1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
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 pass implements a simple loop unroller. It works best when loops have
11 // been canonicalized by the -indvars pass, allowing it to determine the trip
12 // counts of loops easily.
14 // This pass will multi-block loops only if they contain no non-unrolled
15 // subloops. The process of unrolling can produce extraneous basic blocks
16 // linked with unconditional branches. This will be corrected in the future.
18 //===----------------------------------------------------------------------===//
20 #define DEBUG_TYPE "loop-unroll"
21 #include "llvm/Transforms/Scalar.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Transforms/Utils/Cloning.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/Support/CFG.h"
31 #include "llvm/Support/Compiler.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/STLExtras.h"
37 #include "llvm/ADT/SmallPtrSet.h"
38 #include "llvm/IntrinsicInst.h"
44 STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
45 STATISTIC(NumUnrolled, "Number of loops unrolled (completely or otherwise)");
50 ("unroll-threshold", cl::init(100), cl::Hidden,
51 cl::desc("The cut-off point for automatic loop unrolling"));
55 ("unroll-count", cl::init(0), cl::Hidden,
56 cl::desc("Use this unroll count for all loops, for testing purposes"));
58 class VISIBILITY_HIDDEN LoopUnroll : public LoopPass {
59 LoopInfo *LI; // The current loop information
61 static char ID; // Pass ID, replacement for typeid
62 LoopUnroll() : LoopPass((intptr_t)&ID) {}
64 /// A magic value for use with the Threshold parameter to indicate
65 /// that the loop unroll should be performed regardless of how much
66 /// code expansion would result.
67 static const unsigned NoThreshold = UINT_MAX;
69 bool runOnLoop(Loop *L, LPPassManager &LPM);
70 bool unrollLoop(Loop *L, unsigned Count, unsigned Threshold);
71 BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB);
73 /// This transformation requires natural loop information & requires that
74 /// loop preheaders be inserted into the CFG...
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 AU.addRequiredID(LoopSimplifyID);
78 AU.addRequiredID(LCSSAID);
79 AU.addRequired<LoopInfo>();
80 AU.addPreservedID(LCSSAID);
81 AU.addPreserved<LoopInfo>();
84 char LoopUnroll::ID = 0;
85 RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
88 LoopPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
90 /// ApproximateLoopSize - Approximate the size of the loop.
91 static unsigned ApproximateLoopSize(const Loop *L) {
93 for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
94 BasicBlock *BB = L->getBlocks()[i];
95 Instruction *Term = BB->getTerminator();
96 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
97 if (isa<PHINode>(I) && BB == L->getHeader()) {
98 // Ignore PHI nodes in the header.
99 } else if (I->hasOneUse() && I->use_back() == Term) {
100 // Ignore instructions only used by the loop terminator.
101 } else if (isa<DbgInfoIntrinsic>(I)) {
102 // Ignore debug instructions
103 } else if (isa<CallInst>(I)) {
104 // Estimate size overhead introduced by call instructions which
105 // is higher than other instructions. Here 3 and 10 are magic
106 // numbers that help one isolated test case from PR2067 without
107 // negatively impacting measured benchmarks.
108 if (isa<IntrinsicInst>(I))
116 // TODO: Ignore expressions derived from PHI and constants if inval of phi
117 // is a constant, or if operation is associative. This will get induction
125 // RemapInstruction - Convert the instruction operands from referencing the
126 // current values into those specified by ValueMap.
128 static inline void RemapInstruction(Instruction *I,
129 DenseMap<const Value *, Value*> &ValueMap) {
130 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
131 Value *Op = I->getOperand(op);
132 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
133 if (It != ValueMap.end()) Op = It->second;
134 I->setOperand(op, Op);
138 // FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
139 // only has one predecessor, and that predecessor only has one successor.
140 // Returns the new combined block.
141 BasicBlock *LoopUnroll::FoldBlockIntoPredecessor(BasicBlock *BB) {
142 // Merge basic blocks into their predecessor if there is only one distinct
143 // pred, and if there is only one distinct successor of the predecessor, and
144 // if there are no PHI nodes.
146 BasicBlock *OnlyPred = BB->getSinglePredecessor();
147 if (!OnlyPred) return 0;
149 if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
152 DOUT << "Merging: " << *BB << "into: " << *OnlyPred;
154 // Resolve any PHI nodes at the start of the block. They are all
155 // guaranteed to have exactly one entry if they exist, unless there are
156 // multiple duplicate (but guaranteed to be equal) entries for the
157 // incoming edges. This occurs when there are multiple edges from
158 // OnlyPred to OnlySucc.
160 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
161 PN->replaceAllUsesWith(PN->getIncomingValue(0));
162 BB->getInstList().pop_front(); // Delete the phi node...
165 // Delete the unconditional branch from the predecessor...
166 OnlyPred->getInstList().pop_back();
168 // Move all definitions in the successor to the predecessor...
169 OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
171 // Make all PHI nodes that referred to BB now refer to Pred as their
173 BB->replaceAllUsesWith(OnlyPred);
175 std::string OldName = BB->getName();
177 // Erase basic block from the function...
179 BB->eraseFromParent();
181 // Inherit predecessor's name if it exists...
182 if (BB->hasName() && !OnlyPred->hasName())
183 OnlyPred->takeName(BB);
188 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
189 LI = &getAnalysis<LoopInfo>();
192 if (!unrollLoop(L, UnrollCount, UnrollThreshold))
195 // Update the loop information for this loop.
196 // If we completely unrolled the loop, remove it from the parent.
197 if (L->getNumBackEdges() == 0)
198 LPM.deleteLoopFromQueue(L);
203 /// Unroll the given loop by UnrollCount, or by a heuristically-determined
204 /// value if Count is zero. If Threshold is not NoThreshold, it is a value
205 /// to limit code size expansion. If the loop size would expand beyond the
206 /// threshold value, unrolling is suppressed. The return value is true if
207 /// any transformations are performed.
209 bool LoopUnroll::unrollLoop(Loop *L, unsigned Count, unsigned Threshold) {
210 assert(L->isLCSSAForm());
212 BasicBlock *Header = L->getHeader();
213 BasicBlock *LatchBlock = L->getLoopLatch();
214 BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());
216 DOUT << "Loop Unroll: F[" << Header->getParent()->getName()
217 << "] Loop %" << Header->getName() << "\n";
219 if (!BI || BI->isUnconditional()) {
220 // The loop-rotate pass can be helpful to avoid this in many cases.
221 DOUT << " Can't unroll; loop not terminated by a conditional branch.\n";
225 // Determine the trip count and/or trip multiple. A TripCount value of zero
226 // is used to mean an unknown trip count. The TripMultiple value is the
227 // greatest known integer multiple of the trip count.
228 unsigned TripCount = 0;
229 unsigned TripMultiple = 1;
230 if (Value *TripCountValue = L->getTripCount()) {
231 if (ConstantInt *TripCountC = dyn_cast<ConstantInt>(TripCountValue)) {
232 // Guard against huge trip counts. This also guards against assertions in
233 // APInt from the use of getZExtValue, below.
234 if (TripCountC->getValue().getActiveBits() <= 32) {
235 TripCount = (unsigned)TripCountC->getZExtValue();
237 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TripCountValue)) {
238 switch (BO->getOpcode()) {
239 case BinaryOperator::Mul:
240 if (ConstantInt *MultipleC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
241 if (MultipleC->getValue().getActiveBits() <= 32) {
242 TripMultiple = (unsigned)MultipleC->getZExtValue();
251 DOUT << " Trip Count = " << TripCount << "\n";
252 if (TripMultiple != 1)
253 DOUT << " Trip Multiple = " << TripMultiple << "\n";
255 // Automatically select an unroll count.
257 // Conservative heuristic: if we know the trip count, see if we can
258 // completely unroll (subject to the threshold, checked below); otherwise
260 if (TripCount != 0) {
267 // Effectively "DCE" unrolled iterations that are beyond the tripcount
268 // and will never be executed.
269 if (TripCount != 0 && Count > TripCount)
273 assert(TripMultiple > 0);
274 assert(TripCount == 0 || TripCount % TripMultiple == 0);
276 // Enforce the threshold.
277 if (Threshold != NoThreshold) {
278 unsigned LoopSize = ApproximateLoopSize(L);
279 DOUT << " Loop Size = " << LoopSize << "\n";
280 uint64_t Size = (uint64_t)LoopSize*Count;
281 if (TripCount != 1 && Size > Threshold) {
282 DOUT << " TOO LARGE TO UNROLL: "
283 << Size << ">" << Threshold << "\n";
288 // Are we eliminating the loop control altogether?
289 bool CompletelyUnroll = Count == TripCount;
291 // If we know the trip count, we know the multiple...
292 unsigned BreakoutTrip = 0;
293 if (TripCount != 0) {
294 BreakoutTrip = TripCount % Count;
297 // Figure out what multiple to use.
298 BreakoutTrip = TripMultiple =
299 (unsigned)GreatestCommonDivisor64(Count, TripMultiple);
302 if (CompletelyUnroll) {
303 DOUT << "COMPLETELY UNROLLING loop %" << Header->getName()
304 << " with trip count " << TripCount << "!\n";
306 DOUT << "UNROLLING loop %" << Header->getName()
308 if (TripMultiple == 0 || BreakoutTrip != TripMultiple) {
309 DOUT << " with a breakout at trip " << BreakoutTrip;
310 } else if (TripMultiple != 1) {
311 DOUT << " with " << TripMultiple << " trips per branch";
316 std::vector<BasicBlock*> LoopBlocks = L->getBlocks();
318 bool ContinueOnTrue = L->contains(BI->getSuccessor(0));
319 BasicBlock *LoopExit = BI->getSuccessor(ContinueOnTrue);
321 // For the first iteration of the loop, we should use the precloned values for
322 // PHI nodes. Insert associations now.
323 typedef DenseMap<const Value*, Value*> ValueMapTy;
324 ValueMapTy LastValueMap;
325 std::vector<PHINode*> OrigPHINode;
326 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
327 PHINode *PN = cast<PHINode>(I);
328 OrigPHINode.push_back(PN);
330 dyn_cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)))
331 if (L->contains(I->getParent()))
335 std::vector<BasicBlock*> Headers;
336 std::vector<BasicBlock*> Latches;
337 Headers.push_back(Header);
338 Latches.push_back(LatchBlock);
340 for (unsigned It = 1; It != Count; ++It) {
341 char SuffixBuffer[100];
342 sprintf(SuffixBuffer, ".%d", It);
344 std::vector<BasicBlock*> NewBlocks;
346 for (std::vector<BasicBlock*>::iterator BB = LoopBlocks.begin(),
347 E = LoopBlocks.end(); BB != E; ++BB) {
349 BasicBlock *New = CloneBasicBlock(*BB, ValueMap, SuffixBuffer);
350 Header->getParent()->getBasicBlockList().push_back(New);
352 // Loop over all of the PHI nodes in the block, changing them to use the
353 // incoming values from the previous block.
355 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
356 PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
357 Value *InVal = NewPHI->getIncomingValueForBlock(LatchBlock);
358 if (Instruction *InValI = dyn_cast<Instruction>(InVal))
359 if (It > 1 && L->contains(InValI->getParent()))
360 InVal = LastValueMap[InValI];
361 ValueMap[OrigPHINode[i]] = InVal;
362 New->getInstList().erase(NewPHI);
365 // Update our running map of newest clones
366 LastValueMap[*BB] = New;
367 for (ValueMapTy::iterator VI = ValueMap.begin(), VE = ValueMap.end();
369 LastValueMap[VI->first] = VI->second;
371 L->addBasicBlockToLoop(New, LI->getBase());
373 // Add phi entries for newly created values to all exit blocks except
374 // the successor of the latch block. The successor of the exit block will
375 // be updated specially after unrolling all the way.
376 if (*BB != LatchBlock)
377 for (Value::use_iterator UI = (*BB)->use_begin(), UE = (*BB)->use_end();
379 Instruction *UseInst = cast<Instruction>(*UI);
381 if (isa<PHINode>(UseInst) && !L->contains(UseInst->getParent())) {
382 PHINode *phi = cast<PHINode>(UseInst);
383 Value *Incoming = phi->getIncomingValueForBlock(*BB);
384 phi->addIncoming(Incoming, New);
388 // Keep track of new headers and latches as we create them, so that
389 // we can insert the proper branches later.
391 Headers.push_back(New);
392 if (*BB == LatchBlock) {
393 Latches.push_back(New);
395 // Also, clear out the new latch's back edge so that it doesn't look
396 // like a new loop, so that it's amenable to being merged with adjacent
398 TerminatorInst *Term = New->getTerminator();
399 assert(L->contains(Term->getSuccessor(!ContinueOnTrue)));
400 assert(Term->getSuccessor(ContinueOnTrue) == LoopExit);
401 Term->setSuccessor(!ContinueOnTrue, NULL);
404 NewBlocks.push_back(New);
407 // Remap all instructions in the most recent iteration
408 for (unsigned i = 0; i < NewBlocks.size(); ++i) {
409 BasicBlock *NB = NewBlocks[i];
410 if (BasicBlock *UnwindDest = NB->getUnwindDest())
411 NB->setUnwindDest(cast<BasicBlock>(LastValueMap[UnwindDest]));
413 for (BasicBlock::iterator I = NB->begin(), E = NB->end(); I != E; ++I)
414 RemapInstruction(I, LastValueMap);
418 // The latch block exits the loop. If there are any PHI nodes in the
419 // successor blocks, update them to use the appropriate values computed as the
420 // last iteration of the loop.
422 SmallPtrSet<PHINode*, 8> Users;
423 for (Value::use_iterator UI = LatchBlock->use_begin(),
424 UE = LatchBlock->use_end(); UI != UE; ++UI)
425 if (PHINode *phi = dyn_cast<PHINode>(*UI))
428 BasicBlock *LastIterationBB = cast<BasicBlock>(LastValueMap[LatchBlock]);
429 for (SmallPtrSet<PHINode*,8>::iterator SI = Users.begin(), SE = Users.end();
432 Value *InVal = PN->removeIncomingValue(LatchBlock, false);
433 // If this value was defined in the loop, take the value defined by the
434 // last iteration of the loop.
435 if (Instruction *InValI = dyn_cast<Instruction>(InVal)) {
436 if (L->contains(InValI->getParent()))
437 InVal = LastValueMap[InVal];
439 PN->addIncoming(InVal, LastIterationBB);
443 // Now, if we're doing complete unrolling, loop over the PHI nodes in the
444 // original block, setting them to their incoming values.
445 if (CompletelyUnroll) {
446 BasicBlock *Preheader = L->getLoopPreheader();
447 for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
448 PHINode *PN = OrigPHINode[i];
449 PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
450 Header->getInstList().erase(PN);
454 // Now that all the basic blocks for the unrolled iterations are in place,
455 // set up the branches to connect them.
456 for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
457 // The original branch was replicated in each unrolled iteration.
458 BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
460 // The branch destination.
461 unsigned j = (i + 1) % e;
462 BasicBlock *Dest = Headers[j];
463 bool NeedConditional = true;
465 // For a complete unroll, make the last iteration end with a branch
466 // to the exit block.
467 if (CompletelyUnroll && j == 0) {
469 NeedConditional = false;
472 // If we know the trip count or a multiple of it, we can safely use an
473 // unconditional branch for some iterations.
474 if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
475 NeedConditional = false;
478 if (NeedConditional) {
479 // Update the conditional branch's successor for the following
481 Term->setSuccessor(!ContinueOnTrue, Dest);
483 Term->setUnconditionalDest(Dest);
484 // Merge adjacent basic blocks, if possible.
485 if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest)) {
486 std::replace(Latches.begin(), Latches.end(), Dest, Fold);
487 std::replace(Headers.begin(), Headers.end(), Dest, Fold);
492 // At this point, the code is well formed. We now do a quick sweep over the
493 // inserted code, doing constant propagation and dead code elimination as we
495 const std::vector<BasicBlock*> &NewLoopBlocks = L->getBlocks();
496 for (std::vector<BasicBlock*>::const_iterator BB = NewLoopBlocks.begin(),
497 BBE = NewLoopBlocks.end(); BB != BBE; ++BB)
498 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ) {
499 Instruction *Inst = I++;
501 if (isInstructionTriviallyDead(Inst))
502 (*BB)->getInstList().erase(Inst);
503 else if (Constant *C = ConstantFoldInstruction(Inst)) {
504 Inst->replaceAllUsesWith(C);
505 (*BB)->getInstList().erase(Inst);
509 NumCompletelyUnrolled += CompletelyUnroll;