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 #define DEBUG_TYPE "loop-unroll"
20 #include "llvm/Transforms/Utils/UnrollLoop.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/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Cloning.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
35 // TODO: Should these be here or in LoopUnroll?
36 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
37 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
39 /// RemapInstruction - Convert the instruction operands from referencing the
40 /// current values into those specified by VMap.
41 static inline void RemapInstruction(Instruction *I,
42 ValueToValueMapTy &VMap) {
43 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
44 Value *Op = I->getOperand(op);
45 ValueToValueMapTy::iterator It = VMap.find(Op);
47 I->setOperand(op, It->second);
50 if (PHINode *PN = dyn_cast<PHINode>(I)) {
51 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
52 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
54 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
59 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
60 /// only has one predecessor, and that predecessor only has one successor.
61 /// The LoopInfo Analysis that is passed will be kept consistent.
62 /// Returns the new combined block.
63 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI,
65 // Merge basic blocks into their predecessor if there is only one distinct
66 // pred, and if there is only one distinct successor of the predecessor, and
67 // if there are no PHI nodes.
68 BasicBlock *OnlyPred = BB->getSinglePredecessor();
69 if (!OnlyPred) return 0;
71 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
74 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
76 // Resolve any PHI nodes at the start of the block. They are all
77 // guaranteed to have exactly one entry if they exist, unless there are
78 // multiple duplicate (but guaranteed to be equal) entries for the
79 // incoming edges. This occurs when there are multiple edges from
80 // OnlyPred to OnlySucc.
81 FoldSingleEntryPHINodes(BB);
83 // Delete the unconditional branch from the predecessor...
84 OnlyPred->getInstList().pop_back();
86 // Make all PHI nodes that referred to BB now refer to Pred as their
88 BB->replaceAllUsesWith(OnlyPred);
90 // Move all definitions in the successor to the predecessor...
91 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
93 StringRef OldName = BB->getName();
95 // Erase basic block from the function...
97 // ScalarEvolution holds references to loop exit blocks.
99 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
100 if (Loop *L = LI->getLoopFor(BB))
105 BB->eraseFromParent();
107 // Inherit predecessor's name if it exists...
108 if (!OldName.empty() && !OnlyPred->hasName())
109 OnlyPred->setName(OldName);
114 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
115 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
116 /// can only fail when the loop's latch block is not terminated by a conditional
117 /// branch instruction. However, if the trip count (and multiple) are not known,
118 /// loop unrolling will mostly produce more code that is no faster.
120 /// TripCount is generally defined as the number of times the loop header
121 /// executes. UnrollLoop relaxes the definition to permit early exits: here
122 /// TripCount is the iteration on which control exits LatchBlock if no early
123 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
124 /// terminates LatchBlock in order to remove unnecesssary instances of the
125 /// test. In other words, control may exit the loop prior to TripCount
126 /// iterations via an early branch, but control may not exit the loop from the
127 /// LatchBlock's terminator prior to TripCount iterations.
129 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
130 /// execute without exiting the loop.
132 /// The LoopInfo Analysis that is passed will be kept consistent.
134 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
135 /// removed from the LoopPassManager as well. LPM can also be NULL.
137 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
138 /// available it must also preserve those analyses.
139 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
140 bool AllowRuntime, unsigned TripMultiple,
141 LoopInfo *LI, LPPassManager *LPM) {
142 BasicBlock *Preheader = L->getLoopPreheader();
144 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
148 BasicBlock *LatchBlock = L->getLoopLatch();
150 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
154 // Loops with indirectbr cannot be cloned.
155 if (!L->isSafeToClone()) {
156 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
160 BasicBlock *Header = L->getHeader();
161 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
163 if (!BI || BI->isUnconditional()) {
164 // The loop-rotate pass can be helpful to avoid this in many cases.
166 " Can't unroll; loop not terminated by a conditional branch.\n");
170 if (Header->hasAddressTaken()) {
171 // The loop-rotate pass can be helpful to avoid this in many cases.
173 " Won't unroll loop: address of header block is taken.\n");
178 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
179 if (TripMultiple != 1)
180 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
182 // Effectively "DCE" unrolled iterations that are beyond the tripcount
183 // and will never be executed.
184 if (TripCount != 0 && Count > TripCount)
187 // Don't enter the unroll code if there is nothing to do. This way we don't
188 // need to support "partial unrolling by 1".
189 if (TripCount == 0 && Count < 2)
193 assert(TripMultiple > 0);
194 assert(TripCount == 0 || TripCount % TripMultiple == 0);
196 // Are we eliminating the loop control altogether?
197 bool CompletelyUnroll = Count == TripCount;
199 // We assume a run-time trip count if the compiler cannot
200 // figure out the loop trip count and the unroll-runtime
201 // flag is specified.
202 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
204 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
207 // Notify ScalarEvolution that the loop will be substantially changed,
208 // if not outright eliminated.
210 ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
215 // If we know the trip count, we know the multiple...
216 unsigned BreakoutTrip = 0;
217 if (TripCount != 0) {
218 BreakoutTrip = TripCount % Count;
221 // Figure out what multiple to use.
222 BreakoutTrip = TripMultiple =
223 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
226 if (CompletelyUnroll) {
227 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
228 << " with trip count " << TripCount << "!\n");
230 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
232 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
233 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
234 } else if (TripMultiple != 1) {
235 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
236 } else if (RuntimeTripCount) {
237 DEBUG(dbgs() << " with run-time trip count");
239 DEBUG(dbgs() << "!\n");
242 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
243 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
245 // For the first iteration of the loop, we should use the precloned values for
246 // PHI nodes. Insert associations now.
247 ValueToValueMapTy LastValueMap;
248 std::vector<PHINode*> OrigPHINode;
249 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
250 OrigPHINode.push_back(cast<PHINode>(I));
253 std::vector<BasicBlock*> Headers;
254 std::vector<BasicBlock*> Latches;
255 Headers.push_back(Header);
256 Latches.push_back(LatchBlock);
258 // The current on-the-fly SSA update requires blocks to be processed in
259 // reverse postorder so that LastValueMap contains the correct value at each
261 LoopBlocksDFS DFS(L);
264 // Stash the DFS iterators before adding blocks to the loop.
265 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
266 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
268 for (unsigned It = 1; It != Count; ++It) {
269 std::vector<BasicBlock*> NewBlocks;
271 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
272 ValueToValueMapTy VMap;
273 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
274 Header->getParent()->getBasicBlockList().push_back(New);
276 // Loop over all of the PHI nodes in the block, changing them to use the
277 // incoming values from the previous block.
279 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
280 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
281 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
282 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
283 if (It > 1 && L->contains(InValI))
284 InVal = LastValueMap[InValI];
285 VMap[OrigPHINode[i]] = InVal;
286 New->getInstList().erase(NewPHI);
289 // Update our running map of newest clones
290 LastValueMap[*BB] = New;
291 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
293 LastValueMap[VI->first] = VI->second;
295 L->addBasicBlockToLoop(New, LI->getBase());
297 // Add phi entries for newly created values to all exit blocks.
298 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
300 if (L->contains(*SI))
302 for (BasicBlock::iterator BBI = (*SI)->begin();
303 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
304 Value *Incoming = phi->getIncomingValueForBlock(*BB);
305 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
306 if (It != LastValueMap.end())
307 Incoming = It->second;
308 phi->addIncoming(Incoming, New);
311 // Keep track of new headers and latches as we create them, so that
312 // we can insert the proper branches later.
314 Headers.push_back(New);
315 if (*BB == LatchBlock)
316 Latches.push_back(New);
318 NewBlocks.push_back(New);
321 // Remap all instructions in the most recent iteration
322 for (unsigned i = 0; i < NewBlocks.size(); ++i)
323 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
324 E = NewBlocks[i]->end(); I != E; ++I)
325 ::RemapInstruction(I, LastValueMap);
328 // Loop over the PHI nodes in the original block, setting incoming values.
329 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
330 PHINode *PN = OrigPHINode[i];
331 if (CompletelyUnroll) {
332 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
333 Header->getInstList().erase(PN);
335 else if (Count > 1) {
336 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
337 // If this value was defined in the loop, take the value defined by the
338 // last iteration of the loop.
339 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
340 if (L->contains(InValI))
341 InVal = LastValueMap[InVal];
343 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
344 PN->addIncoming(InVal, Latches.back());
348 // Now that all the basic blocks for the unrolled iterations are in place,
349 // set up the branches to connect them.
350 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
351 // The original branch was replicated in each unrolled iteration.
352 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
354 // The branch destination.
355 unsigned j = (i + 1) % e;
356 BasicBlock *Dest = Headers[j];
357 bool NeedConditional = true;
359 if (RuntimeTripCount && j != 0) {
360 NeedConditional = false;
363 // For a complete unroll, make the last iteration end with a branch
364 // to the exit block.
365 if (CompletelyUnroll && j == 0) {
367 NeedConditional = false;
370 // If we know the trip count or a multiple of it, we can safely use an
371 // unconditional branch for some iterations.
372 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
373 NeedConditional = false;
376 if (NeedConditional) {
377 // Update the conditional branch's successor for the following
379 Term->setSuccessor(!ContinueOnTrue, Dest);
381 // Remove phi operands at this loop exit
382 if (Dest != LoopExit) {
383 BasicBlock *BB = Latches[i];
384 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
386 if (*SI == Headers[i])
388 for (BasicBlock::iterator BBI = (*SI)->begin();
389 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
390 Phi->removeIncomingValue(BB, false);
394 // Replace the conditional branch with an unconditional one.
395 BranchInst::Create(Dest, Term);
396 Term->eraseFromParent();
400 // Merge adjacent basic blocks, if possible.
401 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
402 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
403 if (Term->isUnconditional()) {
404 BasicBlock *Dest = Term->getSuccessor(0);
405 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
406 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
411 // FIXME: Reconstruct dom info, because it is not preserved properly.
412 // Incrementally updating domtree after loop unrolling would be easy.
413 if (DominatorTree *DT = LPM->getAnalysisIfAvailable<DominatorTree>())
414 DT->runOnFunction(*L->getHeader()->getParent());
416 // Simplify any new induction variables in the partially unrolled loop.
417 ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
418 if (SE && !CompletelyUnroll) {
419 SmallVector<WeakVH, 16> DeadInsts;
420 simplifyLoopIVs(L, SE, LPM, DeadInsts);
422 // Aggressively clean up dead instructions that simplifyLoopIVs already
423 // identified. Any remaining should be cleaned up below.
424 while (!DeadInsts.empty())
425 if (Instruction *Inst =
426 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
427 RecursivelyDeleteTriviallyDeadInstructions(Inst);
430 // At this point, the code is well formed. We now do a quick sweep over the
431 // inserted code, doing constant propagation and dead code elimination as we
433 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
434 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
435 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
436 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
437 Instruction *Inst = I++;
439 if (isInstructionTriviallyDead(Inst))
440 (*BB)->getInstList().erase(Inst);
441 else if (Value *V = SimplifyInstruction(Inst))
442 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
443 Inst->replaceAllUsesWith(V);
444 (*BB)->getInstList().erase(Inst);
448 NumCompletelyUnrolled += CompletelyUnroll;
450 // Remove the loop from the LoopPassManager if it's completely removed.
451 if (CompletelyUnroll && LPM != NULL)
452 LPM->deleteLoopFromQueue(L);