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/BasicBlock.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LoopIterator.h"
25 #include "llvm/Analysis/LoopPass.h"
26 #include "llvm/Analysis/ScalarEvolution.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 std::string OldName = BB->getName();
95 // Erase basic block from the function...
97 // ScalarEvolution holds references to loop exit blocks.
98 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
99 if (Loop *L = LI->getLoopFor(BB))
103 BB->eraseFromParent();
105 // Inherit predecessor's name if it exists...
106 if (!OldName.empty() && !OnlyPred->hasName())
107 OnlyPred->setName(OldName);
112 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
113 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
114 /// can only fail when the loop's latch block is not terminated by a conditional
115 /// branch instruction. However, if the trip count (and multiple) are not known,
116 /// loop unrolling will mostly produce more code that is no faster.
118 /// TripCount is generally defined as the number of times the loop header
119 /// executes. UnrollLoop relaxes the definition to permit early exits: here
120 /// TripCount is the iteration on which control exits LatchBlock if no early
121 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
122 /// terminates LatchBlock in order to remove unnecesssary instances of the
123 /// test. In other words, control may exit the loop prior to TripCount
124 /// iterations via an early branch, but control may not exit the loop from the
125 /// LatchBlock's terminator prior to TripCount iterations.
127 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
128 /// execute without exiting the loop.
130 /// The LoopInfo Analysis that is passed will be kept consistent.
132 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
133 /// removed from the LoopPassManager as well. LPM can also be NULL.
135 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
136 /// available it must also preserve those analyses.
137 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
138 bool AllowRuntime, unsigned TripMultiple,
139 LoopInfo *LI, LPPassManager *LPM) {
140 BasicBlock *Preheader = L->getLoopPreheader();
142 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
146 BasicBlock *LatchBlock = L->getLoopLatch();
148 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
152 BasicBlock *Header = L->getHeader();
153 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
155 if (!BI || BI->isUnconditional()) {
156 // The loop-rotate pass can be helpful to avoid this in many cases.
158 " Can't unroll; loop not terminated by a conditional branch.\n");
162 if (Header->hasAddressTaken()) {
163 // The loop-rotate pass can be helpful to avoid this in many cases.
165 " Won't unroll loop: address of header block is taken.\n");
170 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
171 if (TripMultiple != 1)
172 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
174 // Effectively "DCE" unrolled iterations that are beyond the tripcount
175 // and will never be executed.
176 if (TripCount != 0 && Count > TripCount)
179 // Don't enter the unroll code if there is nothing to do. This way we don't
180 // need to support "partial unrolling by 1".
181 if (TripCount == 0 && Count < 2)
185 assert(TripMultiple > 0);
186 assert(TripCount == 0 || TripCount % TripMultiple == 0);
188 // Are we eliminating the loop control altogether?
189 bool CompletelyUnroll = Count == TripCount;
191 // We assume a run-time trip count if the compiler cannot
192 // figure out the loop trip count and the unroll-runtime
193 // flag is specified.
194 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
196 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
199 // Notify ScalarEvolution that the loop will be substantially changed,
200 // if not outright eliminated.
201 ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>();
205 // If we know the trip count, we know the multiple...
206 unsigned BreakoutTrip = 0;
207 if (TripCount != 0) {
208 BreakoutTrip = TripCount % Count;
211 // Figure out what multiple to use.
212 BreakoutTrip = TripMultiple =
213 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
216 if (CompletelyUnroll) {
217 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
218 << " with trip count " << TripCount << "!\n");
220 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
222 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
223 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
224 } else if (TripMultiple != 1) {
225 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
226 } else if (RuntimeTripCount) {
227 DEBUG(dbgs() << " with run-time trip count");
229 DEBUG(dbgs() << "!\n");
232 std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
234 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
235 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
237 // For the first iteration of the loop, we should use the precloned values for
238 // PHI nodes. Insert associations now.
239 ValueToValueMapTy LastValueMap;
240 std::vector<PHINode*> OrigPHINode;
241 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
242 OrigPHINode.push_back(cast<PHINode>(I));
245 std::vector<BasicBlock*> Headers;
246 std::vector<BasicBlock*> Latches;
247 Headers.push_back(Header);
248 Latches.push_back(LatchBlock);
250 // The current on-the-fly SSA update requires blocks to be processed in
251 // reverse postorder so that LastValueMap contains the correct value at each
253 LoopBlocksDFS DFS(L);
256 // Stash the DFS iterators before adding blocks to the loop.
257 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
258 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
260 for (unsigned It = 1; It != Count; ++It) {
261 std::vector<BasicBlock*> NewBlocks;
263 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
264 ValueToValueMapTy VMap;
265 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
266 Header->getParent()->getBasicBlockList().push_back(New);
268 // Loop over all of the PHI nodes in the block, changing them to use the
269 // incoming values from the previous block.
271 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
272 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
273 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
274 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
275 if (It > 1 && L->contains(InValI))
276 InVal = LastValueMap[InValI];
277 VMap[OrigPHINode[i]] = InVal;
278 New->getInstList().erase(NewPHI);
281 // Update our running map of newest clones
282 LastValueMap[*BB] = New;
283 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
285 LastValueMap[VI->first] = VI->second;
287 L->addBasicBlockToLoop(New, LI->getBase());
289 // Add phi entries for newly created values to all exit blocks.
290 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
292 if (L->contains(*SI))
294 for (BasicBlock::iterator BBI = (*SI)->begin();
295 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
296 Value *Incoming = phi->getIncomingValueForBlock(*BB);
297 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
298 if (It != LastValueMap.end())
299 Incoming = It->second;
300 phi->addIncoming(Incoming, New);
303 // Keep track of new headers and latches as we create them, so that
304 // we can insert the proper branches later.
306 Headers.push_back(New);
307 if (*BB == LatchBlock)
308 Latches.push_back(New);
310 NewBlocks.push_back(New);
313 // Remap all instructions in the most recent iteration
314 for (unsigned i = 0; i < NewBlocks.size(); ++i)
315 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
316 E = NewBlocks[i]->end(); I != E; ++I)
317 ::RemapInstruction(I, LastValueMap);
320 // Loop over the PHI nodes in the original block, setting incoming values.
321 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
322 PHINode *PN = OrigPHINode[i];
323 if (CompletelyUnroll) {
324 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
325 Header->getInstList().erase(PN);
327 else if (Count > 1) {
328 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
329 // If this value was defined in the loop, take the value defined by the
330 // last iteration of the loop.
331 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
332 if (L->contains(InValI))
333 InVal = LastValueMap[InVal];
335 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
336 PN->addIncoming(InVal, Latches.back());
340 // Now that all the basic blocks for the unrolled iterations are in place,
341 // set up the branches to connect them.
342 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
343 // The original branch was replicated in each unrolled iteration.
344 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
346 // The branch destination.
347 unsigned j = (i + 1) % e;
348 BasicBlock *Dest = Headers[j];
349 bool NeedConditional = true;
351 if (RuntimeTripCount && j != 0) {
352 NeedConditional = false;
355 // For a complete unroll, make the last iteration end with a branch
356 // to the exit block.
357 if (CompletelyUnroll && j == 0) {
359 NeedConditional = false;
362 // If we know the trip count or a multiple of it, we can safely use an
363 // unconditional branch for some iterations.
364 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
365 NeedConditional = false;
368 if (NeedConditional) {
369 // Update the conditional branch's successor for the following
371 Term->setSuccessor(!ContinueOnTrue, Dest);
373 // Remove phi operands at this loop exit
374 if (Dest != LoopExit) {
375 BasicBlock *BB = Latches[i];
376 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
378 if (*SI == Headers[i])
380 for (BasicBlock::iterator BBI = (*SI)->begin();
381 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
382 Phi->removeIncomingValue(BB, false);
386 // Replace the conditional branch with an unconditional one.
387 BranchInst::Create(Dest, Term);
388 Term->eraseFromParent();
392 // Merge adjacent basic blocks, if possible.
393 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
394 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
395 if (Term->isUnconditional()) {
396 BasicBlock *Dest = Term->getSuccessor(0);
397 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
398 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
402 // FIXME: Reconstruct dom info, because it is not preserved properly.
403 // Incrementally updating domtree after loop unrolling would be easy.
404 if (DominatorTree *DT = LPM->getAnalysisIfAvailable<DominatorTree>())
405 DT->runOnFunction(*L->getHeader()->getParent());
407 // Simplify any new induction variables in the partially unrolled loop.
408 if (SE && !CompletelyUnroll) {
409 SmallVector<WeakVH, 16> DeadInsts;
410 simplifyLoopIVs(L, SE, LPM, DeadInsts);
412 // Aggressively clean up dead instructions that simplifyLoopIVs already
413 // identified. Any remaining should be cleaned up below.
414 while (!DeadInsts.empty())
415 if (Instruction *Inst =
416 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
417 RecursivelyDeleteTriviallyDeadInstructions(Inst);
420 // At this point, the code is well formed. We now do a quick sweep over the
421 // inserted code, doing constant propagation and dead code elimination as we
423 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
424 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
425 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
426 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
427 Instruction *Inst = I++;
429 if (isInstructionTriviallyDead(Inst))
430 (*BB)->getInstList().erase(Inst);
431 else if (Value *V = SimplifyInstruction(Inst))
432 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
433 Inst->replaceAllUsesWith(V);
434 (*BB)->getInstList().erase(Inst);
438 NumCompletelyUnrolled += CompletelyUnroll;
440 // Remove the loop from the LoopPassManager if it's completely removed.
441 if (CompletelyUnroll && LPM != NULL)
442 LPM->deleteLoopFromQueue(L);