1 //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
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 an idiom recognizer that transforms simple loops into a
11 // non-loop form. In cases that this kicks in, it can be a significant
14 //===----------------------------------------------------------------------===//
18 // Future loop memory idioms to recognize:
19 // memcmp, memmove, strlen, etc.
20 // Future floating point idioms to recognize in -ffast-math mode:
22 // Future integer operation idioms to recognize:
25 // Beware that isel's default lowering for ctpop is highly inefficient for
26 // i64 and larger types when i64 is legal and the value has few bits set. It
27 // would be good to enhance isel to emit a loop for ctpop in this case.
29 // We should enhance the memset/memcpy recognition to handle multiple stores in
30 // the loop. This would handle things like:
31 // void foo(_Complex float *P)
32 // for (i) { __real__(*P) = 0; __imag__(*P) = 0; }
33 // this is also "Example 2" from http://blog.regehr.org/archives/320
35 // This could recognize common matrix multiplies and dot product idioms and
36 // replace them with calls to BLAS (if linked in??).
38 //===----------------------------------------------------------------------===//
40 #define DEBUG_TYPE "loop-idiom"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/IntrinsicInst.h"
43 #include "llvm/Analysis/AliasAnalysis.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
46 #include "llvm/Analysis/ScalarEvolutionExpander.h"
47 #include "llvm/Analysis/ValueTracking.h"
48 #include "llvm/Target/TargetData.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/IRBuilder.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/ADT/Statistic.h"
56 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
57 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
60 class LoopIdiomRecognize : public LoopPass {
67 explicit LoopIdiomRecognize() : LoopPass(ID) {
68 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
71 bool runOnLoop(Loop *L, LPPassManager &LPM);
72 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
73 SmallVectorImpl<BasicBlock*> &ExitBlocks);
75 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
76 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
78 bool processLoopStoreOfSplatValue(Value *DestPtr, unsigned StoreSize,
79 unsigned StoreAlignment,
80 Value *SplatValue, Instruction *TheStore,
81 const SCEVAddRecExpr *Ev,
83 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
84 const SCEVAddRecExpr *StoreEv,
85 const SCEVAddRecExpr *LoadEv,
88 /// This transformation requires natural loop information & requires that
89 /// loop preheaders be inserted into the CFG.
91 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
92 AU.addRequired<LoopInfo>();
93 AU.addPreserved<LoopInfo>();
94 AU.addRequiredID(LoopSimplifyID);
95 AU.addPreservedID(LoopSimplifyID);
96 AU.addRequiredID(LCSSAID);
97 AU.addPreservedID(LCSSAID);
98 AU.addRequired<AliasAnalysis>();
99 AU.addPreserved<AliasAnalysis>();
100 AU.addRequired<ScalarEvolution>();
101 AU.addPreserved<ScalarEvolution>();
102 AU.addPreserved<DominatorTree>();
103 AU.addRequired<DominatorTree>();
108 char LoopIdiomRecognize::ID = 0;
109 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
111 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
112 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
113 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
114 INITIALIZE_PASS_DEPENDENCY(LCSSA)
115 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
116 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
117 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
120 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
122 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
123 /// and zero out all the operands of this instruction. If any of them become
124 /// dead, delete them and the computation tree that feeds them.
126 static void DeleteDeadInstruction(Instruction *I, ScalarEvolution &SE) {
127 SmallVector<Instruction*, 32> NowDeadInsts;
129 NowDeadInsts.push_back(I);
131 // Before we touch this instruction, remove it from SE!
133 Instruction *DeadInst = NowDeadInsts.pop_back_val();
135 // This instruction is dead, zap it, in stages. Start by removing it from
137 SE.forgetValue(DeadInst);
139 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
140 Value *Op = DeadInst->getOperand(op);
141 DeadInst->setOperand(op, 0);
143 // If this operand just became dead, add it to the NowDeadInsts list.
144 if (!Op->use_empty()) continue;
146 if (Instruction *OpI = dyn_cast<Instruction>(Op))
147 if (isInstructionTriviallyDead(OpI))
148 NowDeadInsts.push_back(OpI);
151 DeadInst->eraseFromParent();
153 } while (!NowDeadInsts.empty());
156 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
159 // The trip count of the loop must be analyzable.
160 SE = &getAnalysis<ScalarEvolution>();
161 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
163 const SCEV *BECount = SE->getBackedgeTakenCount(L);
164 if (isa<SCEVCouldNotCompute>(BECount)) return false;
166 // If this loop executes exactly one time, then it should be peeled, not
167 // optimized by this pass.
168 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
169 if (BECst->getValue()->getValue() == 0)
172 // We require target data for now.
173 TD = getAnalysisIfAvailable<TargetData>();
174 if (TD == 0) return false;
176 DT = &getAnalysis<DominatorTree>();
177 LoopInfo &LI = getAnalysis<LoopInfo>();
179 SmallVector<BasicBlock*, 8> ExitBlocks;
180 CurLoop->getUniqueExitBlocks(ExitBlocks);
182 DEBUG(dbgs() << "loop-idiom Scanning: F["
183 << L->getHeader()->getParent()->getName()
184 << "] Loop %" << L->getHeader()->getName() << "\n");
186 bool MadeChange = false;
187 // Scan all the blocks in the loop that are not in subloops.
188 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
190 // Ignore blocks in subloops.
191 if (LI.getLoopFor(*BI) != CurLoop)
194 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
199 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
200 /// with the specified backedge count. This block is known to be in the current
201 /// loop and not in any subloops.
202 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
203 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
204 // We can only promote stores in this block if they are unconditionally
205 // executed in the loop. For a block to be unconditionally executed, it has
206 // to dominate all the exit blocks of the loop. Verify this now.
207 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
208 if (!DT->dominates(BB, ExitBlocks[i]))
211 bool MadeChange = false;
212 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
213 Instruction *Inst = I++;
214 // Look for store instructions, which may be optimized to memset/memcpy.
215 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
217 if (!processLoopStore(SI, BECount)) continue;
220 // If processing the store invalidated our iterator, start over from the
227 // Look for memset instructions, which may be optimized to a larger memset.
228 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
230 if (!processLoopMemSet(MSI, BECount)) continue;
233 // If processing the memset invalidated our iterator, start over from the
245 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
246 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
247 if (SI->isVolatile()) return false;
249 Value *StoredVal = SI->getValueOperand();
250 Value *StorePtr = SI->getPointerOperand();
252 // Reject stores that are so large that they overflow an unsigned.
253 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
254 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
257 // See if the pointer expression is an AddRec like {base,+,1} on the current
258 // loop, which indicates a strided store. If we have something else, it's a
259 // random store we can't handle.
260 const SCEVAddRecExpr *StoreEv =
261 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
262 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
265 // Check to see if the stride matches the size of the store. If so, then we
266 // know that every byte is touched in the loop.
267 unsigned StoreSize = (unsigned)SizeInBits >> 3;
268 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
270 // TODO: Could also handle negative stride here someday, that will require the
271 // validity check in mayLoopAccessLocation to be updated though.
272 if (Stride == 0 || StoreSize != Stride->getValue()->getValue())
275 // If the stored value is a byte-wise value (like i32 -1), then it may be
276 // turned into a memset of i8 -1, assuming that all the consecutive bytes
277 // are stored. A store of i32 0x01020304 can never be turned into a memset.
278 if (Value *SplatValue = isBytewiseValue(StoredVal))
279 if (processLoopStoreOfSplatValue(StorePtr, StoreSize, SI->getAlignment(),
280 SplatValue, SI, StoreEv, BECount))
283 // If the stored value is a strided load in the same loop with the same stride
284 // this this may be transformable into a memcpy. This kicks in for stuff like
285 // for (i) A[i] = B[i];
286 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
287 const SCEVAddRecExpr *LoadEv =
288 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
289 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
290 StoreEv->getOperand(1) == LoadEv->getOperand(1) && !LI->isVolatile())
291 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
294 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
299 /// processLoopMemSet - See if this memset can be promoted to a large memset.
300 bool LoopIdiomRecognize::
301 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
302 // We can only handle non-volatile memsets with a constant size.
303 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
305 Value *Pointer = MSI->getDest();
307 // See if the pointer expression is an AddRec like {base,+,1} on the current
308 // loop, which indicates a strided store. If we have something else, it's a
309 // random store we can't handle.
310 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
311 if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine())
314 // Reject memsets that are so large that they overflow an unsigned.
315 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
316 if ((SizeInBytes >> 32) != 0)
319 // Check to see if the stride matches the size of the memset. If so, then we
320 // know that every byte is touched in the loop.
321 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
323 // TODO: Could also handle negative stride here someday, that will require the
324 // validity check in mayLoopAccessLocation to be updated though.
325 if (Stride == 0 || MSI->getLength() != Stride->getValue())
328 return processLoopStoreOfSplatValue(Pointer, (unsigned)SizeInBytes,
329 MSI->getAlignment(), MSI->getValue(),
334 /// mayLoopAccessLocation - Return true if the specified loop might access the
335 /// specified pointer location, which is a loop-strided access. The 'Access'
336 /// argument specifies what the verboten forms of access are (read or write).
337 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
338 Loop *L, const SCEV *BECount,
339 unsigned StoreSize, AliasAnalysis &AA,
340 Instruction *IgnoredStore) {
341 // Get the location that may be stored across the loop. Since the access is
342 // strided positively through memory, we say that the modified location starts
343 // at the pointer and has infinite size.
344 uint64_t AccessSize = AliasAnalysis::UnknownSize;
346 // If the loop iterates a fixed number of times, we can refine the access size
347 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
348 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
349 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
351 // TODO: For this to be really effective, we have to dive into the pointer
352 // operand in the store. Store to &A[i] of 100 will always return may alias
353 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
354 // which will then no-alias a store to &A[100].
355 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
357 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
359 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
360 if (&*I != IgnoredStore &&
361 (AA.getModRefInfo(I, StoreLoc) & Access))
367 /// processLoopStoreOfSplatValue - We see a strided store of a memsetable value.
368 /// If we can transform this into a memset in the loop preheader, do so.
369 bool LoopIdiomRecognize::
370 processLoopStoreOfSplatValue(Value *DestPtr, unsigned StoreSize,
371 unsigned StoreAlignment, Value *SplatValue,
372 Instruction *TheStore,
373 const SCEVAddRecExpr *Ev, const SCEV *BECount) {
374 // Verify that the stored value is loop invariant. If not, we can't promote
376 if (!CurLoop->isLoopInvariant(SplatValue))
379 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
380 // this into a memset in the loop preheader now if we want. However, this
381 // would be unsafe to do if there is anything else in the loop that may read
382 // or write to the aliased location. Check for an alias.
383 if (mayLoopAccessLocation(DestPtr, AliasAnalysis::ModRef,
385 StoreSize, getAnalysis<AliasAnalysis>(), TheStore))
388 // Okay, everything looks good, insert the memset.
389 BasicBlock *Preheader = CurLoop->getLoopPreheader();
391 IRBuilder<> Builder(Preheader->getTerminator());
393 // The trip count of the loop and the base pointer of the addrec SCEV is
394 // guaranteed to be loop invariant, which means that it should dominate the
395 // header. Just insert code for it in the preheader.
396 SCEVExpander Expander(*SE);
398 unsigned AddrSpace = cast<PointerType>(DestPtr->getType())->getAddressSpace();
400 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
401 Preheader->getTerminator());
403 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
404 // pointer size if it isn't already.
405 const Type *IntPtr = TD->getIntPtrType(SplatValue->getContext());
406 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
408 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
409 true /*no unsigned overflow*/);
411 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
412 true /*no unsigned overflow*/);
415 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
418 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
420 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
421 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
424 // Okay, the memset has been formed. Zap the original store and anything that
426 DeleteDeadInstruction(TheStore, *SE);
431 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
432 /// same-strided load.
433 bool LoopIdiomRecognize::
434 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
435 const SCEVAddRecExpr *StoreEv,
436 const SCEVAddRecExpr *LoadEv,
437 const SCEV *BECount) {
438 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
440 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
441 // this into a memcpy in the loop preheader now if we want. However, this
442 // would be unsafe to do if there is anything else in the loop that may read
443 // or write to the stored location (including the load feeding the stores).
444 // Check for an alias.
445 if (mayLoopAccessLocation(SI->getPointerOperand(), AliasAnalysis::ModRef,
446 CurLoop, BECount, StoreSize,
447 getAnalysis<AliasAnalysis>(), SI))
450 // For a memcpy, we have to make sure that the input array is not being
451 // mutated by the loop.
452 if (mayLoopAccessLocation(LI->getPointerOperand(), AliasAnalysis::Mod,
453 CurLoop, BECount, StoreSize,
454 getAnalysis<AliasAnalysis>(), SI))
457 // Okay, everything looks good, insert the memcpy.
458 BasicBlock *Preheader = CurLoop->getLoopPreheader();
460 IRBuilder<> Builder(Preheader->getTerminator());
462 // The trip count of the loop and the base pointer of the addrec SCEV is
463 // guaranteed to be loop invariant, which means that it should dominate the
464 // header. Just insert code for it in the preheader.
465 SCEVExpander Expander(*SE);
468 Expander.expandCodeFor(LoadEv->getStart(),
469 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
470 Preheader->getTerminator());
471 Value *StoreBasePtr =
472 Expander.expandCodeFor(StoreEv->getStart(),
473 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
474 Preheader->getTerminator());
476 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
477 // pointer size if it isn't already.
478 const Type *IntPtr = TD->getIntPtrType(SI->getContext());
479 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
481 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
482 true /*no unsigned overflow*/);
484 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
485 true /*no unsigned overflow*/);
488 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
491 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
492 std::min(SI->getAlignment(), LI->getAlignment()));
494 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
495 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
496 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
499 // Okay, the memset has been formed. Zap the original store and anything that
501 DeleteDeadInstruction(SI, *SE);