While DAE can't modify the function signature of an externally visible function,
[llvm.git] / lib / Transforms / Utils / LoopUnroll.cpp
blobac59b4d7b3e8ec18dfcb9ffe7e86049424db5e0d
1 //===-- UnrollLoop.cpp - Loop unrolling utilities -------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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
12 // unrolling.
14 // It works best when loops have been canonicalized by the -indvars pass,
15 // allowing it to determine the trip counts of loops easily.
17 // The process of unrolling can produce extraneous basic blocks linked with
18 // unconditional branches. This will be corrected in the future.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "loop-unroll"
22 #include "llvm/Transforms/Utils/UnrollLoop.h"
23 #include "llvm/BasicBlock.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/LoopPass.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"
33 using namespace llvm;
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 ValueMap.
41 static inline void RemapInstruction(Instruction *I,
42 DenseMap<const Value *, Value*> &ValueMap) {
43 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
44 Value *Op = I->getOperand(op);
45 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
46 if (It != ValueMap.end())
47 I->setOperand(op, It->second);
51 /// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
52 /// only has one predecessor, and that predecessor only has one successor.
53 /// The LoopInfo Analysis that is passed will be kept consistent.
54 /// Returns the new combined block.
55 static BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI) {
56 // Merge basic blocks into their predecessor if there is only one distinct
57 // pred, and if there is only one distinct successor of the predecessor, and
58 // if there are no PHI nodes.
59 BasicBlock *OnlyPred = BB->getSinglePredecessor();
60 if (!OnlyPred) return 0;
62 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
63 return 0;
65 DEBUG(dbgs() << "Merging: " << *BB << "into: " << *OnlyPred);
67 // Resolve any PHI nodes at the start of the block. They are all
68 // guaranteed to have exactly one entry if they exist, unless there are
69 // multiple duplicate (but guaranteed to be equal) entries for the
70 // incoming edges. This occurs when there are multiple edges from
71 // OnlyPred to OnlySucc.
72 FoldSingleEntryPHINodes(BB);
74 // Delete the unconditional branch from the predecessor...
75 OnlyPred->getInstList().pop_back();
77 // Move all definitions in the successor to the predecessor...
78 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
80 // Make all PHI nodes that referred to BB now refer to Pred as their
81 // source...
82 BB->replaceAllUsesWith(OnlyPred);
84 std::string OldName = BB->getName();
86 // Erase basic block from the function...
87 LI->removeBlock(BB);
88 BB->eraseFromParent();
90 // Inherit predecessor's name if it exists...
91 if (!OldName.empty() && !OnlyPred->hasName())
92 OnlyPred->setName(OldName);
94 return OnlyPred;
97 /// Unroll the given loop by Count. The loop must be in LCSSA form. Returns true
98 /// if unrolling was succesful, or false if the loop was unmodified. Unrolling
99 /// can only fail when the loop's latch block is not terminated by a conditional
100 /// branch instruction. However, if the trip count (and multiple) are not known,
101 /// loop unrolling will mostly produce more code that is no faster.
103 /// The LoopInfo Analysis that is passed will be kept consistent.
105 /// If a LoopPassManager is passed in, and the loop is fully removed, it will be
106 /// removed from the LoopPassManager as well. LPM can also be NULL.
107 bool llvm::UnrollLoop(Loop *L, unsigned Count, LoopInfo* LI, LPPassManager* LPM) {
108 BasicBlock *Preheader = L->getLoopPreheader();
109 if (!Preheader) {
110 DEBUG(dbgs() << " Can't unroll; loop preheader-insertion failed.\n");
111 return false;
114 BasicBlock *LatchBlock = L->getLoopLatch();
115 if (!LatchBlock) {
116 DEBUG(dbgs() << " Can't unroll; loop exit-block-insertion failed.\n");
117 return false;
120 BasicBlock *Header = L->getHeader();
121 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
123 if (!BI || BI->isUnconditional()) {
124 // The loop-rotate pass can be helpful to avoid this in many cases.
125 DEBUG(dbgs() <<
126 " Can't unroll; loop not terminated by a conditional branch.\n");
127 return false;
130 // Find trip count
131 unsigned TripCount = L->getSmallConstantTripCount();
132 // Find trip multiple if count is not available
133 unsigned TripMultiple = 1;
134 if (TripCount == 0)
135 TripMultiple = L->getSmallConstantTripMultiple();
137 if (TripCount != 0)
138 DEBUG(dbgs() << " Trip Count = " << TripCount << "\n");
139 if (TripMultiple != 1)
140 DEBUG(dbgs() << " Trip Multiple = " << TripMultiple << "\n");
142 // Effectively "DCE" unrolled iterations that are beyond the tripcount
143 // and will never be executed.
144 if (TripCount != 0 && Count > TripCount)
145 Count = TripCount;
147 assert(Count > 0);
148 assert(TripMultiple > 0);
149 assert(TripCount == 0 || TripCount % TripMultiple == 0);
151 // Are we eliminating the loop control altogether?
152 bool CompletelyUnroll = Count == TripCount;
154 // If we know the trip count, we know the multiple...
155 unsigned BreakoutTrip = 0;
156 if (TripCount != 0) {
157 BreakoutTrip = TripCount % Count;
158 TripMultiple = 0;
159 } else {
160 // Figure out what multiple to use.
161 BreakoutTrip = TripMultiple =
162 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
165 if (CompletelyUnroll) {
166 DEBUG(dbgs() << "COMPLETELY UNROLLING loop %" << Header->getName()
167 << " with trip count " << TripCount << "!\n");
168 } else {
169 DEBUG(dbgs() << "UNROLLING loop %" << Header->getName()
170 << " by " << Count);
171 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
172 DEBUG(dbgs() << " with a breakout at trip " << BreakoutTrip);
173 } else if (TripMultiple != 1) {
174 DEBUG(dbgs() << " with " << TripMultiple << " trips per branch");
176 DEBUG(dbgs() << "!\n");
179 std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
181 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
182 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
184 // For the first iteration of the loop, we should use the precloned values for
185 // PHI nodes. Insert associations now.
186 typedef DenseMap<const Value*, Value*> ValueMapTy;
187 ValueMapTy LastValueMap;
188 std::vector<PHINode*> OrigPHINode;
189 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
190 PHINode *PN = cast<PHINode>(I);
191 OrigPHINode.push_back(PN);
192 if (Instruction *I =
193 dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
194 if (L->contains(I))
195 LastValueMap[I] = I;
198 std::vector<BasicBlock*> Headers;
199 std::vector<BasicBlock*> Latches;
200 Headers.push_back(Header);
201 Latches.push_back(LatchBlock);
203 for (unsigned It = 1; It != Count; ++It) {
204 std::vector<BasicBlock*> NewBlocks;
206 for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
207 E = LoopBlocks.end(); BB != E; ++BB) {
208 ValueMapTy ValueMap;
209 BasicBlock *New = CloneBasicBlock(*BB, ValueMap, "." + Twine(It));
210 Header->getParent()->getBasicBlockList().push_back(New);
212 // Loop over all of the PHI nodes in the block, changing them to use the
213 // incoming values from the previous block.
214 if (*BB == Header)
215 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
216 PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
217 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
218 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
219 if (It > 1 && L->contains(InValI))
220 InVal = LastValueMap[InValI];
221 ValueMap[OrigPHINode[i]] = InVal;
222 New->getInstList().erase(NewPHI);
225 // Update our running map of newest clones
226 LastValueMap[*BB] = New;
227 for (ValueMapTy::iterator VI = ValueMap.begin(), VE = ValueMap.end();
228 VI != VE; ++VI)
229 LastValueMap[VI->first] = VI->second;
231 L->addBasicBlockToLoop(New, LI->getBase());
233 // Add phi entries for newly created values to all exit blocks except
234 // the successor of the latch block. The successor of the exit block will
235 // be updated specially after unrolling all the way.
236 if (*BB != LatchBlock)
237 for (Value::use_iterator UI = (*BB)->use_begin(), UE = (*BB)->use_end();
238 UI != UE;) {
239 Instruction *UseInst = cast<Instruction>(*UI);
240 ++UI;
241 if (isa<PHINode>(UseInst) && !L->contains(UseInst)) {
242 PHINode *phi = cast<PHINode>(UseInst);
243 Value *Incoming = phi->getIncomingValueForBlock(*BB);
244 phi->addIncoming(Incoming, New);
248 // Keep track of new headers and latches as we create them, so that
249 // we can insert the proper branches later.
250 if (*BB == Header)
251 Headers.push_back(New);
252 if (*BB == LatchBlock) {
253 Latches.push_back(New);
255 // Also, clear out the new latch's back edge so that it doesn't look
256 // like a new loop, so that it's amenable to being merged with adjacent
257 // blocks later on.
258 TerminatorInst *Term = New->getTerminator();
259 assert(L->contains(Term->getSuccessor(!ContinueOnTrue)));
260 assert(Term->getSuccessor(ContinueOnTrue) == LoopExit);
261 Term->setSuccessor(!ContinueOnTrue, NULL);
264 NewBlocks.push_back(New);
267 // Remap all instructions in the most recent iteration
268 for (unsigned i = 0; i < NewBlocks.size(); ++i)
269 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
270 E = NewBlocks[i]->end(); I != E; ++I)
271 RemapInstruction(I, LastValueMap);
274 // The latch block exits the loop. If there are any PHI nodes in the
275 // successor blocks, update them to use the appropriate values computed as the
276 // last iteration of the loop.
277 if (Count != 1) {
278 SmallPtrSet<PHINode*, 8> Users;
279 for (Value::use_iterator UI = LatchBlock->use_begin(),
280 UE = LatchBlock->use_end(); UI != UE; ++UI)
281 if (PHINode *phi = dyn_cast<PHINode>(*UI))
282 Users.insert(phi);
284 BasicBlock *LastIterationBB = cast<BasicBlock>(LastValueMap[LatchBlock]);
285 for (SmallPtrSet<PHINode*,8>::iterator SI = Users.begin(), SE = Users.end();
286 SI != SE; ++SI) {
287 PHINode *PN = *SI;
288 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
289 // If this value was defined in the loop, take the value defined by the
290 // last iteration of the loop.
291 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
292 if (L->contains(InValI))
293 InVal = LastValueMap[InVal];
295 PN->addIncoming(InVal, LastIterationBB);
299 // Now, if we're doing complete unrolling, loop over the PHI nodes in the
300 // original block, setting them to their incoming values.
301 if (CompletelyUnroll) {
302 BasicBlock *Preheader = L->getLoopPreheader();
303 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
304 PHINode *PN = OrigPHINode[i];
305 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
306 Header->getInstList().erase(PN);
310 // Now that all the basic blocks for the unrolled iterations are in place,
311 // set up the branches to connect them.
312 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
313 // The original branch was replicated in each unrolled iteration.
314 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
316 // The branch destination.
317 unsigned j = (i + 1) % e;
318 BasicBlock *Dest = Headers[j];
319 bool NeedConditional = true;
321 // For a complete unroll, make the last iteration end with a branch
322 // to the exit block.
323 if (CompletelyUnroll && j == 0) {
324 Dest = LoopExit;
325 NeedConditional = false;
328 // If we know the trip count or a multiple of it, we can safely use an
329 // unconditional branch for some iterations.
330 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
331 NeedConditional = false;
334 if (NeedConditional) {
335 // Update the conditional branch's successor for the following
336 // iteration.
337 Term->setSuccessor(!ContinueOnTrue, Dest);
338 } else {
339 Term->setUnconditionalDest(Dest);
340 // Merge adjacent basic blocks, if possible.
341 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI)) {
342 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
343 std::replace(Headers.begin(), Headers.end(), Dest, Fold);
348 // At this point, the code is well formed. We now do a quick sweep over the
349 // inserted code, doing constant propagation and dead code elimination as we
350 // go.
351 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
352 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
353 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
354 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
355 Instruction *Inst = I++;
357 if (isInstructionTriviallyDead(Inst))
358 (*BB)->getInstList().erase(Inst);
359 else if (Constant *C = ConstantFoldInstruction(Inst)) {
360 Inst->replaceAllUsesWith(C);
361 (*BB)->getInstList().erase(Inst);
365 NumCompletelyUnrolled += CompletelyUnroll;
366 ++NumUnrolled;
367 // Remove the loop from the LoopPassManager if it's completely removed.
368 if (CompletelyUnroll && LPM != NULL)
369 LPM->deleteLoopFromQueue(L);
371 return true;