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/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/LoopUtils.h"
33 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
36 #define DEBUG_TYPE "loop-unroll"
38 // TODO: Should these be here or in LoopUnroll?
39 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
40 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
42 /// RemapInstruction - Convert the instruction operands from referencing the
43 /// current values into those specified by VMap.
44 static inline void RemapInstruction(Instruction *I,
45 ValueToValueMapTy &VMap) {
46 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
47 Value *Op = I->getOperand(op);
48 ValueToValueMapTy::iterator It = VMap.find(Op);
50 I->setOperand(op, It->second);
53 if (PHINode *PN = dyn_cast<PHINode>(I)) {
54 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
55 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
57 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
62 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
63 /// only has one predecessor, and that predecessor only has one successor.
64 /// The LoopInfo Analysis that is passed will be kept consistent.
65 /// Returns the new combined block.
66 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI,
68 // Merge basic blocks into their predecessor if there is only one distinct
69 // pred, and if there is only one distinct successor of the predecessor, and
70 // if there are no PHI nodes.
71 BasicBlock *OnlyPred = BB->getSinglePredecessor();
72 if (!OnlyPred) return 0;
74 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
77 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
79 // Resolve any PHI nodes at the start of the block. They are all
80 // guaranteed to have exactly one entry if they exist, unless there are
81 // multiple duplicate (but guaranteed to be equal) entries for the
82 // incoming edges. This occurs when there are multiple edges from
83 // OnlyPred to OnlySucc.
84 FoldSingleEntryPHINodes(BB);
86 // Delete the unconditional branch from the predecessor...
87 OnlyPred->getInstList().pop_back();
89 // Make all PHI nodes that referred to BB now refer to Pred as their
91 BB->replaceAllUsesWith(OnlyPred);
93 // Move all definitions in the successor to the predecessor...
94 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
96 // OldName will be valid until erased.
97 StringRef OldName = BB->getName();
99 // Erase basic block from the function...
101 // ScalarEvolution holds references to loop exit blocks.
103 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
104 if (Loop *L = LI->getLoopFor(BB))
110 // Inherit predecessor's name if it exists...
111 if (!OldName.empty() && !OnlyPred->hasName())
112 OnlyPred->setName(OldName);
114 BB->eraseFromParent();
119 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
120 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
121 /// can only fail when the loop's latch block is not terminated by a conditional
122 /// branch instruction. However, if the trip count (and multiple) are not known,
123 /// loop unrolling will mostly produce more code that is no faster.
125 /// TripCount is generally defined as the number of times the loop header
126 /// executes. UnrollLoop relaxes the definition to permit early exits: here
127 /// TripCount is the iteration on which control exits LatchBlock if no early
128 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
129 /// terminates LatchBlock in order to remove unnecesssary instances of the
130 /// test. In other words, control may exit the loop prior to TripCount
131 /// iterations via an early branch, but control may not exit the loop from the
132 /// LatchBlock's terminator prior to TripCount iterations.
134 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
135 /// execute without exiting the loop.
137 /// The LoopInfo Analysis that is passed will be kept consistent.
139 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
140 /// removed from the LoopPassManager as well. LPM can also be NULL.
142 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
143 /// available from the Pass it must also preserve those analyses.
144 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
145 bool AllowRuntime, unsigned TripMultiple,
146 LoopInfo *LI, Pass *PP, LPPassManager *LPM) {
147 BasicBlock *Preheader = L->getLoopPreheader();
149 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
153 BasicBlock *LatchBlock = L->getLoopLatch();
155 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
159 // Loops with indirectbr cannot be cloned.
160 if (!L->isSafeToClone()) {
161 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
165 BasicBlock *Header = L->getHeader();
166 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
168 if (!BI || BI->isUnconditional()) {
169 // The loop-rotate pass can be helpful to avoid this in many cases.
171 " Can't unroll; loop not terminated by a conditional branch.\n");
175 if (Header->hasAddressTaken()) {
176 // The loop-rotate pass can be helpful to avoid this in many cases.
178 " Won't unroll loop: address of header block is taken.\n");
183 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
184 if (TripMultiple != 1)
185 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
187 // Effectively "DCE" unrolled iterations that are beyond the tripcount
188 // and will never be executed.
189 if (TripCount != 0 && Count > TripCount)
192 // Don't enter the unroll code if there is nothing to do. This way we don't
193 // need to support "partial unrolling by 1".
194 if (TripCount == 0 && Count < 2)
198 assert(TripMultiple > 0);
199 assert(TripCount == 0 || TripCount % TripMultiple == 0);
201 // Are we eliminating the loop control altogether?
202 bool CompletelyUnroll = Count == TripCount;
204 // We assume a run-time trip count if the compiler cannot
205 // figure out the loop trip count and the unroll-runtime
206 // flag is specified.
207 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
209 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
212 // Notify ScalarEvolution that the loop will be substantially changed,
213 // if not outright eliminated.
215 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
220 // If we know the trip count, we know the multiple...
221 unsigned BreakoutTrip = 0;
222 if (TripCount != 0) {
223 BreakoutTrip = TripCount % Count;
226 // Figure out what multiple to use.
227 BreakoutTrip = TripMultiple =
228 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
231 if (CompletelyUnroll) {
232 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
233 << " with trip count " << TripCount << "!\n");
235 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
237 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
238 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
239 } else if (TripMultiple != 1) {
240 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
241 } else if (RuntimeTripCount) {
242 DEBUG(dbgs() << " with run-time trip count");
244 DEBUG(dbgs() << "!\n");
247 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
248 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
250 // For the first iteration of the loop, we should use the precloned values for
251 // PHI nodes. Insert associations now.
252 ValueToValueMapTy LastValueMap;
253 std::vector<PHINode*> OrigPHINode;
254 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
255 OrigPHINode.push_back(cast<PHINode>(I));
258 std::vector<BasicBlock*> Headers;
259 std::vector<BasicBlock*> Latches;
260 Headers.push_back(Header);
261 Latches.push_back(LatchBlock);
263 // The current on-the-fly SSA update requires blocks to be processed in
264 // reverse postorder so that LastValueMap contains the correct value at each
266 LoopBlocksDFS DFS(L);
269 // Stash the DFS iterators before adding blocks to the loop.
270 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
271 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
273 for (unsigned It = 1; It != Count; ++It) {
274 std::vector<BasicBlock*> NewBlocks;
276 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
277 ValueToValueMapTy VMap;
278 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
279 Header->getParent()->getBasicBlockList().push_back(New);
281 // Loop over all of the PHI nodes in the block, changing them to use the
282 // incoming values from the previous block.
284 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
285 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
286 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
287 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
288 if (It > 1 && L->contains(InValI))
289 InVal = LastValueMap[InValI];
290 VMap[OrigPHINode[i]] = InVal;
291 New->getInstList().erase(NewPHI);
294 // Update our running map of newest clones
295 LastValueMap[*BB] = New;
296 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
298 LastValueMap[VI->first] = VI->second;
300 L->addBasicBlockToLoop(New, LI->getBase());
302 // Add phi entries for newly created values to all exit blocks.
303 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
305 if (L->contains(*SI))
307 for (BasicBlock::iterator BBI = (*SI)->begin();
308 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
309 Value *Incoming = phi->getIncomingValueForBlock(*BB);
310 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
311 if (It != LastValueMap.end())
312 Incoming = It->second;
313 phi->addIncoming(Incoming, New);
316 // Keep track of new headers and latches as we create them, so that
317 // we can insert the proper branches later.
319 Headers.push_back(New);
320 if (*BB == LatchBlock)
321 Latches.push_back(New);
323 NewBlocks.push_back(New);
326 // Remap all instructions in the most recent iteration
327 for (unsigned i = 0; i < NewBlocks.size(); ++i)
328 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
329 E = NewBlocks[i]->end(); I != E; ++I)
330 ::RemapInstruction(I, LastValueMap);
333 // Loop over the PHI nodes in the original block, setting incoming values.
334 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
335 PHINode *PN = OrigPHINode[i];
336 if (CompletelyUnroll) {
337 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
338 Header->getInstList().erase(PN);
340 else if (Count > 1) {
341 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
342 // If this value was defined in the loop, take the value defined by the
343 // last iteration of the loop.
344 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
345 if (L->contains(InValI))
346 InVal = LastValueMap[InVal];
348 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
349 PN->addIncoming(InVal, Latches.back());
353 // Now that all the basic blocks for the unrolled iterations are in place,
354 // set up the branches to connect them.
355 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
356 // The original branch was replicated in each unrolled iteration.
357 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
359 // The branch destination.
360 unsigned j = (i + 1) % e;
361 BasicBlock *Dest = Headers[j];
362 bool NeedConditional = true;
364 if (RuntimeTripCount && j != 0) {
365 NeedConditional = false;
368 // For a complete unroll, make the last iteration end with a branch
369 // to the exit block.
370 if (CompletelyUnroll && j == 0) {
372 NeedConditional = false;
375 // If we know the trip count or a multiple of it, we can safely use an
376 // unconditional branch for some iterations.
377 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
378 NeedConditional = false;
381 if (NeedConditional) {
382 // Update the conditional branch's successor for the following
384 Term->setSuccessor(!ContinueOnTrue, Dest);
386 // Remove phi operands at this loop exit
387 if (Dest != LoopExit) {
388 BasicBlock *BB = Latches[i];
389 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
391 if (*SI == Headers[i])
393 for (BasicBlock::iterator BBI = (*SI)->begin();
394 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
395 Phi->removeIncomingValue(BB, false);
399 // Replace the conditional branch with an unconditional one.
400 BranchInst::Create(Dest, Term);
401 Term->eraseFromParent();
405 // Merge adjacent basic blocks, if possible.
406 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
407 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
408 if (Term->isUnconditional()) {
409 BasicBlock *Dest = Term->getSuccessor(0);
410 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
411 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
415 DominatorTree *DT = 0;
417 // FIXME: Reconstruct dom info, because it is not preserved properly.
418 // Incrementally updating domtree after loop unrolling would be easy.
419 if (DominatorTreeWrapperPass *DTWP =
420 PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
421 DT = &DTWP->getDomTree();
422 DT->recalculate(*L->getHeader()->getParent());
425 // Simplify any new induction variables in the partially unrolled loop.
426 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
427 if (SE && !CompletelyUnroll) {
428 SmallVector<WeakVH, 16> DeadInsts;
429 simplifyLoopIVs(L, SE, LPM, DeadInsts);
431 // Aggressively clean up dead instructions that simplifyLoopIVs already
432 // identified. Any remaining should be cleaned up below.
433 while (!DeadInsts.empty())
434 if (Instruction *Inst =
435 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
436 RecursivelyDeleteTriviallyDeadInstructions(Inst);
439 // At this point, the code is well formed. We now do a quick sweep over the
440 // inserted code, doing constant propagation and dead code elimination as we
442 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
443 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
444 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
445 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
446 Instruction *Inst = I++;
448 if (isInstructionTriviallyDead(Inst))
449 (*BB)->getInstList().erase(Inst);
450 else if (Value *V = SimplifyInstruction(Inst))
451 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
452 Inst->replaceAllUsesWith(V);
453 (*BB)->getInstList().erase(Inst);
457 NumCompletelyUnrolled += CompletelyUnroll;
460 Loop *OuterL = L->getParentLoop();
461 // Remove the loop from the LoopPassManager if it's completely removed.
462 if (CompletelyUnroll && LPM != NULL)
463 LPM->deleteLoopFromQueue(L);
465 // If we have a pass and a DominatorTree we should re-simplify impacted loops
466 // to ensure subsequent analyses can rely on this form. We want to simplify
467 // at least one layer outside of the loop that was unrolled so that any
468 // changes to the parent loop exposed by the unrolling are considered.
470 if (!OuterL && !CompletelyUnroll)
473 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
474 simplifyLoop(OuterL, DT, LI, PP, /*AliasAnalysis*/ 0, SE);
475 formLCSSARecursively(*OuterL, *DT, SE);