propagation and dead code elimination as we
461 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
462 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
463 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
464 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
465 Instruction *Inst = I++;
467 if (isInstructionTriviallyDead(Inst))
468 (*BB)->getInstList().erase(Inst);
469 else if (Value *V = SimplifyInstruction(Inst))
470 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
471 Inst->replaceAllUsesWith(V);
472 (*BB)->getInstList().erase(Inst);
476 NumCompletelyUnrolled += CompletelyUnroll;
479 Loop *OuterL = L->getParentLoop();
480 // Remove the loop from the LoopPassManager if it's completely removed.
481 if (CompletelyUnroll && LPM != nullptr)
482 LPM->deleteLoopFromQueue(L);
484 // If we have a pass and a DominatorTree we should re-simplify impacted loops
485 // to ensure subsequent analyses can rely on this form. We want to simplify
486 // at least one layer outside of the loop that was unrolled so that any
487 // changes to the parent loop exposed by the unrolling are considered.
489 if (!OuterL && !CompletelyUnroll)
492 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
493 simplifyLoop(OuterL, DT, LI, PP, /*AliasAnalysis*/ nullptr, SE);
495 // LCSSA must be performed on the outermost affected loop. The unrolled
496 // loop's last loop latch is guaranteed to be in the outermost loop after
497 // deleteLoopFromQueue updates LoopInfo.
498 Loop *LatchLoop = LI->getLoopFor(Latches.back());
499 if (!OuterL->contains(LatchLoop))
500 while (OuterL->getParentLoop() != LatchLoop)
501 OuterL = OuterL->getParentLoop();
503 formLCSSARecursively(*OuterL, *DT, SE);