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/IR/Dominators.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Cloning.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Transforms/Utils/LoopUtils.h"
34 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
37 // TODO: Should these be here or in LoopUnroll?
38 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
39 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
41 /// RemapInstruction - Convert the instruction operands from referencing the
42 /// current values into those specified by VMap.
43 static inline void RemapInstruction(Instruction *I,
44 ValueToValueMapTy &VMap) {
45 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
46 Value *Op = I->getOperand(op);
47 ValueToValueMapTy::iterator It = VMap.find(Op);
49 I->setOperand(op, It->second);
52 if (PHINode *PN = dyn_cast<PHINode>(I)) {
53 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
54 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
56 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
61 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
62 /// only has one predecessor, and that predecessor only has one successor.
63 /// The LoopInfo Analysis that is passed will be kept consistent.
64 /// Returns the new combined block.
65 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI,
67 // Merge basic blocks into their predecessor if there is only one distinct
68 // pred, and if there is only one distinct successor of the predecessor, and
69 // if there are no PHI nodes.
70 BasicBlock *OnlyPred = BB->getSinglePredecessor();
71 if (!OnlyPred) return 0;
73 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
76 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
78 // Resolve any PHI nodes at the start of the block. They are all
79 // guaranteed to have exactly one entry if they exist, unless there are
80 // multiple duplicate (but guaranteed to be equal) entries for the
81 // incoming edges. This occurs when there are multiple edges from
82 // OnlyPred to OnlySucc.
83 FoldSingleEntryPHINodes(BB);
85 // Delete the unconditional branch from the predecessor...
86 OnlyPred->getInstList().pop_back();
88 // Make all PHI nodes that referred to BB now refer to Pred as their
90 BB->replaceAllUsesWith(OnlyPred);
92 // Move all definitions in the successor to the predecessor...
93 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
95 // OldName will be valid until erased.
96 StringRef OldName = BB->getName();
98 // Erase basic block from the function...
100 // ScalarEvolution holds references to loop exit blocks.
102 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>()) {
103 if (Loop *L = LI->getLoopFor(BB))
109 // Inherit predecessor's name if it exists...
110 if (!OldName.empty() && !OnlyPred->hasName())
111 OnlyPred->setName(OldName);
113 BB->eraseFromParent();
118 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
119 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
120 /// can only fail when the loop's latch block is not terminated by a conditional
121 /// branch instruction. However, if the trip count (and multiple) are not known,
122 /// loop unrolling will mostly produce more code that is no faster.
124 /// TripCount is generally defined as the number of times the loop header
125 /// executes. UnrollLoop relaxes the definition to permit early exits: here
126 /// TripCount is the iteration on which control exits LatchBlock if no early
127 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
128 /// terminates LatchBlock in order to remove unnecesssary instances of the
129 /// test. In other words, control may exit the loop prior to TripCount
130 /// iterations via an early branch, but control may not exit the loop from the
131 /// LatchBlock's terminator prior to TripCount iterations.
133 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
134 /// execute without exiting the loop.
136 /// The LoopInfo Analysis that is passed will be kept consistent.
138 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
139 /// removed from the LoopPassManager as well. LPM can also be NULL.
141 /// This utility preserves LoopInfo. If DominatorTree or ScalarEvolution are
142 /// available from the Pass it must also preserve those analyses.
143 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
144 bool AllowRuntime, unsigned TripMultiple,
145 LoopInfo *LI, Pass *PP, LPPassManager *LPM) {
146 BasicBlock *Preheader = L->getLoopPreheader();
148 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
152 BasicBlock *LatchBlock = L->getLoopLatch();
154 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
158 // Loops with indirectbr cannot be cloned.
159 if (!L->isSafeToClone()) {
160 DEBUG(dbgs() << " Can't unroll; Loop body cannot be cloned.\n");
164 BasicBlock *Header = L->getHeader();
165 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
167 if (!BI || BI->isUnconditional()) {
168 // The loop-rotate pass can be helpful to avoid this in many cases.
170 " Can't unroll; loop not terminated by a conditional branch.\n");
174 if (Header->hasAddressTaken()) {
175 // The loop-rotate pass can be helpful to avoid this in many cases.
177 " Won't unroll loop: address of header block is taken.\n");
182 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
183 if (TripMultiple != 1)
184 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
186 // Effectively "DCE" unrolled iterations that are beyond the tripcount
187 // and will never be executed.
188 if (TripCount != 0 && Count > TripCount)
191 // Don't enter the unroll code if there is nothing to do. This way we don't
192 // need to support "partial unrolling by 1".
193 if (TripCount == 0 && Count < 2)
197 assert(TripMultiple > 0);
198 assert(TripCount == 0 || TripCount % TripMultiple == 0);
200 // Are we eliminating the loop control altogether?
201 bool CompletelyUnroll = Count == TripCount;
203 // We assume a run-time trip count if the compiler cannot
204 // figure out the loop trip count and the unroll-runtime
205 // flag is specified.
206 bool RuntimeTripCount = (TripCount == 0 && Count > 0 && AllowRuntime);
208 if (RuntimeTripCount && !UnrollRuntimeLoopProlog(L, Count, LI, LPM))
211 // Notify ScalarEvolution that the loop will be substantially changed,
212 // if not outright eliminated.
214 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
219 // If we know the trip count, we know the multiple...
220 unsigned BreakoutTrip = 0;
221 if (TripCount != 0) {
222 BreakoutTrip = TripCount % Count;
225 // Figure out what multiple to use.
226 BreakoutTrip = TripMultiple =
227 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
230 if (CompletelyUnroll) {
231 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
232 << " with trip count " << TripCount << "!\n");
234 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
236 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
237 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
238 } else if (TripMultiple != 1) {
239 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
240 } else if (RuntimeTripCount) {
241 DEBUG(dbgs() << " with run-time trip count");
243 DEBUG(dbgs() << "!\n");
246 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
247 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
249 // For the first iteration of the loop, we should use the precloned values for
250 // PHI nodes. Insert associations now.
251 ValueToValueMapTy LastValueMap;
252 std::vector<PHINode*> OrigPHINode;
253 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
254 OrigPHINode.push_back(cast<PHINode>(I));
257 std::vector<BasicBlock*> Headers;
258 std::vector<BasicBlock*> Latches;
259 Headers.push_back(Header);
260 Latches.push_back(LatchBlock);
262 // The current on-the-fly SSA update requires blocks to be processed in
263 // reverse postorder so that LastValueMap contains the correct value at each
265 LoopBlocksDFS DFS(L);
268 // Stash the DFS iterators before adding blocks to the loop.
269 LoopBlocksDFS::RPOIterator BlockBegin = DFS.beginRPO();
270 LoopBlocksDFS::RPOIterator BlockEnd = DFS.endRPO();
272 for (unsigned It = 1; It != Count; ++It) {
273 std::vector<BasicBlock*> NewBlocks;
275 for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
276 ValueToValueMapTy VMap;
277 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
278 Header->getParent()->getBasicBlockList().push_back(New);
280 // Loop over all of the PHI nodes in the block, changing them to use the
281 // incoming values from the previous block.
283 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
284 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
285 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
286 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
287 if (It > 1 && L->contains(InValI))
288 InVal = LastValueMap[InValI];
289 VMap[OrigPHINode[i]] = InVal;
290 New->getInstList().erase(NewPHI);
293 // Update our running map of newest clones
294 LastValueMap[*BB] = New;
295 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
297 LastValueMap[VI->first] = VI->second;
299 L->addBasicBlockToLoop(New, LI->getBase());
301 // Add phi entries for newly created values to all exit blocks.
302 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB);
304 if (L->contains(*SI))
306 for (BasicBlock::iterator BBI = (*SI)->begin();
307 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
308 Value *Incoming = phi->getIncomingValueForBlock(*BB);
309 ValueToValueMapTy::iterator It = LastValueMap.find(Incoming);
310 if (It != LastValueMap.end())
311 Incoming = It->second;
312 phi->addIncoming(Incoming, New);
315 // Keep track of new headers and latches as we create them, so that
316 // we can insert the proper branches later.
318 Headers.push_back(New);
319 if (*BB == LatchBlock)
320 Latches.push_back(New);
322 NewBlocks.push_back(New);
325 // Remap all instructions in the most recent iteration
326 for (unsigned i = 0; i < NewBlocks.size(); ++i)
327 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
328 E = NewBlocks[i]->end(); I != E; ++I)
329 ::RemapInstruction(I, LastValueMap);
332 // Loop over the PHI nodes in the original block, setting incoming values.
333 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
334 PHINode *PN = OrigPHINode[i];
335 if (CompletelyUnroll) {
336 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
337 Header->getInstList().erase(PN);
339 else if (Count > 1) {
340 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
341 // If this value was defined in the loop, take the value defined by the
342 // last iteration of the loop.
343 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
344 if (L->contains(InValI))
345 InVal = LastValueMap[InVal];
347 assert(Latches.back() == LastValueMap[LatchBlock] && "bad last latch");
348 PN->addIncoming(InVal, Latches.back());
352 // Now that all the basic blocks for the unrolled iterations are in place,
353 // set up the branches to connect them.
354 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
355 // The original branch was replicated in each unrolled iteration.
356 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
358 // The branch destination.
359 unsigned j = (i + 1) % e;
360 BasicBlock *Dest = Headers[j];
361 bool NeedConditional = true;
363 if (RuntimeTripCount && j != 0) {
364 NeedConditional = false;
367 // For a complete unroll, make the last iteration end with a branch
368 // to the exit block.
369 if (CompletelyUnroll && j == 0) {
371 NeedConditional = false;
374 // If we know the trip count or a multiple of it, we can safely use an
375 // unconditional branch for some iterations.
376 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
377 NeedConditional = false;
380 if (NeedConditional) {
381 // Update the conditional branch's successor for the following
383 Term->setSuccessor(!ContinueOnTrue, Dest);
385 // Remove phi operands at this loop exit
386 if (Dest != LoopExit) {
387 BasicBlock *BB = Latches[i];
388 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
390 if (*SI == Headers[i])
392 for (BasicBlock::iterator BBI = (*SI)->begin();
393 PHINode *Phi = dyn_cast<PHINode>(BBI); ++BBI) {
394 Phi->removeIncomingValue(BB, false);
398 // Replace the conditional branch with an unconditional one.
399 BranchInst::Create(Dest, Term);
400 Term->eraseFromParent();
404 // Merge adjacent basic blocks, if possible.
405 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
406 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
407 if (Term->isUnconditional()) {
408 BasicBlock *Dest = Term->getSuccessor(0);
409 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, LPM))
410 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
414 DominatorTree *DT = 0;
416 // FIXME: Reconstruct dom info, because it is not preserved properly.
417 // Incrementally updating domtree after loop unrolling would be easy.
418 if (DominatorTreeWrapperPass *DTWP =
419 PP->getAnalysisIfAvailable<DominatorTreeWrapperPass>()) {
420 DT = &DTWP->getDomTree();
421 DT->recalculate(*L->getHeader()->getParent());
424 // Simplify any new induction variables in the partially unrolled loop.
425 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
426 if (SE && !CompletelyUnroll) {
427 SmallVector<WeakVH, 16> DeadInsts;
428 simplifyLoopIVs(L, SE, LPM, DeadInsts);
430 // Aggressively clean up dead instructions that simplifyLoopIVs already
431 // identified. Any remaining should be cleaned up below.
432 while (!DeadInsts.empty())
433 if (Instruction *Inst =
434 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
435 RecursivelyDeleteTriviallyDeadInstructions(Inst);
438 // At this point, the code is well formed. We now do a quick sweep over the
439 // inserted code, doing constant propagation and dead code elimination as we
441 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
442 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
443 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
444 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
445 Instruction *Inst = I++;
447 if (isInstructionTriviallyDead(Inst))
448 (*BB)->getInstList().erase(Inst);
449 else if (Value *V = SimplifyInstruction(Inst))
450 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
451 Inst->replaceAllUsesWith(V);
452 (*BB)->getInstList().erase(Inst);
456 NumCompletelyUnrolled += CompletelyUnroll;
459 Loop *OuterL = L->getParentLoop();
460 // Remove the loop from the LoopPassManager if it's completely removed.
461 if (CompletelyUnroll && LPM != NULL)
462 LPM->deleteLoopFromQueue(L);
464 // If we have a pass and a DominatorTree we should re-simplify impacted loops
465 // to ensure subsequent analyses can rely on this form. We want to simplify
466 // at least one layer outside of the loop that was unrolled so that any
467 // changes to the parent loop exposed by the unrolling are considered.
469 if (!OuterL && !CompletelyUnroll)
472 ScalarEvolution *SE = PP->getAnalysisIfAvailable<ScalarEvolution>();
473 simplifyLoop(OuterL, DT, LI, PP, /*AliasAnalysis*/ 0, SE);
474 formLCSSARecursively(*OuterL, *DT, SE);