1//===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
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 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
12// performance win.
13//
14//===----------------------------------------------------------------------===//
15//
16// TODO List:
17//
18// Future loop memory idioms to recognize:
19//   memcmp, memmove, strlen, etc.
20// Future floating point idioms to recognize in -ffast-math mode:
21//   fpowi
22// Future integer operation idioms to recognize:
23//   ctpop, ctlz, cttz
24//
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.
28//
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//
34// We should enhance this to handle negative strides through memory.
35// Alternatively (and perhaps better) we could rely on an earlier pass to force
36// forward iteration through memory, which is generally better for cache
37// behavior.  Negative strides *do* happen for memset/memcpy loops.
38//
39// This could recognize common matrix multiplies and dot product idioms and
40// replace them with calls to BLAS (if linked in??).
41//
42//===----------------------------------------------------------------------===//
43
44#define DEBUG_TYPE "loop-idiom"
45#include "llvm/Transforms/Scalar.h"
46#include "llvm/IntrinsicInst.h"
47#include "llvm/Module.h"
48#include "llvm/Analysis/AliasAnalysis.h"
49#include "llvm/Analysis/LoopPass.h"
50#include "llvm/Analysis/ScalarEvolutionExpressions.h"
51#include "llvm/Analysis/ScalarEvolutionExpander.h"
52#include "llvm/Analysis/ValueTracking.h"
53#include "llvm/Target/TargetData.h"
54#include "llvm/Target/TargetLibraryInfo.h"
55#include "llvm/Transforms/Utils/Local.h"
56#include "llvm/Support/Debug.h"
57#include "llvm/Support/IRBuilder.h"
58#include "llvm/Support/raw_ostream.h"
59#include "llvm/ADT/Statistic.h"
60using namespace llvm;
61
62STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
63STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
64
65namespace {
66  class LoopIdiomRecognize : public LoopPass {
67    Loop *CurLoop;
68    const TargetData *TD;
69    DominatorTree *DT;
70    ScalarEvolution *SE;
71    TargetLibraryInfo *TLI;
72  public:
73    static char ID;
74    explicit LoopIdiomRecognize() : LoopPass(ID) {
75      initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
76    }
77
78    bool runOnLoop(Loop *L, LPPassManager &LPM);
79    bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
80                        SmallVectorImpl<BasicBlock*> &ExitBlocks);
81
82    bool processLoopStore(StoreInst *SI, const SCEV *BECount);
83    bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
84
85    bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
86                                 unsigned StoreAlignment,
87                                 Value *SplatValue, Instruction *TheStore,
88                                 const SCEVAddRecExpr *Ev,
89                                 const SCEV *BECount);
90    bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
91                                    const SCEVAddRecExpr *StoreEv,
92                                    const SCEVAddRecExpr *LoadEv,
93                                    const SCEV *BECount);
94
95    /// This transformation requires natural loop information & requires that
96    /// loop preheaders be inserted into the CFG.
97    ///
98    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
99      AU.addRequired<LoopInfo>();
100      AU.addPreserved<LoopInfo>();
101      AU.addRequiredID(LoopSimplifyID);
102      AU.addPreservedID(LoopSimplifyID);
103      AU.addRequiredID(LCSSAID);
104      AU.addPreservedID(LCSSAID);
105      AU.addRequired<AliasAnalysis>();
106      AU.addPreserved<AliasAnalysis>();
107      AU.addRequired<ScalarEvolution>();
108      AU.addPreserved<ScalarEvolution>();
109      AU.addPreserved<DominatorTree>();
110      AU.addRequired<DominatorTree>();
111      AU.addRequired<TargetLibraryInfo>();
112    }
113  };
114}
115
116char LoopIdiomRecognize::ID = 0;
117INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
118                      false, false)
119INITIALIZE_PASS_DEPENDENCY(LoopInfo)
120INITIALIZE_PASS_DEPENDENCY(DominatorTree)
121INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
122INITIALIZE_PASS_DEPENDENCY(LCSSA)
123INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
124INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
125INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
126INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
127                    false, false)
128
129Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
130
131/// deleteDeadInstruction - Delete this instruction.  Before we do, go through
132/// and zero out all the operands of this instruction.  If any of them become
133/// dead, delete them and the computation tree that feeds them.
134///
135static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE) {
136  SmallVector<Instruction*, 32> NowDeadInsts;
137
138  NowDeadInsts.push_back(I);
139
140  // Before we touch this instruction, remove it from SE!
141  do {
142    Instruction *DeadInst = NowDeadInsts.pop_back_val();
143
144    // This instruction is dead, zap it, in stages.  Start by removing it from
145    // SCEV.
146    SE.forgetValue(DeadInst);
147
148    for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
149      Value *Op = DeadInst->getOperand(op);
150      DeadInst->setOperand(op, 0);
151
152      // If this operand just became dead, add it to the NowDeadInsts list.
153      if (!Op->use_empty()) continue;
154
155      if (Instruction *OpI = dyn_cast<Instruction>(Op))
156        if (isInstructionTriviallyDead(OpI))
157          NowDeadInsts.push_back(OpI);
158    }
159
160    DeadInst->eraseFromParent();
161
162  } while (!NowDeadInsts.empty());
163}
164
165/// deleteIfDeadInstruction - If the specified value is a dead instruction,
166/// delete it and any recursively used instructions.
167static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE) {
168  if (Instruction *I = dyn_cast<Instruction>(V))
169    if (isInstructionTriviallyDead(I))
170      deleteDeadInstruction(I, SE);
171}
172
173bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
174  CurLoop = L;
175
176  // Disable loop idiom recognition if the function's name is a common idiom.
177  StringRef Name = L->getHeader()->getParent()->getName();
178  if (Name == "memset" || Name == "memcpy")
179    return false;
180
181  // The trip count of the loop must be analyzable.
182  SE = &getAnalysis<ScalarEvolution>();
183  if (!SE->hasLoopInvariantBackedgeTakenCount(L))
184    return false;
185  const SCEV *BECount = SE->getBackedgeTakenCount(L);
186  if (isa<SCEVCouldNotCompute>(BECount)) return false;
187
188  // If this loop executes exactly one time, then it should be peeled, not
189  // optimized by this pass.
190  if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
191    if (BECst->getValue()->getValue() == 0)
192      return false;
193
194  // We require target data for now.
195  TD = getAnalysisIfAvailable<TargetData>();
196  if (TD == 0) return false;
197
198  DT = &getAnalysis<DominatorTree>();
199  LoopInfo &LI = getAnalysis<LoopInfo>();
200  TLI = &getAnalysis<TargetLibraryInfo>();
201
202  SmallVector<BasicBlock*, 8> ExitBlocks;
203  CurLoop->getUniqueExitBlocks(ExitBlocks);
204
205  DEBUG(dbgs() << "loop-idiom Scanning: F["
206               << L->getHeader()->getParent()->getName()
207               << "] Loop %" << L->getHeader()->getName() << "\n");
208
209  bool MadeChange = false;
210  // Scan all the blocks in the loop that are not in subloops.
211  for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
212       ++BI) {
213    // Ignore blocks in subloops.
214    if (LI.getLoopFor(*BI) != CurLoop)
215      continue;
216
217    MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
218  }
219  return MadeChange;
220}
221
222/// runOnLoopBlock - Process the specified block, which lives in a counted loop
223/// with the specified backedge count.  This block is known to be in the current
224/// loop and not in any subloops.
225bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
226                                     SmallVectorImpl<BasicBlock*> &ExitBlocks) {
227  // We can only promote stores in this block if they are unconditionally
228  // executed in the loop.  For a block to be unconditionally executed, it has
229  // to dominate all the exit blocks of the loop.  Verify this now.
230  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
231    if (!DT->dominates(BB, ExitBlocks[i]))
232      return false;
233
234  bool MadeChange = false;
235  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
236    Instruction *Inst = I++;
237    // Look for store instructions, which may be optimized to memset/memcpy.
238    if (StoreInst *SI = dyn_cast<StoreInst>(Inst))  {
239      WeakVH InstPtr(I);
240      if (!processLoopStore(SI, BECount)) continue;
241      MadeChange = true;
242
243      // If processing the store invalidated our iterator, start over from the
244      // top of the block.
245      if (InstPtr == 0)
246        I = BB->begin();
247      continue;
248    }
249
250    // Look for memset instructions, which may be optimized to a larger memset.
251    if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst))  {
252      WeakVH InstPtr(I);
253      if (!processLoopMemSet(MSI, BECount)) continue;
254      MadeChange = true;
255
256      // If processing the memset invalidated our iterator, start over from the
257      // top of the block.
258      if (InstPtr == 0)
259        I = BB->begin();
260      continue;
261    }
262  }
263
264  return MadeChange;
265}
266
267
268/// processLoopStore - See if this store can be promoted to a memset or memcpy.
269bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
270  if (!SI->isSimple()) return false;
271
272  Value *StoredVal = SI->getValueOperand();
273  Value *StorePtr = SI->getPointerOperand();
274
275  // Reject stores that are so large that they overflow an unsigned.
276  uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
277  if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
278    return false;
279
280  // See if the pointer expression is an AddRec like {base,+,1} on the current
281  // loop, which indicates a strided store.  If we have something else, it's a
282  // random store we can't handle.
283  const SCEVAddRecExpr *StoreEv =
284    dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
285  if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
286    return false;
287
288  // Check to see if the stride matches the size of the store.  If so, then we
289  // know that every byte is touched in the loop.
290  unsigned StoreSize = (unsigned)SizeInBits >> 3;
291  const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
292
293  if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) {
294    // TODO: Could also handle negative stride here someday, that will require
295    // the validity check in mayLoopAccessLocation to be updated though.
296    // Enable this to print exact negative strides.
297    if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
298      dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
299      dbgs() << "BB: " << *SI->getParent();
300    }
301
302    return false;
303  }
304
305  // See if we can optimize just this store in isolation.
306  if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
307                              StoredVal, SI, StoreEv, BECount))
308    return true;
309
310  // If the stored value is a strided load in the same loop with the same stride
311  // this this may be transformable into a memcpy.  This kicks in for stuff like
312  //   for (i) A[i] = B[i];
313  if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
314    const SCEVAddRecExpr *LoadEv =
315      dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
316    if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
317        StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
318      if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
319        return true;
320  }
321  //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
322
323  return false;
324}
325
326/// processLoopMemSet - See if this memset can be promoted to a large memset.
327bool LoopIdiomRecognize::
328processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
329  // We can only handle non-volatile memsets with a constant size.
330  if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
331
332  // If we're not allowed to hack on memset, we fail.
333  if (!TLI->has(LibFunc::memset))
334    return false;
335
336  Value *Pointer = MSI->getDest();
337
338  // See if the pointer expression is an AddRec like {base,+,1} on the current
339  // loop, which indicates a strided store.  If we have something else, it's a
340  // random store we can't handle.
341  const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
342  if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine())
343    return false;
344
345  // Reject memsets that are so large that they overflow an unsigned.
346  uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
347  if ((SizeInBytes >> 32) != 0)
348    return false;
349
350  // Check to see if the stride matches the size of the memset.  If so, then we
351  // know that every byte is touched in the loop.
352  const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
353
354  // TODO: Could also handle negative stride here someday, that will require the
355  // validity check in mayLoopAccessLocation to be updated though.
356  if (Stride == 0 || MSI->getLength() != Stride->getValue())
357    return false;
358
359  return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
360                                 MSI->getAlignment(), MSI->getValue(),
361                                 MSI, Ev, BECount);
362}
363
364
365/// mayLoopAccessLocation - Return true if the specified loop might access the
366/// specified pointer location, which is a loop-strided access.  The 'Access'
367/// argument specifies what the verboten forms of access are (read or write).
368static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
369                                  Loop *L, const SCEV *BECount,
370                                  unsigned StoreSize, AliasAnalysis &AA,
371                                  Instruction *IgnoredStore) {
372  // Get the location that may be stored across the loop.  Since the access is
373  // strided positively through memory, we say that the modified location starts
374  // at the pointer and has infinite size.
375  uint64_t AccessSize = AliasAnalysis::UnknownSize;
376
377  // If the loop iterates a fixed number of times, we can refine the access size
378  // to be exactly the size of the memset, which is (BECount+1)*StoreSize
379  if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
380    AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
381
382  // TODO: For this to be really effective, we have to dive into the pointer
383  // operand in the store.  Store to &A[i] of 100 will always return may alias
384  // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
385  // which will then no-alias a store to &A[100].
386  AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
387
388  for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
389       ++BI)
390    for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
391      if (&*I != IgnoredStore &&
392          (AA.getModRefInfo(I, StoreLoc) & Access))
393        return true;
394
395  return false;
396}
397
398/// getMemSetPatternValue - If a strided store of the specified value is safe to
399/// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
400/// be passed in.  Otherwise, return null.
401///
402/// Note that we don't ever attempt to use memset_pattern8 or 4, because these
403/// just replicate their input array and then pass on to memset_pattern16.
404static Constant *getMemSetPatternValue(Value *V, const TargetData &TD) {
405  // If the value isn't a constant, we can't promote it to being in a constant
406  // array.  We could theoretically do a store to an alloca or something, but
407  // that doesn't seem worthwhile.
408  Constant *C = dyn_cast<Constant>(V);
409  if (C == 0) return 0;
410
411  // Only handle simple values that are a power of two bytes in size.
412  uint64_t Size = TD.getTypeSizeInBits(V->getType());
413  if (Size == 0 || (Size & 7) || (Size & (Size-1)))
414    return 0;
415
416  // Don't care enough about darwin/ppc to implement this.
417  if (TD.isBigEndian())
418    return 0;
419
420  // Convert to size in bytes.
421  Size /= 8;
422
423  // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
424  // if the top and bottom are the same (e.g. for vectors and large integers).
425  if (Size > 16) return 0;
426
427  // If the constant is exactly 16 bytes, just use it.
428  if (Size == 16) return C;
429
430  // Otherwise, we'll use an array of the constants.
431  unsigned ArraySize = 16/Size;
432  ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
433  return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
434}
435
436
437/// processLoopStridedStore - We see a strided store of some value.  If we can
438/// transform this into a memset or memset_pattern in the loop preheader, do so.
439bool LoopIdiomRecognize::
440processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
441                        unsigned StoreAlignment, Value *StoredVal,
442                        Instruction *TheStore, const SCEVAddRecExpr *Ev,
443                        const SCEV *BECount) {
444
445  // If the stored value is a byte-wise value (like i32 -1), then it may be
446  // turned into a memset of i8 -1, assuming that all the consecutive bytes
447  // are stored.  A store of i32 0x01020304 can never be turned into a memset,
448  // but it can be turned into memset_pattern if the target supports it.
449  Value *SplatValue = isBytewiseValue(StoredVal);
450  Constant *PatternValue = 0;
451
452  // If we're allowed to form a memset, and the stored value would be acceptable
453  // for memset, use it.
454  if (SplatValue && TLI->has(LibFunc::memset) &&
455      // Verify that the stored value is loop invariant.  If not, we can't
456      // promote the memset.
457      CurLoop->isLoopInvariant(SplatValue)) {
458    // Keep and use SplatValue.
459    PatternValue = 0;
460  } else if (TLI->has(LibFunc::memset_pattern16) &&
461             (PatternValue = getMemSetPatternValue(StoredVal, *TD))) {
462    // It looks like we can use PatternValue!
463    SplatValue = 0;
464  } else {
465    // Otherwise, this isn't an idiom we can transform.  For example, we can't
466    // do anything with a 3-byte store.
467    return false;
468  }
469
470  // The trip count of the loop and the base pointer of the addrec SCEV is
471  // guaranteed to be loop invariant, which means that it should dominate the
472  // header.  This allows us to insert code for it in the preheader.
473  BasicBlock *Preheader = CurLoop->getLoopPreheader();
474  IRBuilder<> Builder(Preheader->getTerminator());
475  SCEVExpander Expander(*SE, "loop-idiom");
476
477  // Okay, we have a strided store "p[i]" of a splattable value.  We can turn
478  // this into a memset in the loop preheader now if we want.  However, this
479  // would be unsafe to do if there is anything else in the loop that may read
480  // or write to the aliased location.  Check for any overlap by generating the
481  // base pointer and checking the region.
482  unsigned AddrSpace = cast<PointerType>(DestPtr->getType())->getAddressSpace();
483  Value *BasePtr =
484    Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
485                           Preheader->getTerminator());
486
487
488  if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
489                            CurLoop, BECount,
490                            StoreSize, getAnalysis<AliasAnalysis>(), TheStore)){
491    Expander.clear();
492    // If we generated new code for the base pointer, clean up.
493    deleteIfDeadInstruction(BasePtr, *SE);
494    return false;
495  }
496
497  // Okay, everything looks good, insert the memset.
498
499  // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
500  // pointer size if it isn't already.
501  Type *IntPtr = TD->getIntPtrType(DestPtr->getContext());
502  BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
503
504  const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
505                                         SCEV::FlagNUW);
506  if (StoreSize != 1)
507    NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
508                               SCEV::FlagNUW);
509
510  Value *NumBytes =
511    Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
512
513  CallInst *NewCall;
514  if (SplatValue)
515    NewCall = Builder.CreateMemSet(BasePtr, SplatValue,NumBytes,StoreAlignment);
516  else {
517    Module *M = TheStore->getParent()->getParent()->getParent();
518    Value *MSP = M->getOrInsertFunction("memset_pattern16",
519                                        Builder.getVoidTy(),
520                                        Builder.getInt8PtrTy(),
521                                        Builder.getInt8PtrTy(), IntPtr,
522                                        (void*)0);
523
524    // Otherwise we should form a memset_pattern16.  PatternValue is known to be
525    // an constant array of 16-bytes.  Plop the value into a mergable global.
526    GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
527                                            GlobalValue::InternalLinkage,
528                                            PatternValue, ".memset_pattern");
529    GV->setUnnamedAddr(true); // Ok to merge these.
530    GV->setAlignment(16);
531    Value *PatternPtr = ConstantExpr::getBitCast(GV, Builder.getInt8PtrTy());
532    NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes);
533  }
534
535  DEBUG(dbgs() << "  Formed memset: " << *NewCall << "\n"
536               << "    from store to: " << *Ev << " at: " << *TheStore << "\n");
537  NewCall->setDebugLoc(TheStore->getDebugLoc());
538
539  // Okay, the memset has been formed.  Zap the original store and anything that
540  // feeds into it.
541  deleteDeadInstruction(TheStore, *SE);
542  ++NumMemSet;
543  return true;
544}
545
546/// processLoopStoreOfLoopLoad - We see a strided store whose value is a
547/// same-strided load.
548bool LoopIdiomRecognize::
549processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
550                           const SCEVAddRecExpr *StoreEv,
551                           const SCEVAddRecExpr *LoadEv,
552                           const SCEV *BECount) {
553  // If we're not allowed to form memcpy, we fail.
554  if (!TLI->has(LibFunc::memcpy))
555    return false;
556
557  LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
558
559  // The trip count of the loop and the base pointer of the addrec SCEV is
560  // guaranteed to be loop invariant, which means that it should dominate the
561  // header.  This allows us to insert code for it in the preheader.
562  BasicBlock *Preheader = CurLoop->getLoopPreheader();
563  IRBuilder<> Builder(Preheader->getTerminator());
564  SCEVExpander Expander(*SE, "loop-idiom");
565
566  // Okay, we have a strided store "p[i]" of a loaded value.  We can turn
567  // this into a memcpy in the loop preheader now if we want.  However, this
568  // would be unsafe to do if there is anything else in the loop that may read
569  // or write the memory region we're storing to.  This includes the load that
570  // feeds the stores.  Check for an alias by generating the base address and
571  // checking everything.
572  Value *StoreBasePtr =
573    Expander.expandCodeFor(StoreEv->getStart(),
574                           Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
575                           Preheader->getTerminator());
576
577  if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
578                            CurLoop, BECount, StoreSize,
579                            getAnalysis<AliasAnalysis>(), SI)) {
580    Expander.clear();
581    // If we generated new code for the base pointer, clean up.
582    deleteIfDeadInstruction(StoreBasePtr, *SE);
583    return false;
584  }
585
586  // For a memcpy, we have to make sure that the input array is not being
587  // mutated by the loop.
588  Value *LoadBasePtr =
589    Expander.expandCodeFor(LoadEv->getStart(),
590                           Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
591                           Preheader->getTerminator());
592
593  if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
594                            StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
595    Expander.clear();
596    // If we generated new code for the base pointer, clean up.
597    deleteIfDeadInstruction(LoadBasePtr, *SE);
598    deleteIfDeadInstruction(StoreBasePtr, *SE);
599    return false;
600  }
601
602  // Okay, everything is safe, we can transform this!
603
604
605  // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
606  // pointer size if it isn't already.
607  Type *IntPtr = TD->getIntPtrType(SI->getContext());
608  BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
609
610  const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
611                                         SCEV::FlagNUW);
612  if (StoreSize != 1)
613    NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
614                               SCEV::FlagNUW);
615
616  Value *NumBytes =
617    Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
618
619  CallInst *NewCall =
620    Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
621                         std::min(SI->getAlignment(), LI->getAlignment()));
622  NewCall->setDebugLoc(SI->getDebugLoc());
623
624  DEBUG(dbgs() << "  Formed memcpy: " << *NewCall << "\n"
625               << "    from load ptr=" << *LoadEv << " at: " << *LI << "\n"
626               << "    from store ptr=" << *StoreEv << " at: " << *SI << "\n");
627
628
629  // Okay, the memset has been formed.  Zap the original store and anything that
630  // feeds into it.
631  deleteDeadInstruction(SI, *SE);
632  ++NumMemCpy;
633  return true;
634}
635