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/LoopPass.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
29 #include "llvm/Transforms/Utils/Cloning.h"
30 #include "llvm/Transforms/Utils/Local.h"
33 // TODO: Should these be here or in LoopUnroll?
34 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
35 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
37 /// RemapInstruction - Convert the instruction operands from referencing the
38 /// current values into those specified by VMap.
39 static inline void RemapInstruction(Instruction *I,
40 ValueToValueMapTy &VMap) {
41 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
42 Value *Op = I->getOperand(op);
43 ValueToValueMapTy::iterator It = VMap.find(Op);
45 I->setOperand(op, It->second);
48 if (PHINode *PN = dyn_cast<PHINode>(I)) {
49 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
50 ValueToValueMapTy::iterator It = VMap.find(PN->getIncomingBlock(i));
52 PN->setIncomingBlock(i, cast<BasicBlock>(It->second));
57 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
58 /// only has one predecessor, and that predecessor only has one successor.
59 /// The LoopInfo Analysis that is passed will be kept consistent.
60 /// Returns the new combined block.
61 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI) {
62 // Merge basic blocks into their predecessor if there is only one distinct
63 // pred, and if there is only one distinct successor of the predecessor, and
64 // if there are no PHI nodes.
65 BasicBlock *OnlyPred = BB->getSinglePredecessor();
66 if (!OnlyPred) return 0;
68 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
71 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
73 // Resolve any PHI nodes at the start of the block. They are all
74 // guaranteed to have exactly one entry if they exist, unless there are
75 // multiple duplicate (but guaranteed to be equal) entries for the
76 // incoming edges. This occurs when there are multiple edges from
77 // OnlyPred to OnlySucc.
78 FoldSingleEntryPHINodes(BB);
80 // Delete the unconditional branch from the predecessor...
81 OnlyPred->getInstList().pop_back();
83 // Make all PHI nodes that referred to BB now refer to Pred as their
85 BB->replaceAllUsesWith(OnlyPred);
87 // Move all definitions in the successor to the predecessor...
88 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
90 std::string OldName = BB->getName();
92 // Erase basic block from the function...
94 BB->eraseFromParent();
96 // Inherit predecessor's name if it exists...
97 if (!OldName.empty() && !OnlyPred->hasName())
98 OnlyPred->setName(OldName);
103 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
104 /// if unrolling was successful, or false if the loop was unmodified. Unrolling
105 /// can only fail when the loop's latch block is not terminated by a conditional
106 /// branch instruction. However, if the trip count (and multiple) are not known,
107 /// loop unrolling will mostly produce more code that is no faster.
109 /// The LoopInfo Analysis that is passed will be kept consistent.
111 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
112 /// removed from the LoopPassManager as well. LPM can also be NULL.
113 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount,
114 unsigned TripMultiple, LoopInfo *LI, LPPassManager *LPM) {
115 BasicBlock *Preheader = L->getLoopPreheader();
117 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
121 BasicBlock *LatchBlock = L->getLoopLatch();
123 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
127 BasicBlock *Header = L->getHeader();
128 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
130 if (!BI || BI->isUnconditional()) {
131 // The loop-rotate pass can be helpful to avoid this in many cases.
133 " Can't unroll; loop not terminated by a conditional branch.\n");
137 if (Header->hasAddressTaken()) {
138 // The loop-rotate pass can be helpful to avoid this in many cases.
140 " Won't unroll loop: address of header block is taken.\n");
144 // Notify ScalarEvolution that the loop will be substantially changed,
145 // if not outright eliminated.
146 if (ScalarEvolution *SE = LPM->getAnalysisIfAvailable<ScalarEvolution>())
150 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
151 if (TripMultiple != 1)
152 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
154 // Effectively "DCE" unrolled iterations that are beyond the tripcount
155 // and will never be executed.
156 if (TripCount != 0 && Count > TripCount)
160 assert(TripMultiple > 0);
161 assert(TripCount == 0 || TripCount % TripMultiple == 0);
163 // Are we eliminating the loop control altogether?
164 bool CompletelyUnroll = Count == TripCount;
166 // If we know the trip count, we know the multiple...
167 unsigned BreakoutTrip = 0;
168 if (TripCount != 0) {
169 BreakoutTrip = TripCount % Count;
172 // Figure out what multiple to use.
173 BreakoutTrip = TripMultiple =
174 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
177 if (CompletelyUnroll) {
178 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
179 << " with trip count " << TripCount << "!\n");
181 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
183 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
184 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
185 } else if (TripMultiple != 1) {
186 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
188 DEBUG(dbgs() << "!\n");
191 std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
193 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
194 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
196 // For the first iteration of the loop, we should use the precloned values for
197 // PHI nodes. Insert associations now.
198 ValueToValueMapTy LastValueMap;
199 std::vector<PHINode*> OrigPHINode;
200 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
201 PHINode *PN = cast<PHINode>(I);
202 OrigPHINode.push_back(PN);
204 dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
209 std::vector<BasicBlock*> Headers;
210 std::vector<BasicBlock*> Latches;
211 Headers.push_back(Header);
212 Latches.push_back(LatchBlock);
214 for (unsigned It = 1; It != Count; ++It) {
215 std::vector<BasicBlock*> NewBlocks;
217 for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
218 E = LoopBlocks.end(); BB != E; ++BB) {
219 ValueToValueMapTy VMap;
220 BasicBlock *New = CloneBasicBlock(*BB, VMap, "." + Twine(It));
221 Header->getParent()->getBasicBlockList().push_back(New);
223 // Loop over all of the PHI nodes in the block, changing them to use the
224 // incoming values from the previous block.
226 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
227 PHINode *NewPHI = cast<PHINode>(VMap[OrigPHINode[i]]);
228 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
229 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
230 if (It > 1 && L->contains(InValI))
231 InVal = LastValueMap[InValI];
232 VMap[OrigPHINode[i]] = InVal;
233 New->getInstList().erase(NewPHI);
236 // Update our running map of newest clones
237 LastValueMap[*BB] = New;
238 for (ValueToValueMapTy::iterator VI = VMap.begin(), VE = VMap.end();
240 LastValueMap[VI->first] = VI->second;
242 L->addBasicBlockToLoop(New, LI->getBase());
244 // Add phi entries for newly created values to all exit blocks except
245 // the successor of the latch block. The successor of the exit block will
246 // be updated specially after unrolling all the way.
247 if (*BB != LatchBlock)
248 for (succ_iterator SI = succ_begin(*BB), SE = succ_end(*BB); SI != SE;
250 if (!L->contains(*SI))
251 for (BasicBlock::iterator BBI = (*SI)->begin();
252 PHINode *phi = dyn_cast<PHINode>(BBI); ++BBI) {
253 Value *Incoming = phi->getIncomingValueForBlock(*BB);
254 phi->addIncoming(Incoming, New);
257 // Keep track of new headers and latches as we create them, so that
258 // we can insert the proper branches later.
260 Headers.push_back(New);
261 if (*BB == LatchBlock) {
262 Latches.push_back(New);
264 // Also, clear out the new latch's back edge so that it doesn't look
265 // like a new loop, so that it's amenable to being merged with adjacent
267 TerminatorInst *Term = New->getTerminator();
268 assert(L->contains(Term->getSuccessor(!ContinueOnTrue)));
269 assert(Term->getSuccessor(ContinueOnTrue) == LoopExit);
270 Term->setSuccessor(!ContinueOnTrue, NULL);
273 NewBlocks.push_back(New);
276 // Remap all instructions in the most recent iteration
277 for (unsigned i = 0; i < NewBlocks.size(); ++i)
278 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
279 E = NewBlocks[i]->end(); I != E; ++I)
280 ::RemapInstruction(I, LastValueMap);
283 // The latch block exits the loop. If there are any PHI nodes in the
284 // successor blocks, update them to use the appropriate values computed as the
285 // last iteration of the loop.
287 BasicBlock *LastIterationBB = cast<BasicBlock>(LastValueMap[LatchBlock]);
288 for (succ_iterator SI = succ_begin(LatchBlock), SE = succ_end(LatchBlock);
290 for (BasicBlock::iterator BBI = (*SI)->begin();
291 PHINode *PN = dyn_cast<PHINode>(BBI); ++BBI) {
292 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
293 // If this value was defined in the loop, take the value defined by the
294 // last iteration of the loop.
295 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
296 if (L->contains(InValI))
297 InVal = LastValueMap[InVal];
299 PN->addIncoming(InVal, LastIterationBB);
304 // Now, if we're doing complete unrolling, loop over the PHI nodes in the
305 // original block, setting them to their incoming values.
306 if (CompletelyUnroll) {
307 BasicBlock *Preheader = L->getLoopPreheader();
308 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
309 PHINode *PN = OrigPHINode[i];
310 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
311 Header->getInstList().erase(PN);
315 // Now that all the basic blocks for the unrolled iterations are in place,
316 // set up the branches to connect them.
317 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
318 // The original branch was replicated in each unrolled iteration.
319 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
321 // The branch destination.
322 unsigned j = (i + 1) % e;
323 BasicBlock *Dest = Headers[j];
324 bool NeedConditional = true;
326 // For a complete unroll, make the last iteration end with a branch
327 // to the exit block.
328 if (CompletelyUnroll && j == 0) {
330 NeedConditional = false;
333 // If we know the trip count or a multiple of it, we can safely use an
334 // unconditional branch for some iterations.
335 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
336 NeedConditional = false;
339 if (NeedConditional) {
340 // Update the conditional branch's successor for the following
342 Term->setSuccessor(!ContinueOnTrue, Dest);
344 // Replace the conditional branch with an unconditional one.
345 BranchInst::Create(Dest, Term);
346 Term->eraseFromParent();
350 // Merge adjacent basic blocks, if possible.
351 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
352 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
353 if (Term->isUnconditional()) {
354 BasicBlock *Dest = Term->getSuccessor(0);
355 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI))
356 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
360 // At this point, the code is well formed. We now do a quick sweep over the
361 // inserted code, doing constant propagation and dead code elimination as we
363 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
364 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
365 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
366 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
367 Instruction *Inst = I++;
369 if (isInstructionTriviallyDead(Inst))
370 (*BB)->getInstList().erase(Inst);
371 else if (Value *V = SimplifyInstruction(Inst))
372 if (LI->replacementPreservesLCSSAForm(Inst, V)) {
373 Inst->replaceAllUsesWith(V);
374 (*BB)->getInstList().erase(Inst);
378 NumCompletelyUnrolled += CompletelyUnroll;
380 // Remove the loop from the LoopPassManager if it's completely removed.
381 if (CompletelyUnroll && LPM != NULL)
382 LPM->deleteLoopFromQueue(L);