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 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "loop-idiom"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/LoopPass.h"
20 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
21 #include "llvm/Analysis/ScalarEvolutionExpander.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/Transforms/Utils/Local.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/IRBuilder.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/ADT/Statistic.h"
31 // TODO: Recognize "N" size array multiplies: replace with call to blas or
33 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
34 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
37 class LoopIdiomRecognize : public LoopPass {
43 explicit LoopIdiomRecognize() : LoopPass(ID) {
44 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
47 bool runOnLoop(Loop *L, LPPassManager &LPM);
49 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
51 bool processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize,
53 const SCEVAddRecExpr *Ev,
55 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
56 const SCEVAddRecExpr *StoreEv,
57 const SCEVAddRecExpr *LoadEv,
60 /// This transformation requires natural loop information & requires that
61 /// loop preheaders be inserted into the CFG.
63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 AU.addRequired<LoopInfo>();
65 AU.addPreserved<LoopInfo>();
66 AU.addRequiredID(LoopSimplifyID);
67 AU.addPreservedID(LoopSimplifyID);
68 AU.addRequiredID(LCSSAID);
69 AU.addPreservedID(LCSSAID);
70 AU.addRequired<AliasAnalysis>();
71 AU.addPreserved<AliasAnalysis>();
72 AU.addRequired<ScalarEvolution>();
73 AU.addPreserved<ScalarEvolution>();
74 AU.addPreserved<DominatorTree>();
79 char LoopIdiomRecognize::ID = 0;
80 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
82 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
83 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
84 INITIALIZE_PASS_DEPENDENCY(LCSSA)
85 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
86 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
87 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
90 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
92 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
93 /// and zero out all the operands of this instruction. If any of them become
94 /// dead, delete them and the computation tree that feeds them.
96 static void DeleteDeadInstruction(Instruction *I, ScalarEvolution &SE) {
97 SmallVector<Instruction*, 32> NowDeadInsts;
99 NowDeadInsts.push_back(I);
101 // Before we touch this instruction, remove it from SE!
103 Instruction *DeadInst = NowDeadInsts.pop_back_val();
105 // This instruction is dead, zap it, in stages. Start by removing it from
107 SE.forgetValue(DeadInst);
109 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
110 Value *Op = DeadInst->getOperand(op);
111 DeadInst->setOperand(op, 0);
113 // If this operand just became dead, add it to the NowDeadInsts list.
114 if (!Op->use_empty()) continue;
116 if (Instruction *OpI = dyn_cast<Instruction>(Op))
117 if (isInstructionTriviallyDead(OpI))
118 NowDeadInsts.push_back(OpI);
121 DeadInst->eraseFromParent();
123 } while (!NowDeadInsts.empty());
126 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
129 // We only look at trivial single basic block loops.
130 // TODO: eventually support more complex loops, scanning the header.
131 if (L->getBlocks().size() != 1)
134 // The trip count of the loop must be analyzable.
135 SE = &getAnalysis<ScalarEvolution>();
136 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
138 const SCEV *BECount = SE->getBackedgeTakenCount(L);
139 if (isa<SCEVCouldNotCompute>(BECount)) return false;
141 // We require target data for now.
142 TD = getAnalysisIfAvailable<TargetData>();
143 if (TD == 0) return false;
145 BasicBlock *BB = L->getHeader();
146 DEBUG(dbgs() << "loop-idiom Scanning: F[" << BB->getParent()->getName()
147 << "] Loop %" << BB->getName() << "\n");
149 bool MadeChange = false;
150 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
151 // Look for store instructions, which may be memsets.
152 StoreInst *SI = dyn_cast<StoreInst>(I++);
153 if (SI == 0 || SI->isVolatile()) continue;
156 if (!processLoopStore(SI, BECount)) continue;
160 // If processing the store invalidated our iterator, start over from the
169 /// scanBlock - Look over a block to see if we can promote anything out of it.
170 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
171 Value *StoredVal = SI->getValueOperand();
172 Value *StorePtr = SI->getPointerOperand();
174 // Reject stores that are so large that they overflow an unsigned.
175 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
176 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
179 // See if the pointer expression is an AddRec like {base,+,1} on the current
180 // loop, which indicates a strided store. If we have something else, it's a
181 // random store we can't handle.
182 const SCEVAddRecExpr *StoreEv =
183 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
184 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
187 // Check to see if the stride matches the size of the store. If so, then we
188 // know that every byte is touched in the loop.
189 unsigned StoreSize = (unsigned)SizeInBits >> 3;
190 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
192 // TODO: Could also handle negative stride here someday, that will require the
193 // validity check in mayLoopModRefLocation to be updated though.
194 if (Stride == 0 || StoreSize != Stride->getValue()->getValue())
197 // If the stored value is a byte-wise value (like i32 -1), then it may be
198 // turned into a memset of i8 -1, assuming that all the consequtive bytes
199 // are stored. A store of i32 0x01020304 can never be turned into a memset.
200 if (Value *SplatValue = isBytewiseValue(StoredVal))
201 if (processLoopStoreOfSplatValue(SI, StoreSize, SplatValue, StoreEv,
205 // If the stored value is a strided load in the same loop with the same stride
206 // this this may be transformable into a memcpy. This kicks in for stuff like
207 // for (i) A[i] = B[i];
208 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
209 const SCEVAddRecExpr *LoadEv =
210 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
211 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
212 StoreEv->getOperand(1) == LoadEv->getOperand(1) && !LI->isVolatile())
213 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
216 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
221 /// mayLoopModRefLocation - Return true if the specified loop might do a load or
222 /// store to the same location that the specified store could store to, which is
223 /// a loop-strided access.
224 static bool mayLoopModRefLocation(Value *Ptr, Loop *L, const SCEV *BECount,
225 unsigned StoreSize, AliasAnalysis &AA,
226 StoreInst *IgnoredStore) {
227 // Get the location that may be stored across the loop. Since the access is
228 // strided positively through memory, we say that the modified location starts
229 // at the pointer and has infinite size.
230 uint64_t AccessSize = AliasAnalysis::UnknownSize;
232 // If the loop iterates a fixed number of times, we can refine the access size
233 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
234 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
235 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
237 // TODO: For this to be really effective, we have to dive into the pointer
238 // operand in the store. Store to &A[i] of 100 will always return may alias
239 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
240 // which will then no-alias a store to &A[100].
241 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
243 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
245 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
246 if (&*I != IgnoredStore &&
247 AA.getModRefInfo(I, StoreLoc) != AliasAnalysis::NoModRef)
253 /// processLoopStoreOfSplatValue - We see a strided store of a memsetable value.
254 /// If we can transform this into a memset in the loop preheader, do so.
255 bool LoopIdiomRecognize::
256 processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize,
258 const SCEVAddRecExpr *Ev, const SCEV *BECount) {
259 // Verify that the stored value is loop invariant. If not, we can't promote
261 if (!CurLoop->isLoopInvariant(SplatValue))
264 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
265 // this into a memset in the loop preheader now if we want. However, this
266 // would be unsafe to do if there is anything else in the loop that may read
267 // or write to the aliased location. Check for an alias.
268 if (mayLoopModRefLocation(SI->getPointerOperand(), CurLoop, BECount,
269 StoreSize, getAnalysis<AliasAnalysis>(), SI))
272 // Okay, everything looks good, insert the memset.
273 BasicBlock *Preheader = CurLoop->getLoopPreheader();
275 IRBuilder<> Builder(Preheader->getTerminator());
277 // The trip count of the loop and the base pointer of the addrec SCEV is
278 // guaranteed to be loop invariant, which means that it should dominate the
279 // header. Just insert code for it in the preheader.
280 SCEVExpander Expander(*SE);
282 unsigned AddrSpace = SI->getPointerAddressSpace();
284 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
285 Preheader->getTerminator());
287 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
288 // pointer size if it isn't already.
289 const Type *IntPtr = TD->getIntPtrType(SI->getContext());
290 unsigned BESize = SE->getTypeSizeInBits(BECount->getType());
291 if (BESize < TD->getPointerSizeInBits())
292 BECount = SE->getZeroExtendExpr(BECount, IntPtr);
293 else if (BESize > TD->getPointerSizeInBits())
294 BECount = SE->getTruncateExpr(BECount, IntPtr);
296 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
297 true, true /*nooverflow*/);
299 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
300 true, true /*nooverflow*/);
303 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
306 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, SI->getAlignment());
308 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
309 << " from store to: " << *Ev << " at: " << *SI << "\n");
312 // Okay, the memset has been formed. Zap the original store and anything that
314 DeleteDeadInstruction(SI, *SE);
319 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
320 /// same-strided load.
321 bool LoopIdiomRecognize::
322 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
323 const SCEVAddRecExpr *StoreEv,
324 const SCEVAddRecExpr *LoadEv,
325 const SCEV *BECount) {
326 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
328 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
329 // this into a memcmp in the loop preheader now if we want. However, this
330 // would be unsafe to do if there is anything else in the loop that may read
331 // or write to the aliased location (including the load feeding the stores).
332 // Check for an alias.
333 if (mayLoopModRefLocation(SI->getPointerOperand(), CurLoop, BECount,
334 StoreSize, getAnalysis<AliasAnalysis>(), SI))
337 // Okay, everything looks good, insert the memcpy.
338 BasicBlock *Preheader = CurLoop->getLoopPreheader();
340 IRBuilder<> Builder(Preheader->getTerminator());
342 // The trip count of the loop and the base pointer of the addrec SCEV is
343 // guaranteed to be loop invariant, which means that it should dominate the
344 // header. Just insert code for it in the preheader.
345 SCEVExpander Expander(*SE);
348 Expander.expandCodeFor(LoadEv->getStart(),
349 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
350 Preheader->getTerminator());
351 Value *StoreBasePtr =
352 Expander.expandCodeFor(StoreEv->getStart(),
353 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
354 Preheader->getTerminator());
356 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
357 // pointer size if it isn't already.
358 const Type *IntPtr = TD->getIntPtrType(SI->getContext());
359 unsigned BESize = SE->getTypeSizeInBits(BECount->getType());
360 if (BESize < TD->getPointerSizeInBits())
361 BECount = SE->getZeroExtendExpr(BECount, IntPtr);
362 else if (BESize > TD->getPointerSizeInBits())
363 BECount = SE->getTruncateExpr(BECount, IntPtr);
365 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
366 true, true /*nooverflow*/);
368 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
369 true, true /*nooverflow*/);
372 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
375 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
376 std::min(SI->getAlignment(), LI->getAlignment()));
378 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
379 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
380 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
383 // Okay, the memset has been formed. Zap the original store and anything that
385 DeleteDeadInstruction(SI, *SE);