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 // This could recognize common matrix multiplies and dot product idioms and
35 // replace them with calls to BLAS (if linked in??).
36 //
37 //===----------------------------------------------------------------------===//
38
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/Analysis/BasicAliasAnalysis.h"
43 #include "llvm/Analysis/GlobalsModRef.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
46 #include "llvm/Analysis/ScalarEvolutionExpander.h"
47 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/TargetTransformInfo.h"
50 #include "llvm/Analysis/ValueTracking.h"
51 #include "llvm/IR/DataLayout.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/IRBuilder.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Module.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 using namespace llvm;
60
61 #define DEBUG_TYPE "loop-idiom"
62
63 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
64 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
65
66 namespace {
67
68 class LoopIdiomRecognize : public LoopPass {
69 Loop *CurLoop;
70 AliasAnalysis *AA;
71 DominatorTree *DT;
72 LoopInfo *LI;
73 ScalarEvolution *SE;
74 TargetLibraryInfo *TLI;
75 const TargetTransformInfo *TTI;
76 const DataLayout *DL;
77
78 public:
79 static char ID;
LoopIdiomRecognize()80 explicit LoopIdiomRecognize() : LoopPass(ID) {
81 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
82 }
83
84 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
85
86 /// This transformation requires natural loop information & requires that
87 /// loop preheaders be inserted into the CFG.
88 ///
getAnalysisUsage(AnalysisUsage & AU) const89 void getAnalysisUsage(AnalysisUsage &AU) const override {
90 AU.addRequired<LoopInfoWrapperPass>();
91 AU.addPreserved<LoopInfoWrapperPass>();
92 AU.addRequiredID(LoopSimplifyID);
93 AU.addPreservedID(LoopSimplifyID);
94 AU.addRequiredID(LCSSAID);
95 AU.addPreservedID(LCSSAID);
96 AU.addRequired<AAResultsWrapperPass>();
97 AU.addPreserved<AAResultsWrapperPass>();
98 AU.addRequired<ScalarEvolutionWrapperPass>();
99 AU.addPreserved<ScalarEvolutionWrapperPass>();
100 AU.addPreserved<SCEVAAWrapperPass>();
101 AU.addRequired<DominatorTreeWrapperPass>();
102 AU.addPreserved<DominatorTreeWrapperPass>();
103 AU.addRequired<TargetLibraryInfoWrapperPass>();
104 AU.addRequired<TargetTransformInfoWrapperPass>();
105 AU.addPreserved<BasicAAWrapperPass>();
106 AU.addPreserved<GlobalsAAWrapperPass>();
107 }
108
109 private:
110 typedef SmallVector<StoreInst *, 8> StoreList;
111 StoreList StoreRefs;
112
113 /// \name Countable Loop Idiom Handling
114 /// @{
115
116 bool runOnCountableLoop();
117 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
118 SmallVectorImpl<BasicBlock *> &ExitBlocks);
119
120 void collectStores(BasicBlock *BB);
121 bool isLegalStore(StoreInst *SI);
122 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
123 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
124
125 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
126 unsigned StoreAlignment, Value *SplatValue,
127 Instruction *TheStore, const SCEVAddRecExpr *Ev,
128 const SCEV *BECount, bool NegStride);
129 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
130 const SCEVAddRecExpr *StoreEv,
131 const SCEV *BECount, bool NegStride);
132
133 /// @}
134 /// \name Noncountable Loop Idiom Handling
135 /// @{
136
137 bool runOnNoncountableLoop();
138
139 bool recognizePopcount();
140 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
141 PHINode *CntPhi, Value *Var);
142
143 /// @}
144 };
145
146 } // End anonymous namespace.
147
148 char LoopIdiomRecognize::ID = 0;
149 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
150 false, false)
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)151 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
152 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
153 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
154 INITIALIZE_PASS_DEPENDENCY(LCSSA)
155 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
156 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
157 INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
158 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
159 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
160 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
162 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
163 false, false)
164
165 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
166
167 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
168 /// and zero out all the operands of this instruction. If any of them become
169 /// dead, delete them and the computation tree that feeds them.
170 ///
deleteDeadInstruction(Instruction * I,const TargetLibraryInfo * TLI)171 static void deleteDeadInstruction(Instruction *I,
172 const TargetLibraryInfo *TLI) {
173 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
174 I->replaceAllUsesWith(UndefValue::get(I->getType()));
175 I->eraseFromParent();
176 for (Value *Op : Operands)
177 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
178 }
179
180 //===----------------------------------------------------------------------===//
181 //
182 // Implementation of LoopIdiomRecognize
183 //
184 //===----------------------------------------------------------------------===//
185
runOnLoop(Loop * L,LPPassManager & LPM)186 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
187 if (skipOptnoneFunction(L))
188 return false;
189
190 CurLoop = L;
191 // If the loop could not be converted to canonical form, it must have an
192 // indirectbr in it, just give up.
193 if (!L->getLoopPreheader())
194 return false;
195
196 // Disable loop idiom recognition if the function's name is a common idiom.
197 StringRef Name = L->getHeader()->getParent()->getName();
198 if (Name == "memset" || Name == "memcpy")
199 return false;
200
201 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
202 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
203 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
204 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
205 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
206 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
207 *CurLoop->getHeader()->getParent());
208 DL = &CurLoop->getHeader()->getModule()->getDataLayout();
209
210 if (SE->hasLoopInvariantBackedgeTakenCount(L))
211 return runOnCountableLoop();
212
213 return runOnNoncountableLoop();
214 }
215
runOnCountableLoop()216 bool LoopIdiomRecognize::runOnCountableLoop() {
217 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
218 assert(!isa<SCEVCouldNotCompute>(BECount) &&
219 "runOnCountableLoop() called on a loop without a predictable"
220 "backedge-taken count");
221
222 // If this loop executes exactly one time, then it should be peeled, not
223 // optimized by this pass.
224 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
225 if (BECst->getAPInt() == 0)
226 return false;
227
228 SmallVector<BasicBlock *, 8> ExitBlocks;
229 CurLoop->getUniqueExitBlocks(ExitBlocks);
230
231 DEBUG(dbgs() << "loop-idiom Scanning: F["
232 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
233 << CurLoop->getHeader()->getName() << "\n");
234
235 bool MadeChange = false;
236 // Scan all the blocks in the loop that are not in subloops.
237 for (auto *BB : CurLoop->getBlocks()) {
238 // Ignore blocks in subloops.
239 if (LI->getLoopFor(BB) != CurLoop)
240 continue;
241
242 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
243 }
244 return MadeChange;
245 }
246
getStoreSizeInBytes(StoreInst * SI,const DataLayout * DL)247 static unsigned getStoreSizeInBytes(StoreInst *SI, const DataLayout *DL) {
248 uint64_t SizeInBits = DL->getTypeSizeInBits(SI->getValueOperand()->getType());
249 assert(((SizeInBits & 7) || (SizeInBits >> 32) == 0) &&
250 "Don't overflow unsigned.");
251 return (unsigned)SizeInBits >> 3;
252 }
253
getStoreStride(const SCEVAddRecExpr * StoreEv)254 static unsigned getStoreStride(const SCEVAddRecExpr *StoreEv) {
255 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
256 return ConstStride->getAPInt().getZExtValue();
257 }
258
259 /// getMemSetPatternValue - If a strided store of the specified value is safe to
260 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
261 /// be passed in. Otherwise, return null.
262 ///
263 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
264 /// just replicate their input array and then pass on to memset_pattern16.
getMemSetPatternValue(Value * V,const DataLayout * DL)265 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
266 // If the value isn't a constant, we can't promote it to being in a constant
267 // array. We could theoretically do a store to an alloca or something, but
268 // that doesn't seem worthwhile.
269 Constant *C = dyn_cast<Constant>(V);
270 if (!C)
271 return nullptr;
272
273 // Only handle simple values that are a power of two bytes in size.
274 uint64_t Size = DL->getTypeSizeInBits(V->getType());
275 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
276 return nullptr;
277
278 // Don't care enough about darwin/ppc to implement this.
279 if (DL->isBigEndian())
280 return nullptr;
281
282 // Convert to size in bytes.
283 Size /= 8;
284
285 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
286 // if the top and bottom are the same (e.g. for vectors and large integers).
287 if (Size > 16)
288 return nullptr;
289
290 // If the constant is exactly 16 bytes, just use it.
291 if (Size == 16)
292 return C;
293
294 // Otherwise, we'll use an array of the constants.
295 unsigned ArraySize = 16 / Size;
296 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
297 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
298 }
299
isLegalStore(StoreInst * SI)300 bool LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
301 // Don't touch volatile stores.
302 if (!SI->isSimple())
303 return false;
304
305 Value *StoredVal = SI->getValueOperand();
306 Value *StorePtr = SI->getPointerOperand();
307
308 // Reject stores that are so large that they overflow an unsigned.
309 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
310 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
311 return false;
312
313 // See if the pointer expression is an AddRec like {base,+,1} on the current
314 // loop, which indicates a strided store. If we have something else, it's a
315 // random store we can't handle.
316 const SCEVAddRecExpr *StoreEv =
317 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
318 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
319 return false;
320
321 // Check to see if we have a constant stride.
322 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
323 return false;
324
325 return true;
326 }
327
collectStores(BasicBlock * BB)328 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
329 StoreRefs.clear();
330 for (Instruction &I : *BB) {
331 StoreInst *SI = dyn_cast<StoreInst>(&I);
332 if (!SI)
333 continue;
334
335 // Make sure this is a strided store with a constant stride.
336 if (!isLegalStore(SI))
337 continue;
338
339 // Save the store locations.
340 StoreRefs.push_back(SI);
341 }
342 }
343
344 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
345 /// with the specified backedge count. This block is known to be in the current
346 /// loop and not in any subloops.
runOnLoopBlock(BasicBlock * BB,const SCEV * BECount,SmallVectorImpl<BasicBlock * > & ExitBlocks)347 bool LoopIdiomRecognize::runOnLoopBlock(
348 BasicBlock *BB, const SCEV *BECount,
349 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
350 // We can only promote stores in this block if they are unconditionally
351 // executed in the loop. For a block to be unconditionally executed, it has
352 // to dominate all the exit blocks of the loop. Verify this now.
353 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
354 if (!DT->dominates(BB, ExitBlocks[i]))
355 return false;
356
357 bool MadeChange = false;
358 // Look for store instructions, which may be optimized to memset/memcpy.
359 collectStores(BB);
360 for (auto &SI : StoreRefs)
361 MadeChange |= processLoopStore(SI, BECount);
362
363 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
364 Instruction *Inst = &*I++;
365 // Look for memset instructions, which may be optimized to a larger memset.
366 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
367 WeakVH InstPtr(&*I);
368 if (!processLoopMemSet(MSI, BECount))
369 continue;
370 MadeChange = true;
371
372 // If processing the memset invalidated our iterator, start over from the
373 // top of the block.
374 if (!InstPtr)
375 I = BB->begin();
376 continue;
377 }
378 }
379
380 return MadeChange;
381 }
382
383 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
processLoopStore(StoreInst * SI,const SCEV * BECount)384 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
385 assert(SI->isSimple() && "Expected only non-volatile stores.");
386
387 Value *StoredVal = SI->getValueOperand();
388 Value *StorePtr = SI->getPointerOperand();
389
390 // Check to see if the stride matches the size of the store. If so, then we
391 // know that every byte is touched in the loop.
392 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
393 unsigned Stride = getStoreStride(StoreEv);
394 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
395 if (StoreSize != Stride && StoreSize != -Stride)
396 return false;
397
398 bool NegStride = StoreSize == -Stride;
399
400 // See if we can optimize just this store in isolation.
401 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
402 StoredVal, SI, StoreEv, BECount, NegStride))
403 return true;
404
405 // Optimize the store into a memcpy, if it feeds an similarly strided load.
406 return processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, BECount, NegStride);
407 }
408
409 /// processLoopMemSet - See if this memset can be promoted to a large memset.
processLoopMemSet(MemSetInst * MSI,const SCEV * BECount)410 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
411 const SCEV *BECount) {
412 // We can only handle non-volatile memsets with a constant size.
413 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
414 return false;
415
416 // If we're not allowed to hack on memset, we fail.
417 if (!TLI->has(LibFunc::memset))
418 return false;
419
420 Value *Pointer = MSI->getDest();
421
422 // See if the pointer expression is an AddRec like {base,+,1} on the current
423 // loop, which indicates a strided store. If we have something else, it's a
424 // random store we can't handle.
425 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
426 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
427 return false;
428
429 // Reject memsets that are so large that they overflow an unsigned.
430 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
431 if ((SizeInBytes >> 32) != 0)
432 return false;
433
434 // Check to see if the stride matches the size of the memset. If so, then we
435 // know that every byte is touched in the loop.
436 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
437
438 // TODO: Could also handle negative stride here someday, that will require the
439 // validity check in mayLoopAccessLocation to be updated though.
440 if (!Stride || MSI->getLength() != Stride->getValue())
441 return false;
442
443 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
444 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
445 BECount, /*NegStride=*/false);
446 }
447
448 /// mayLoopAccessLocation - Return true if the specified loop might access the
449 /// specified pointer location, which is a loop-strided access. The 'Access'
450 /// argument specifies what the verboten forms of access are (read or write).
mayLoopAccessLocation(Value * Ptr,ModRefInfo Access,Loop * L,const SCEV * BECount,unsigned StoreSize,AliasAnalysis & AA,Instruction * IgnoredStore)451 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
452 const SCEV *BECount, unsigned StoreSize,
453 AliasAnalysis &AA,
454 Instruction *IgnoredStore) {
455 // Get the location that may be stored across the loop. Since the access is
456 // strided positively through memory, we say that the modified location starts
457 // at the pointer and has infinite size.
458 uint64_t AccessSize = MemoryLocation::UnknownSize;
459
460 // If the loop iterates a fixed number of times, we can refine the access size
461 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
462 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
463 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
464
465 // TODO: For this to be really effective, we have to dive into the pointer
466 // operand in the store. Store to &A[i] of 100 will always return may alias
467 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
468 // which will then no-alias a store to &A[100].
469 MemoryLocation StoreLoc(Ptr, AccessSize);
470
471 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
472 ++BI)
473 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
474 if (&*I != IgnoredStore && (AA.getModRefInfo(&*I, StoreLoc) & Access))
475 return true;
476
477 return false;
478 }
479
480 // If we have a negative stride, Start refers to the end of the memory location
481 // we're trying to memset. Therefore, we need to recompute the base pointer,
482 // which is just Start - BECount*Size.
getStartForNegStride(const SCEV * Start,const SCEV * BECount,Type * IntPtr,unsigned StoreSize,ScalarEvolution * SE)483 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
484 Type *IntPtr, unsigned StoreSize,
485 ScalarEvolution *SE) {
486 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
487 if (StoreSize != 1)
488 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
489 SCEV::FlagNUW);
490 return SE->getMinusSCEV(Start, Index);
491 }
492
493 /// processLoopStridedStore - We see a strided store of some value. If we can
494 /// transform this into a memset or memset_pattern in the loop preheader, do so.
processLoopStridedStore(Value * DestPtr,unsigned StoreSize,unsigned StoreAlignment,Value * StoredVal,Instruction * TheStore,const SCEVAddRecExpr * Ev,const SCEV * BECount,bool NegStride)495 bool LoopIdiomRecognize::processLoopStridedStore(
496 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
497 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
498 const SCEV *BECount, bool NegStride) {
499
500 // If the stored value is a byte-wise value (like i32 -1), then it may be
501 // turned into a memset of i8 -1, assuming that all the consecutive bytes
502 // are stored. A store of i32 0x01020304 can never be turned into a memset,
503 // but it can be turned into memset_pattern if the target supports it.
504 Value *SplatValue = isBytewiseValue(StoredVal);
505 Constant *PatternValue = nullptr;
506 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
507
508 // If we're allowed to form a memset, and the stored value would be acceptable
509 // for memset, use it.
510 if (SplatValue && TLI->has(LibFunc::memset) &&
511 // Verify that the stored value is loop invariant. If not, we can't
512 // promote the memset.
513 CurLoop->isLoopInvariant(SplatValue)) {
514 // Keep and use SplatValue.
515 PatternValue = nullptr;
516 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
517 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
518 // Don't create memset_pattern16s with address spaces.
519 // It looks like we can use PatternValue!
520 SplatValue = nullptr;
521 } else {
522 // Otherwise, this isn't an idiom we can transform. For example, we can't
523 // do anything with a 3-byte store.
524 return false;
525 }
526
527 // The trip count of the loop and the base pointer of the addrec SCEV is
528 // guaranteed to be loop invariant, which means that it should dominate the
529 // header. This allows us to insert code for it in the preheader.
530 BasicBlock *Preheader = CurLoop->getLoopPreheader();
531 IRBuilder<> Builder(Preheader->getTerminator());
532 SCEVExpander Expander(*SE, *DL, "loop-idiom");
533
534 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
535 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
536
537 const SCEV *Start = Ev->getStart();
538 // Handle negative strided loops.
539 if (NegStride)
540 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
541
542 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
543 // this into a memset in the loop preheader now if we want. However, this
544 // would be unsafe to do if there is anything else in the loop that may read
545 // or write to the aliased location. Check for any overlap by generating the
546 // base pointer and checking the region.
547 Value *BasePtr =
548 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
549 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
550 *AA, TheStore)) {
551 Expander.clear();
552 // If we generated new code for the base pointer, clean up.
553 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
554 return false;
555 }
556
557 // Okay, everything looks good, insert the memset.
558
559 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
560 // pointer size if it isn't already.
561 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
562
563 const SCEV *NumBytesS =
564 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
565 if (StoreSize != 1) {
566 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
567 SCEV::FlagNUW);
568 }
569
570 Value *NumBytes =
571 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
572
573 CallInst *NewCall;
574 if (SplatValue) {
575 NewCall =
576 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
577 } else {
578 // Everything is emitted in default address space
579 Type *Int8PtrTy = DestInt8PtrTy;
580
581 Module *M = TheStore->getModule();
582 Value *MSP =
583 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
584 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
585
586 // Otherwise we should form a memset_pattern16. PatternValue is known to be
587 // an constant array of 16-bytes. Plop the value into a mergable global.
588 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
589 GlobalValue::PrivateLinkage,
590 PatternValue, ".memset_pattern");
591 GV->setUnnamedAddr(true); // Ok to merge these.
592 GV->setAlignment(16);
593 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
594 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
595 }
596
597 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
598 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
599 NewCall->setDebugLoc(TheStore->getDebugLoc());
600
601 // Okay, the memset has been formed. Zap the original store and anything that
602 // feeds into it.
603 deleteDeadInstruction(TheStore, TLI);
604 ++NumMemSet;
605 return true;
606 }
607
608 /// If the stored value is a strided load in the same loop with the same stride
609 /// this may be transformable into a memcpy. This kicks in for stuff like
610 /// for (i) A[i] = B[i];
processLoopStoreOfLoopLoad(StoreInst * SI,unsigned StoreSize,const SCEVAddRecExpr * StoreEv,const SCEV * BECount,bool NegStride)611 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
612 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
613 const SCEV *BECount, bool NegStride) {
614 // If we're not allowed to form memcpy, we fail.
615 if (!TLI->has(LibFunc::memcpy))
616 return false;
617
618 // The store must be feeding a non-volatile load.
619 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
620 if (!LI || !LI->isSimple())
621 return false;
622
623 // See if the pointer expression is an AddRec like {base,+,1} on the current
624 // loop, which indicates a strided load. If we have something else, it's a
625 // random load we can't handle.
626 const SCEVAddRecExpr *LoadEv =
627 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
628 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
629 return false;
630
631 // The store and load must share the same stride.
632 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
633 return false;
634
635 // The trip count of the loop and the base pointer of the addrec SCEV is
636 // guaranteed to be loop invariant, which means that it should dominate the
637 // header. This allows us to insert code for it in the preheader.
638 BasicBlock *Preheader = CurLoop->getLoopPreheader();
639 IRBuilder<> Builder(Preheader->getTerminator());
640 SCEVExpander Expander(*SE, *DL, "loop-idiom");
641
642 const SCEV *StrStart = StoreEv->getStart();
643 unsigned StrAS = SI->getPointerAddressSpace();
644 Type *IntPtrTy = Builder.getIntPtrTy(*DL, StrAS);
645
646 // Handle negative strided loops.
647 if (NegStride)
648 StrStart = getStartForNegStride(StrStart, BECount, IntPtrTy, StoreSize, SE);
649
650 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
651 // this into a memcpy in the loop preheader now if we want. However, this
652 // would be unsafe to do if there is anything else in the loop that may read
653 // or write the memory region we're storing to. This includes the load that
654 // feeds the stores. Check for an alias by generating the base address and
655 // checking everything.
656 Value *StoreBasePtr = Expander.expandCodeFor(
657 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
658
659 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
660 StoreSize, *AA, SI)) {
661 Expander.clear();
662 // If we generated new code for the base pointer, clean up.
663 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
664 return false;
665 }
666
667 const SCEV *LdStart = LoadEv->getStart();
668 unsigned LdAS = LI->getPointerAddressSpace();
669
670 // Handle negative strided loops.
671 if (NegStride)
672 LdStart = getStartForNegStride(LdStart, BECount, IntPtrTy, StoreSize, SE);
673
674 // For a memcpy, we have to make sure that the input array is not being
675 // mutated by the loop.
676 Value *LoadBasePtr = Expander.expandCodeFor(
677 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
678
679 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
680 *AA, SI)) {
681 Expander.clear();
682 // If we generated new code for the base pointer, clean up.
683 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
684 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
685 return false;
686 }
687
688 // Okay, everything is safe, we can transform this!
689
690 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
691 // pointer size if it isn't already.
692 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
693
694 const SCEV *NumBytesS =
695 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
696 if (StoreSize != 1)
697 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
698 SCEV::FlagNUW);
699
700 Value *NumBytes =
701 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
702
703 CallInst *NewCall =
704 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
705 std::min(SI->getAlignment(), LI->getAlignment()));
706 NewCall->setDebugLoc(SI->getDebugLoc());
707
708 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
709 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
710 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
711
712 // Okay, the memcpy has been formed. Zap the original store and anything that
713 // feeds into it.
714 deleteDeadInstruction(SI, TLI);
715 ++NumMemCpy;
716 return true;
717 }
718
runOnNoncountableLoop()719 bool LoopIdiomRecognize::runOnNoncountableLoop() {
720 return recognizePopcount();
721 }
722
723 /// Check if the given conditional branch is based on the comparison between
724 /// a variable and zero, and if the variable is non-zero, the control yields to
725 /// the loop entry. If the branch matches the behavior, the variable involved
726 /// in the comparion is returned. This function will be called to see if the
727 /// precondition and postcondition of the loop are in desirable form.
matchCondition(BranchInst * BI,BasicBlock * LoopEntry)728 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
729 if (!BI || !BI->isConditional())
730 return nullptr;
731
732 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
733 if (!Cond)
734 return nullptr;
735
736 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
737 if (!CmpZero || !CmpZero->isZero())
738 return nullptr;
739
740 ICmpInst::Predicate Pred = Cond->getPredicate();
741 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
742 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
743 return Cond->getOperand(0);
744
745 return nullptr;
746 }
747
748 /// Return true iff the idiom is detected in the loop.
749 ///
750 /// Additionally:
751 /// 1) \p CntInst is set to the instruction counting the population bit.
752 /// 2) \p CntPhi is set to the corresponding phi node.
753 /// 3) \p Var is set to the value whose population bits are being counted.
754 ///
755 /// The core idiom we are trying to detect is:
756 /// \code
757 /// if (x0 != 0)
758 /// goto loop-exit // the precondition of the loop
759 /// cnt0 = init-val;
760 /// do {
761 /// x1 = phi (x0, x2);
762 /// cnt1 = phi(cnt0, cnt2);
763 ///
764 /// cnt2 = cnt1 + 1;
765 /// ...
766 /// x2 = x1 & (x1 - 1);
767 /// ...
768 /// } while(x != 0);
769 ///
770 /// loop-exit:
771 /// \endcode
detectPopcountIdiom(Loop * CurLoop,BasicBlock * PreCondBB,Instruction * & CntInst,PHINode * & CntPhi,Value * & Var)772 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
773 Instruction *&CntInst, PHINode *&CntPhi,
774 Value *&Var) {
775 // step 1: Check to see if the look-back branch match this pattern:
776 // "if (a!=0) goto loop-entry".
777 BasicBlock *LoopEntry;
778 Instruction *DefX2, *CountInst;
779 Value *VarX1, *VarX0;
780 PHINode *PhiX, *CountPhi;
781
782 DefX2 = CountInst = nullptr;
783 VarX1 = VarX0 = nullptr;
784 PhiX = CountPhi = nullptr;
785 LoopEntry = *(CurLoop->block_begin());
786
787 // step 1: Check if the loop-back branch is in desirable form.
788 {
789 if (Value *T = matchCondition(
790 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
791 DefX2 = dyn_cast<Instruction>(T);
792 else
793 return false;
794 }
795
796 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
797 {
798 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
799 return false;
800
801 BinaryOperator *SubOneOp;
802
803 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
804 VarX1 = DefX2->getOperand(1);
805 else {
806 VarX1 = DefX2->getOperand(0);
807 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
808 }
809 if (!SubOneOp)
810 return false;
811
812 Instruction *SubInst = cast<Instruction>(SubOneOp);
813 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
814 if (!Dec ||
815 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
816 (SubInst->getOpcode() == Instruction::Add &&
817 Dec->isAllOnesValue()))) {
818 return false;
819 }
820 }
821
822 // step 3: Check the recurrence of variable X
823 {
824 PhiX = dyn_cast<PHINode>(VarX1);
825 if (!PhiX ||
826 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
827 return false;
828 }
829 }
830
831 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
832 {
833 CountInst = nullptr;
834 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
835 IterE = LoopEntry->end();
836 Iter != IterE; Iter++) {
837 Instruction *Inst = &*Iter;
838 if (Inst->getOpcode() != Instruction::Add)
839 continue;
840
841 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
842 if (!Inc || !Inc->isOne())
843 continue;
844
845 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
846 if (!Phi || Phi->getParent() != LoopEntry)
847 continue;
848
849 // Check if the result of the instruction is live of the loop.
850 bool LiveOutLoop = false;
851 for (User *U : Inst->users()) {
852 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
853 LiveOutLoop = true;
854 break;
855 }
856 }
857
858 if (LiveOutLoop) {
859 CountInst = Inst;
860 CountPhi = Phi;
861 break;
862 }
863 }
864
865 if (!CountInst)
866 return false;
867 }
868
869 // step 5: check if the precondition is in this form:
870 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
871 {
872 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
873 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
874 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
875 return false;
876
877 CntInst = CountInst;
878 CntPhi = CountPhi;
879 Var = T;
880 }
881
882 return true;
883 }
884
885 /// Recognizes a population count idiom in a non-countable loop.
886 ///
887 /// If detected, transforms the relevant code to issue the popcount intrinsic
888 /// function call, and returns true; otherwise, returns false.
recognizePopcount()889 bool LoopIdiomRecognize::recognizePopcount() {
890 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
891 return false;
892
893 // Counting population are usually conducted by few arithmetic instructions.
894 // Such instructions can be easily "absorbed" by vacant slots in a
895 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
896 // in a compact loop.
897
898 // Give up if the loop has multiple blocks or multiple backedges.
899 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
900 return false;
901
902 BasicBlock *LoopBody = *(CurLoop->block_begin());
903 if (LoopBody->size() >= 20) {
904 // The loop is too big, bail out.
905 return false;
906 }
907
908 // It should have a preheader containing nothing but an unconditional branch.
909 BasicBlock *PH = CurLoop->getLoopPreheader();
910 if (!PH)
911 return false;
912 if (&PH->front() != PH->getTerminator())
913 return false;
914 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
915 if (!EntryBI || EntryBI->isConditional())
916 return false;
917
918 // It should have a precondition block where the generated popcount instrinsic
919 // function can be inserted.
920 auto *PreCondBB = PH->getSinglePredecessor();
921 if (!PreCondBB)
922 return false;
923 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
924 if (!PreCondBI || PreCondBI->isUnconditional())
925 return false;
926
927 Instruction *CntInst;
928 PHINode *CntPhi;
929 Value *Val;
930 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
931 return false;
932
933 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
934 return true;
935 }
936
createPopcntIntrinsic(IRBuilder<> & IRBuilder,Value * Val,DebugLoc DL)937 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
938 DebugLoc DL) {
939 Value *Ops[] = {Val};
940 Type *Tys[] = {Val->getType()};
941
942 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
943 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
944 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
945 CI->setDebugLoc(DL);
946
947 return CI;
948 }
949
transformLoopToPopcount(BasicBlock * PreCondBB,Instruction * CntInst,PHINode * CntPhi,Value * Var)950 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
951 Instruction *CntInst,
952 PHINode *CntPhi, Value *Var) {
953 BasicBlock *PreHead = CurLoop->getLoopPreheader();
954 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
955 const DebugLoc DL = CntInst->getDebugLoc();
956
957 // Assuming before transformation, the loop is following:
958 // if (x) // the precondition
959 // do { cnt++; x &= x - 1; } while(x);
960
961 // Step 1: Insert the ctpop instruction at the end of the precondition block
962 IRBuilder<> Builder(PreCondBr);
963 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
964 {
965 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
966 NewCount = PopCntZext =
967 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
968
969 if (NewCount != PopCnt)
970 (cast<Instruction>(NewCount))->setDebugLoc(DL);
971
972 // TripCnt is exactly the number of iterations the loop has
973 TripCnt = NewCount;
974
975 // If the population counter's initial value is not zero, insert Add Inst.
976 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
977 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
978 if (!InitConst || !InitConst->isZero()) {
979 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
980 (cast<Instruction>(NewCount))->setDebugLoc(DL);
981 }
982 }
983
984 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
985 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
986 // function would be partial dead code, and downstream passes will drag
987 // it back from the precondition block to the preheader.
988 {
989 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
990
991 Value *Opnd0 = PopCntZext;
992 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
993 if (PreCond->getOperand(0) != Var)
994 std::swap(Opnd0, Opnd1);
995
996 ICmpInst *NewPreCond = cast<ICmpInst>(
997 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
998 PreCondBr->setCondition(NewPreCond);
999
1000 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
1001 }
1002
1003 // Step 3: Note that the population count is exactly the trip count of the
1004 // loop in question, which enable us to to convert the loop from noncountable
1005 // loop into a countable one. The benefit is twofold:
1006 //
1007 // - If the loop only counts population, the entire loop becomes dead after
1008 // the transformation. It is a lot easier to prove a countable loop dead
1009 // than to prove a noncountable one. (In some C dialects, an infinite loop
1010 // isn't dead even if it computes nothing useful. In general, DCE needs
1011 // to prove a noncountable loop finite before safely delete it.)
1012 //
1013 // - If the loop also performs something else, it remains alive.
1014 // Since it is transformed to countable form, it can be aggressively
1015 // optimized by some optimizations which are in general not applicable
1016 // to a noncountable loop.
1017 //
1018 // After this step, this loop (conceptually) would look like following:
1019 // newcnt = __builtin_ctpop(x);
1020 // t = newcnt;
1021 // if (x)
1022 // do { cnt++; x &= x-1; t--) } while (t > 0);
1023 BasicBlock *Body = *(CurLoop->block_begin());
1024 {
1025 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1026 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1027 Type *Ty = TripCnt->getType();
1028
1029 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1030
1031 Builder.SetInsertPoint(LbCond);
1032 Instruction *TcDec = cast<Instruction>(
1033 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1034 "tcdec", false, true));
1035
1036 TcPhi->addIncoming(TripCnt, PreHead);
1037 TcPhi->addIncoming(TcDec, Body);
1038
1039 CmpInst::Predicate Pred =
1040 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1041 LbCond->setPredicate(Pred);
1042 LbCond->setOperand(0, TcDec);
1043 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1044 }
1045
1046 // Step 4: All the references to the original population counter outside
1047 // the loop are replaced with the NewCount -- the value returned from
1048 // __builtin_ctpop().
1049 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1050
1051 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1052 // loop. The loop would otherwise not be deleted even if it becomes empty.
1053 SE->forgetLoop(CurLoop);
1054 }
1055