1 //===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass implements an idiom recognizer that transforms simple loops into a
10 // non-loop form. In cases that this kicks in, it can be a significant
11 // performance win.
12 //
13 // If compiling for code size we avoid idiom recognition if the resulting
14 // code could be larger than the code for the original loop. One way this could
15 // happen is if the loop is not removable after idiom recognition due to the
16 // presence of non-idiom instructions. The initial implementation of the
17 // heuristics applies to idioms in multi-block loops.
18 //
19 //===----------------------------------------------------------------------===//
20 //
21 // TODO List:
22 //
23 // Future loop memory idioms to recognize:
24 // memcmp, strlen, etc.
25 // Future floating point idioms to recognize in -ffast-math mode:
26 // fpowi
27 // Future integer operation idioms to recognize:
28 // ctpop
29 //
30 // Beware that isel's default lowering for ctpop is highly inefficient for
31 // i64 and larger types when i64 is legal and the value has few bits set. It
32 // would be good to enhance isel to emit a loop for ctpop in this case.
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/LoopIdiomRecognize.h"
40 #include "llvm/ADT/APInt.h"
41 #include "llvm/ADT/ArrayRef.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/MapVector.h"
44 #include "llvm/ADT/SetVector.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/SmallVector.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/StringRef.h"
49 #include "llvm/Analysis/AliasAnalysis.h"
50 #include "llvm/Analysis/CmpInstAnalysis.h"
51 #include "llvm/Analysis/LoopAccessAnalysis.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Analysis/MemoryLocation.h"
55 #include "llvm/Analysis/MemorySSA.h"
56 #include "llvm/Analysis/MemorySSAUpdater.h"
57 #include "llvm/Analysis/MustExecute.h"
58 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
59 #include "llvm/Analysis/ScalarEvolution.h"
60 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
61 #include "llvm/Analysis/TargetLibraryInfo.h"
62 #include "llvm/Analysis/TargetTransformInfo.h"
63 #include "llvm/Analysis/ValueTracking.h"
64 #include "llvm/IR/Attributes.h"
65 #include "llvm/IR/BasicBlock.h"
66 #include "llvm/IR/Constant.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugLoc.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/GlobalValue.h"
73 #include "llvm/IR/GlobalVariable.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/InstrTypes.h"
76 #include "llvm/IR/Instruction.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/Intrinsics.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PassManager.h"
83 #include "llvm/IR/PatternMatch.h"
84 #include "llvm/IR/Type.h"
85 #include "llvm/IR/User.h"
86 #include "llvm/IR/Value.h"
87 #include "llvm/IR/ValueHandle.h"
88 #include "llvm/InitializePasses.h"
89 #include "llvm/Pass.h"
90 #include "llvm/Support/Casting.h"
91 #include "llvm/Support/CommandLine.h"
92 #include "llvm/Support/Debug.h"
93 #include "llvm/Support/InstructionCost.h"
94 #include "llvm/Support/raw_ostream.h"
95 #include "llvm/Transforms/Scalar.h"
96 #include "llvm/Transforms/Utils/BuildLibCalls.h"
97 #include "llvm/Transforms/Utils/Local.h"
98 #include "llvm/Transforms/Utils/LoopUtils.h"
99 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
100 #include <algorithm>
101 #include <cassert>
102 #include <cstdint>
103 #include <utility>
104 #include <vector>
105
106 using namespace llvm;
107
108 #define DEBUG_TYPE "loop-idiom"
109
110 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
111 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
112 STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores");
113 STATISTIC(
114 NumShiftUntilBitTest,
115 "Number of uncountable loops recognized as 'shift until bitttest' idiom");
116 STATISTIC(NumShiftUntilZero,
117 "Number of uncountable loops recognized as 'shift until zero' idiom");
118
119 bool DisableLIRP::All;
120 static cl::opt<bool, true>
121 DisableLIRPAll("disable-" DEBUG_TYPE "-all",
122 cl::desc("Options to disable Loop Idiom Recognize Pass."),
123 cl::location(DisableLIRP::All), cl::init(false),
124 cl::ReallyHidden);
125
126 bool DisableLIRP::Memset;
127 static cl::opt<bool, true>
128 DisableLIRPMemset("disable-" DEBUG_TYPE "-memset",
129 cl::desc("Proceed with loop idiom recognize pass, but do "
130 "not convert loop(s) to memset."),
131 cl::location(DisableLIRP::Memset), cl::init(false),
132 cl::ReallyHidden);
133
134 bool DisableLIRP::Memcpy;
135 static cl::opt<bool, true>
136 DisableLIRPMemcpy("disable-" DEBUG_TYPE "-memcpy",
137 cl::desc("Proceed with loop idiom recognize pass, but do "
138 "not convert loop(s) to memcpy."),
139 cl::location(DisableLIRP::Memcpy), cl::init(false),
140 cl::ReallyHidden);
141
142 static cl::opt<bool> UseLIRCodeSizeHeurs(
143 "use-lir-code-size-heurs",
144 cl::desc("Use loop idiom recognition code size heuristics when compiling"
145 "with -Os/-Oz"),
146 cl::init(true), cl::Hidden);
147
148 namespace {
149
150 class LoopIdiomRecognize {
151 Loop *CurLoop = nullptr;
152 AliasAnalysis *AA;
153 DominatorTree *DT;
154 LoopInfo *LI;
155 ScalarEvolution *SE;
156 TargetLibraryInfo *TLI;
157 const TargetTransformInfo *TTI;
158 const DataLayout *DL;
159 OptimizationRemarkEmitter &ORE;
160 bool ApplyCodeSizeHeuristics;
161 std::unique_ptr<MemorySSAUpdater> MSSAU;
162
163 public:
LoopIdiomRecognize(AliasAnalysis * AA,DominatorTree * DT,LoopInfo * LI,ScalarEvolution * SE,TargetLibraryInfo * TLI,const TargetTransformInfo * TTI,MemorySSA * MSSA,const DataLayout * DL,OptimizationRemarkEmitter & ORE)164 explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
165 LoopInfo *LI, ScalarEvolution *SE,
166 TargetLibraryInfo *TLI,
167 const TargetTransformInfo *TTI, MemorySSA *MSSA,
168 const DataLayout *DL,
169 OptimizationRemarkEmitter &ORE)
170 : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) {
171 if (MSSA)
172 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
173 }
174
175 bool runOnLoop(Loop *L);
176
177 private:
178 using StoreList = SmallVector<StoreInst *, 8>;
179 using StoreListMap = MapVector<Value *, StoreList>;
180
181 StoreListMap StoreRefsForMemset;
182 StoreListMap StoreRefsForMemsetPattern;
183 StoreList StoreRefsForMemcpy;
184 bool HasMemset;
185 bool HasMemsetPattern;
186 bool HasMemcpy;
187
188 /// Return code for isLegalStore()
189 enum LegalStoreKind {
190 None = 0,
191 Memset,
192 MemsetPattern,
193 Memcpy,
194 UnorderedAtomicMemcpy,
195 DontUse // Dummy retval never to be used. Allows catching errors in retval
196 // handling.
197 };
198
199 /// \name Countable Loop Idiom Handling
200 /// @{
201
202 bool runOnCountableLoop();
203 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
204 SmallVectorImpl<BasicBlock *> &ExitBlocks);
205
206 void collectStores(BasicBlock *BB);
207 LegalStoreKind isLegalStore(StoreInst *SI);
208 enum class ForMemset { No, Yes };
209 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
210 ForMemset For);
211
212 template <typename MemInst>
213 bool processLoopMemIntrinsic(
214 BasicBlock *BB,
215 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
216 const SCEV *BECount);
217 bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount);
218 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
219
220 bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV,
221 MaybeAlign StoreAlignment, Value *StoredVal,
222 Instruction *TheStore,
223 SmallPtrSetImpl<Instruction *> &Stores,
224 const SCEVAddRecExpr *Ev, const SCEV *BECount,
225 bool IsNegStride, bool IsLoopMemset = false);
226 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
227 bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr,
228 const SCEV *StoreSize, MaybeAlign StoreAlign,
229 MaybeAlign LoadAlign, Instruction *TheStore,
230 Instruction *TheLoad,
231 const SCEVAddRecExpr *StoreEv,
232 const SCEVAddRecExpr *LoadEv,
233 const SCEV *BECount);
234 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
235 bool IsLoopMemset = false);
236
237 /// @}
238 /// \name Noncountable Loop Idiom Handling
239 /// @{
240
241 bool runOnNoncountableLoop();
242
243 bool recognizePopcount();
244 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
245 PHINode *CntPhi, Value *Var);
246 bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz
247 void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB,
248 Instruction *CntInst, PHINode *CntPhi,
249 Value *Var, Instruction *DefX,
250 const DebugLoc &DL, bool ZeroCheck,
251 bool IsCntPhiUsedOutsideLoop);
252
253 bool recognizeShiftUntilBitTest();
254 bool recognizeShiftUntilZero();
255
256 /// @}
257 };
258
259 class LoopIdiomRecognizeLegacyPass : public LoopPass {
260 public:
261 static char ID;
262
LoopIdiomRecognizeLegacyPass()263 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
264 initializeLoopIdiomRecognizeLegacyPassPass(
265 *PassRegistry::getPassRegistry());
266 }
267
runOnLoop(Loop * L,LPPassManager & LPM)268 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
269 if (DisableLIRP::All)
270 return false;
271
272 if (skipLoop(L))
273 return false;
274
275 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
276 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
277 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
278 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
279 TargetLibraryInfo *TLI =
280 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
281 *L->getHeader()->getParent());
282 const TargetTransformInfo *TTI =
283 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
284 *L->getHeader()->getParent());
285 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
286 auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>();
287 MemorySSA *MSSA = nullptr;
288 if (MSSAAnalysis)
289 MSSA = &MSSAAnalysis->getMSSA();
290
291 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
292 // pass. Function analyses need to be preserved across loop transformations
293 // but ORE cannot be preserved (see comment before the pass definition).
294 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
295
296 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE);
297 return LIR.runOnLoop(L);
298 }
299
300 /// This transformation requires natural loop information & requires that
301 /// loop preheaders be inserted into the CFG.
getAnalysisUsage(AnalysisUsage & AU) const302 void getAnalysisUsage(AnalysisUsage &AU) const override {
303 AU.addRequired<TargetLibraryInfoWrapperPass>();
304 AU.addRequired<TargetTransformInfoWrapperPass>();
305 AU.addPreserved<MemorySSAWrapperPass>();
306 getLoopAnalysisUsage(AU);
307 }
308 };
309
310 } // end anonymous namespace
311
312 char LoopIdiomRecognizeLegacyPass::ID = 0;
313
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater &)314 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
315 LoopStandardAnalysisResults &AR,
316 LPMUpdater &) {
317 if (DisableLIRP::All)
318 return PreservedAnalyses::all();
319
320 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
321
322 // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
323 // pass. Function analyses need to be preserved across loop transformations
324 // but ORE cannot be preserved (see comment before the pass definition).
325 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
326
327 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI,
328 AR.MSSA, DL, ORE);
329 if (!LIR.runOnLoop(&L))
330 return PreservedAnalyses::all();
331
332 auto PA = getLoopPassPreservedAnalyses();
333 if (AR.MSSA)
334 PA.preserve<MemorySSAAnalysis>();
335 return PA;
336 }
337
338 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
339 "Recognize loop idioms", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)340 INITIALIZE_PASS_DEPENDENCY(LoopPass)
341 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
342 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
343 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
344 "Recognize loop idioms", false, false)
345
346 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
347
deleteDeadInstruction(Instruction * I)348 static void deleteDeadInstruction(Instruction *I) {
349 I->replaceAllUsesWith(UndefValue::get(I->getType()));
350 I->eraseFromParent();
351 }
352
353 //===----------------------------------------------------------------------===//
354 //
355 // Implementation of LoopIdiomRecognize
356 //
357 //===----------------------------------------------------------------------===//
358
runOnLoop(Loop * L)359 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
360 CurLoop = L;
361 // If the loop could not be converted to canonical form, it must have an
362 // indirectbr in it, just give up.
363 if (!L->getLoopPreheader())
364 return false;
365
366 // Disable loop idiom recognition if the function's name is a common idiom.
367 StringRef Name = L->getHeader()->getParent()->getName();
368 if (Name == "memset" || Name == "memcpy")
369 return false;
370
371 // Determine if code size heuristics need to be applied.
372 ApplyCodeSizeHeuristics =
373 L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs;
374
375 HasMemset = TLI->has(LibFunc_memset);
376 HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
377 HasMemcpy = TLI->has(LibFunc_memcpy);
378
379 if (HasMemset || HasMemsetPattern || HasMemcpy)
380 if (SE->hasLoopInvariantBackedgeTakenCount(L))
381 return runOnCountableLoop();
382
383 return runOnNoncountableLoop();
384 }
385
runOnCountableLoop()386 bool LoopIdiomRecognize::runOnCountableLoop() {
387 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
388 assert(!isa<SCEVCouldNotCompute>(BECount) &&
389 "runOnCountableLoop() called on a loop without a predictable"
390 "backedge-taken count");
391
392 // If this loop executes exactly one time, then it should be peeled, not
393 // optimized by this pass.
394 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
395 if (BECst->getAPInt() == 0)
396 return false;
397
398 SmallVector<BasicBlock *, 8> ExitBlocks;
399 CurLoop->getUniqueExitBlocks(ExitBlocks);
400
401 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
402 << CurLoop->getHeader()->getParent()->getName()
403 << "] Countable Loop %" << CurLoop->getHeader()->getName()
404 << "\n");
405
406 // The following transforms hoist stores/memsets into the loop pre-header.
407 // Give up if the loop has instructions that may throw.
408 SimpleLoopSafetyInfo SafetyInfo;
409 SafetyInfo.computeLoopSafetyInfo(CurLoop);
410 if (SafetyInfo.anyBlockMayThrow())
411 return false;
412
413 bool MadeChange = false;
414
415 // Scan all the blocks in the loop that are not in subloops.
416 for (auto *BB : CurLoop->getBlocks()) {
417 // Ignore blocks in subloops.
418 if (LI->getLoopFor(BB) != CurLoop)
419 continue;
420
421 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
422 }
423 return MadeChange;
424 }
425
getStoreStride(const SCEVAddRecExpr * StoreEv)426 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
427 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
428 return ConstStride->getAPInt();
429 }
430
431 /// getMemSetPatternValue - If a strided store of the specified value is safe to
432 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
433 /// be passed in. Otherwise, return null.
434 ///
435 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
436 /// just replicate their input array and then pass on to memset_pattern16.
getMemSetPatternValue(Value * V,const DataLayout * DL)437 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
438 // FIXME: This could check for UndefValue because it can be merged into any
439 // other valid pattern.
440
441 // If the value isn't a constant, we can't promote it to being in a constant
442 // array. We could theoretically do a store to an alloca or something, but
443 // that doesn't seem worthwhile.
444 Constant *C = dyn_cast<Constant>(V);
445 if (!C)
446 return nullptr;
447
448 // Only handle simple values that are a power of two bytes in size.
449 uint64_t Size = DL->getTypeSizeInBits(V->getType());
450 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
451 return nullptr;
452
453 // Don't care enough about darwin/ppc to implement this.
454 if (DL->isBigEndian())
455 return nullptr;
456
457 // Convert to size in bytes.
458 Size /= 8;
459
460 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
461 // if the top and bottom are the same (e.g. for vectors and large integers).
462 if (Size > 16)
463 return nullptr;
464
465 // If the constant is exactly 16 bytes, just use it.
466 if (Size == 16)
467 return C;
468
469 // Otherwise, we'll use an array of the constants.
470 unsigned ArraySize = 16 / Size;
471 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
472 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
473 }
474
475 LoopIdiomRecognize::LegalStoreKind
isLegalStore(StoreInst * SI)476 LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
477 // Don't touch volatile stores.
478 if (SI->isVolatile())
479 return LegalStoreKind::None;
480 // We only want simple or unordered-atomic stores.
481 if (!SI->isUnordered())
482 return LegalStoreKind::None;
483
484 // Avoid merging nontemporal stores.
485 if (SI->getMetadata(LLVMContext::MD_nontemporal))
486 return LegalStoreKind::None;
487
488 Value *StoredVal = SI->getValueOperand();
489 Value *StorePtr = SI->getPointerOperand();
490
491 // Don't convert stores of non-integral pointer types to memsets (which stores
492 // integers).
493 if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
494 return LegalStoreKind::None;
495
496 // Reject stores that are so large that they overflow an unsigned.
497 // When storing out scalable vectors we bail out for now, since the code
498 // below currently only works for constant strides.
499 TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
500 if (SizeInBits.isScalable() || (SizeInBits.getFixedSize() & 7) ||
501 (SizeInBits.getFixedSize() >> 32) != 0)
502 return LegalStoreKind::None;
503
504 // See if the pointer expression is an AddRec like {base,+,1} on the current
505 // loop, which indicates a strided store. If we have something else, it's a
506 // random store we can't handle.
507 const SCEVAddRecExpr *StoreEv =
508 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
509 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
510 return LegalStoreKind::None;
511
512 // Check to see if we have a constant stride.
513 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
514 return LegalStoreKind::None;
515
516 // See if the store can be turned into a memset.
517
518 // If the stored value is a byte-wise value (like i32 -1), then it may be
519 // turned into a memset of i8 -1, assuming that all the consecutive bytes
520 // are stored. A store of i32 0x01020304 can never be turned into a memset,
521 // but it can be turned into memset_pattern if the target supports it.
522 Value *SplatValue = isBytewiseValue(StoredVal, *DL);
523
524 // Note: memset and memset_pattern on unordered-atomic is yet not supported
525 bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple();
526
527 // If we're allowed to form a memset, and the stored value would be
528 // acceptable for memset, use it.
529 if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset &&
530 // Verify that the stored value is loop invariant. If not, we can't
531 // promote the memset.
532 CurLoop->isLoopInvariant(SplatValue)) {
533 // It looks like we can use SplatValue.
534 return LegalStoreKind::Memset;
535 }
536 if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset &&
537 // Don't create memset_pattern16s with address spaces.
538 StorePtr->getType()->getPointerAddressSpace() == 0 &&
539 getMemSetPatternValue(StoredVal, DL)) {
540 // It looks like we can use PatternValue!
541 return LegalStoreKind::MemsetPattern;
542 }
543
544 // Otherwise, see if the store can be turned into a memcpy.
545 if (HasMemcpy && !DisableLIRP::Memcpy) {
546 // Check to see if the stride matches the size of the store. If so, then we
547 // know that every byte is touched in the loop.
548 APInt Stride = getStoreStride(StoreEv);
549 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
550 if (StoreSize != Stride && StoreSize != -Stride)
551 return LegalStoreKind::None;
552
553 // The store must be feeding a non-volatile load.
554 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
555
556 // Only allow non-volatile loads
557 if (!LI || LI->isVolatile())
558 return LegalStoreKind::None;
559 // Only allow simple or unordered-atomic loads
560 if (!LI->isUnordered())
561 return LegalStoreKind::None;
562
563 // See if the pointer expression is an AddRec like {base,+,1} on the current
564 // loop, which indicates a strided load. If we have something else, it's a
565 // random load we can't handle.
566 const SCEVAddRecExpr *LoadEv =
567 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
568 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
569 return LegalStoreKind::None;
570
571 // The store and load must share the same stride.
572 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
573 return LegalStoreKind::None;
574
575 // Success. This store can be converted into a memcpy.
576 UnorderedAtomic = UnorderedAtomic || LI->isAtomic();
577 return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy
578 : LegalStoreKind::Memcpy;
579 }
580 // This store can't be transformed into a memset/memcpy.
581 return LegalStoreKind::None;
582 }
583
collectStores(BasicBlock * BB)584 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
585 StoreRefsForMemset.clear();
586 StoreRefsForMemsetPattern.clear();
587 StoreRefsForMemcpy.clear();
588 for (Instruction &I : *BB) {
589 StoreInst *SI = dyn_cast<StoreInst>(&I);
590 if (!SI)
591 continue;
592
593 // Make sure this is a strided store with a constant stride.
594 switch (isLegalStore(SI)) {
595 case LegalStoreKind::None:
596 // Nothing to do
597 break;
598 case LegalStoreKind::Memset: {
599 // Find the base pointer.
600 Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
601 StoreRefsForMemset[Ptr].push_back(SI);
602 } break;
603 case LegalStoreKind::MemsetPattern: {
604 // Find the base pointer.
605 Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
606 StoreRefsForMemsetPattern[Ptr].push_back(SI);
607 } break;
608 case LegalStoreKind::Memcpy:
609 case LegalStoreKind::UnorderedAtomicMemcpy:
610 StoreRefsForMemcpy.push_back(SI);
611 break;
612 default:
613 assert(false && "unhandled return value");
614 break;
615 }
616 }
617 }
618
619 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
620 /// with the specified backedge count. This block is known to be in the current
621 /// loop and not in any subloops.
runOnLoopBlock(BasicBlock * BB,const SCEV * BECount,SmallVectorImpl<BasicBlock * > & ExitBlocks)622 bool LoopIdiomRecognize::runOnLoopBlock(
623 BasicBlock *BB, const SCEV *BECount,
624 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
625 // We can only promote stores in this block if they are unconditionally
626 // executed in the loop. For a block to be unconditionally executed, it has
627 // to dominate all the exit blocks of the loop. Verify this now.
628 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
629 if (!DT->dominates(BB, ExitBlocks[i]))
630 return false;
631
632 bool MadeChange = false;
633 // Look for store instructions, which may be optimized to memset/memcpy.
634 collectStores(BB);
635
636 // Look for a single store or sets of stores with a common base, which can be
637 // optimized into a memset (memset_pattern). The latter most commonly happens
638 // with structs and handunrolled loops.
639 for (auto &SL : StoreRefsForMemset)
640 MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes);
641
642 for (auto &SL : StoreRefsForMemsetPattern)
643 MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No);
644
645 // Optimize the store into a memcpy, if it feeds an similarly strided load.
646 for (auto &SI : StoreRefsForMemcpy)
647 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
648
649 MadeChange |= processLoopMemIntrinsic<MemCpyInst>(
650 BB, &LoopIdiomRecognize::processLoopMemCpy, BECount);
651 MadeChange |= processLoopMemIntrinsic<MemSetInst>(
652 BB, &LoopIdiomRecognize::processLoopMemSet, BECount);
653
654 return MadeChange;
655 }
656
657 /// See if this store(s) can be promoted to a memset.
processLoopStores(SmallVectorImpl<StoreInst * > & SL,const SCEV * BECount,ForMemset For)658 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
659 const SCEV *BECount, ForMemset For) {
660 // Try to find consecutive stores that can be transformed into memsets.
661 SetVector<StoreInst *> Heads, Tails;
662 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
663
664 // Do a quadratic search on all of the given stores and find
665 // all of the pairs of stores that follow each other.
666 SmallVector<unsigned, 16> IndexQueue;
667 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
668 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
669
670 Value *FirstStoredVal = SL[i]->getValueOperand();
671 Value *FirstStorePtr = SL[i]->getPointerOperand();
672 const SCEVAddRecExpr *FirstStoreEv =
673 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
674 APInt FirstStride = getStoreStride(FirstStoreEv);
675 unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType());
676
677 // See if we can optimize just this store in isolation.
678 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
679 Heads.insert(SL[i]);
680 continue;
681 }
682
683 Value *FirstSplatValue = nullptr;
684 Constant *FirstPatternValue = nullptr;
685
686 if (For == ForMemset::Yes)
687 FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL);
688 else
689 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
690
691 assert((FirstSplatValue || FirstPatternValue) &&
692 "Expected either splat value or pattern value.");
693
694 IndexQueue.clear();
695 // If a store has multiple consecutive store candidates, search Stores
696 // array according to the sequence: from i+1 to e, then from i-1 to 0.
697 // This is because usually pairing with immediate succeeding or preceding
698 // candidate create the best chance to find memset opportunity.
699 unsigned j = 0;
700 for (j = i + 1; j < e; ++j)
701 IndexQueue.push_back(j);
702 for (j = i; j > 0; --j)
703 IndexQueue.push_back(j - 1);
704
705 for (auto &k : IndexQueue) {
706 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
707 Value *SecondStorePtr = SL[k]->getPointerOperand();
708 const SCEVAddRecExpr *SecondStoreEv =
709 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
710 APInt SecondStride = getStoreStride(SecondStoreEv);
711
712 if (FirstStride != SecondStride)
713 continue;
714
715 Value *SecondStoredVal = SL[k]->getValueOperand();
716 Value *SecondSplatValue = nullptr;
717 Constant *SecondPatternValue = nullptr;
718
719 if (For == ForMemset::Yes)
720 SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL);
721 else
722 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
723
724 assert((SecondSplatValue || SecondPatternValue) &&
725 "Expected either splat value or pattern value.");
726
727 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
728 if (For == ForMemset::Yes) {
729 if (isa<UndefValue>(FirstSplatValue))
730 FirstSplatValue = SecondSplatValue;
731 if (FirstSplatValue != SecondSplatValue)
732 continue;
733 } else {
734 if (isa<UndefValue>(FirstPatternValue))
735 FirstPatternValue = SecondPatternValue;
736 if (FirstPatternValue != SecondPatternValue)
737 continue;
738 }
739 Tails.insert(SL[k]);
740 Heads.insert(SL[i]);
741 ConsecutiveChain[SL[i]] = SL[k];
742 break;
743 }
744 }
745 }
746
747 // We may run into multiple chains that merge into a single chain. We mark the
748 // stores that we transformed so that we don't visit the same store twice.
749 SmallPtrSet<Value *, 16> TransformedStores;
750 bool Changed = false;
751
752 // For stores that start but don't end a link in the chain:
753 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
754 it != e; ++it) {
755 if (Tails.count(*it))
756 continue;
757
758 // We found a store instr that starts a chain. Now follow the chain and try
759 // to transform it.
760 SmallPtrSet<Instruction *, 8> AdjacentStores;
761 StoreInst *I = *it;
762
763 StoreInst *HeadStore = I;
764 unsigned StoreSize = 0;
765
766 // Collect the chain into a list.
767 while (Tails.count(I) || Heads.count(I)) {
768 if (TransformedStores.count(I))
769 break;
770 AdjacentStores.insert(I);
771
772 StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType());
773 // Move to the next value in the chain.
774 I = ConsecutiveChain[I];
775 }
776
777 Value *StoredVal = HeadStore->getValueOperand();
778 Value *StorePtr = HeadStore->getPointerOperand();
779 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
780 APInt Stride = getStoreStride(StoreEv);
781
782 // Check to see if the stride matches the size of the stores. If so, then
783 // we know that every byte is touched in the loop.
784 if (StoreSize != Stride && StoreSize != -Stride)
785 continue;
786
787 bool IsNegStride = StoreSize == -Stride;
788
789 Type *IntIdxTy = DL->getIndexType(StorePtr->getType());
790 const SCEV *StoreSizeSCEV = SE->getConstant(IntIdxTy, StoreSize);
791 if (processLoopStridedStore(StorePtr, StoreSizeSCEV,
792 MaybeAlign(HeadStore->getAlignment()),
793 StoredVal, HeadStore, AdjacentStores, StoreEv,
794 BECount, IsNegStride)) {
795 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
796 Changed = true;
797 }
798 }
799
800 return Changed;
801 }
802
803 /// processLoopMemIntrinsic - Template function for calling different processor
804 /// functions based on mem instrinsic type.
805 template <typename MemInst>
processLoopMemIntrinsic(BasicBlock * BB,bool (LoopIdiomRecognize::* Processor)(MemInst *,const SCEV *),const SCEV * BECount)806 bool LoopIdiomRecognize::processLoopMemIntrinsic(
807 BasicBlock *BB,
808 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
809 const SCEV *BECount) {
810 bool MadeChange = false;
811 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
812 Instruction *Inst = &*I++;
813 // Look for memory instructions, which may be optimized to a larger one.
814 if (MemInst *MI = dyn_cast<MemInst>(Inst)) {
815 WeakTrackingVH InstPtr(&*I);
816 if (!(this->*Processor)(MI, BECount))
817 continue;
818 MadeChange = true;
819
820 // If processing the instruction invalidated our iterator, start over from
821 // the top of the block.
822 if (!InstPtr)
823 I = BB->begin();
824 }
825 }
826 return MadeChange;
827 }
828
829 /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy
processLoopMemCpy(MemCpyInst * MCI,const SCEV * BECount)830 bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
831 const SCEV *BECount) {
832 // We can only handle non-volatile memcpys with a constant size.
833 if (MCI->isVolatile() || !isa<ConstantInt>(MCI->getLength()))
834 return false;
835
836 // If we're not allowed to hack on memcpy, we fail.
837 if ((!HasMemcpy && !isa<MemCpyInlineInst>(MCI)) || DisableLIRP::Memcpy)
838 return false;
839
840 Value *Dest = MCI->getDest();
841 Value *Source = MCI->getSource();
842 if (!Dest || !Source)
843 return false;
844
845 // See if the load and store pointer expressions are AddRec like {base,+,1} on
846 // the current loop, which indicates a strided load and store. If we have
847 // something else, it's a random load or store we can't handle.
848 const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest));
849 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
850 return false;
851 const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source));
852 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
853 return false;
854
855 // Reject memcpys that are so large that they overflow an unsigned.
856 uint64_t SizeInBytes = cast<ConstantInt>(MCI->getLength())->getZExtValue();
857 if ((SizeInBytes >> 32) != 0)
858 return false;
859
860 // Check if the stride matches the size of the memcpy. If so, then we know
861 // that every byte is touched in the loop.
862 const SCEVConstant *ConstStoreStride =
863 dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
864 const SCEVConstant *ConstLoadStride =
865 dyn_cast<SCEVConstant>(LoadEv->getOperand(1));
866 if (!ConstStoreStride || !ConstLoadStride)
867 return false;
868
869 APInt StoreStrideValue = ConstStoreStride->getAPInt();
870 APInt LoadStrideValue = ConstLoadStride->getAPInt();
871 // Huge stride value - give up
872 if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64)
873 return false;
874
875 if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) {
876 ORE.emit([&]() {
877 return OptimizationRemarkMissed(DEBUG_TYPE, "SizeStrideUnequal", MCI)
878 << ore::NV("Inst", "memcpy") << " in "
879 << ore::NV("Function", MCI->getFunction())
880 << " function will not be hoisted: "
881 << ore::NV("Reason", "memcpy size is not equal to stride");
882 });
883 return false;
884 }
885
886 int64_t StoreStrideInt = StoreStrideValue.getSExtValue();
887 int64_t LoadStrideInt = LoadStrideValue.getSExtValue();
888 // Check if the load stride matches the store stride.
889 if (StoreStrideInt != LoadStrideInt)
890 return false;
891
892 return processLoopStoreOfLoopLoad(
893 Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes),
894 MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv,
895 BECount);
896 }
897
898 /// processLoopMemSet - See if this memset can be promoted to a large memset.
processLoopMemSet(MemSetInst * MSI,const SCEV * BECount)899 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
900 const SCEV *BECount) {
901 // We can only handle non-volatile memsets.
902 if (MSI->isVolatile())
903 return false;
904
905 // If we're not allowed to hack on memset, we fail.
906 if (!HasMemset || DisableLIRP::Memset)
907 return false;
908
909 Value *Pointer = MSI->getDest();
910
911 // See if the pointer expression is an AddRec like {base,+,1} on the current
912 // loop, which indicates a strided store. If we have something else, it's a
913 // random store we can't handle.
914 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
915 if (!Ev || Ev->getLoop() != CurLoop)
916 return false;
917 if (!Ev->isAffine()) {
918 LLVM_DEBUG(dbgs() << " Pointer is not affine, abort\n");
919 return false;
920 }
921
922 const SCEV *PointerStrideSCEV = Ev->getOperand(1);
923 const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength());
924 if (!PointerStrideSCEV || !MemsetSizeSCEV)
925 return false;
926
927 bool IsNegStride = false;
928 const bool IsConstantSize = isa<ConstantInt>(MSI->getLength());
929
930 if (IsConstantSize) {
931 // Memset size is constant.
932 // Check if the pointer stride matches the memset size. If so, then
933 // we know that every byte is touched in the loop.
934 LLVM_DEBUG(dbgs() << " memset size is constant\n");
935 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
936 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
937 if (!ConstStride)
938 return false;
939
940 APInt Stride = ConstStride->getAPInt();
941 if (SizeInBytes != Stride && SizeInBytes != -Stride)
942 return false;
943
944 IsNegStride = SizeInBytes == -Stride;
945 } else {
946 // Memset size is non-constant.
947 // Check if the pointer stride matches the memset size.
948 // To be conservative, the pass would not promote pointers that aren't in
949 // address space zero. Also, the pass only handles memset length and stride
950 // that are invariant for the top level loop.
951 LLVM_DEBUG(dbgs() << " memset size is non-constant\n");
952 if (Pointer->getType()->getPointerAddressSpace() != 0) {
953 LLVM_DEBUG(dbgs() << " pointer is not in address space zero, "
954 << "abort\n");
955 return false;
956 }
957 if (!SE->isLoopInvariant(MemsetSizeSCEV, CurLoop)) {
958 LLVM_DEBUG(dbgs() << " memset size is not a loop-invariant, "
959 << "abort\n");
960 return false;
961 }
962
963 // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV
964 IsNegStride = PointerStrideSCEV->isNonConstantNegative();
965 const SCEV *PositiveStrideSCEV =
966 IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV)
967 : PointerStrideSCEV;
968 LLVM_DEBUG(dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"
969 << " PositiveStrideSCEV: " << *PositiveStrideSCEV
970 << "\n");
971
972 if (PositiveStrideSCEV != MemsetSizeSCEV) {
973 // TODO: folding can be done to the SCEVs
974 // The folding is to fold expressions that is covered by the loop guard
975 // at loop entry. After the folding, compare again and proceed
976 // optimization if equal.
977 LLVM_DEBUG(dbgs() << " SCEV don't match, abort\n");
978 return false;
979 }
980 }
981
982 // Verify that the memset value is loop invariant. If not, we can't promote
983 // the memset.
984 Value *SplatValue = MSI->getValue();
985 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
986 return false;
987
988 SmallPtrSet<Instruction *, 1> MSIs;
989 MSIs.insert(MSI);
990 return processLoopStridedStore(Pointer, SE->getSCEV(MSI->getLength()),
991 MaybeAlign(MSI->getDestAlignment()),
992 SplatValue, MSI, MSIs, Ev, BECount,
993 IsNegStride, /*IsLoopMemset=*/true);
994 }
995
996 /// mayLoopAccessLocation - Return true if the specified loop might access the
997 /// specified pointer location, which is a loop-strided access. The 'Access'
998 /// argument specifies what the verboten forms of access are (read or write).
999 static bool
mayLoopAccessLocation(Value * Ptr,ModRefInfo Access,Loop * L,const SCEV * BECount,const SCEV * StoreSizeSCEV,AliasAnalysis & AA,SmallPtrSetImpl<Instruction * > & IgnoredInsts)1000 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
1001 const SCEV *BECount, const SCEV *StoreSizeSCEV,
1002 AliasAnalysis &AA,
1003 SmallPtrSetImpl<Instruction *> &IgnoredInsts) {
1004 // Get the location that may be stored across the loop. Since the access is
1005 // strided positively through memory, we say that the modified location starts
1006 // at the pointer and has infinite size.
1007 LocationSize AccessSize = LocationSize::afterPointer();
1008
1009 // If the loop iterates a fixed number of times, we can refine the access size
1010 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
1011 const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount);
1012 const SCEVConstant *ConstSize = dyn_cast<SCEVConstant>(StoreSizeSCEV);
1013 if (BECst && ConstSize)
1014 AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) *
1015 ConstSize->getValue()->getZExtValue());
1016
1017 // TODO: For this to be really effective, we have to dive into the pointer
1018 // operand in the store. Store to &A[i] of 100 will always return may alias
1019 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
1020 // which will then no-alias a store to &A[100].
1021 MemoryLocation StoreLoc(Ptr, AccessSize);
1022
1023 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
1024 ++BI)
1025 for (Instruction &I : **BI)
1026 if (IgnoredInsts.count(&I) == 0 &&
1027 isModOrRefSet(
1028 intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access)))
1029 return true;
1030 return false;
1031 }
1032
1033 // If we have a negative stride, Start refers to the end of the memory location
1034 // we're trying to memset. Therefore, we need to recompute the base pointer,
1035 // which is just Start - BECount*Size.
getStartForNegStride(const SCEV * Start,const SCEV * BECount,Type * IntPtr,const SCEV * StoreSizeSCEV,ScalarEvolution * SE)1036 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
1037 Type *IntPtr, const SCEV *StoreSizeSCEV,
1038 ScalarEvolution *SE) {
1039 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
1040 if (!StoreSizeSCEV->isOne()) {
1041 // index = back edge count * store size
1042 Index = SE->getMulExpr(Index,
1043 SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr),
1044 SCEV::FlagNUW);
1045 }
1046 // base pointer = start - index * store size
1047 return SE->getMinusSCEV(Start, Index);
1048 }
1049
1050 /// Compute trip count from the backedge taken count.
getTripCount(const SCEV * BECount,Type * IntPtr,Loop * CurLoop,const DataLayout * DL,ScalarEvolution * SE)1051 static const SCEV *getTripCount(const SCEV *BECount, Type *IntPtr,
1052 Loop *CurLoop, const DataLayout *DL,
1053 ScalarEvolution *SE) {
1054 const SCEV *TripCountS = nullptr;
1055 // The # stored bytes is (BECount+1). Expand the trip count out to
1056 // pointer size if it isn't already.
1057 //
1058 // If we're going to need to zero extend the BE count, check if we can add
1059 // one to it prior to zero extending without overflow. Provided this is safe,
1060 // it allows better simplification of the +1.
1061 if (DL->getTypeSizeInBits(BECount->getType()) <
1062 DL->getTypeSizeInBits(IntPtr) &&
1063 SE->isLoopEntryGuardedByCond(
1064 CurLoop, ICmpInst::ICMP_NE, BECount,
1065 SE->getNegativeSCEV(SE->getOne(BECount->getType())))) {
1066 TripCountS = SE->getZeroExtendExpr(
1067 SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW),
1068 IntPtr);
1069 } else {
1070 TripCountS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr),
1071 SE->getOne(IntPtr), SCEV::FlagNUW);
1072 }
1073
1074 return TripCountS;
1075 }
1076
1077 /// Compute the number of bytes as a SCEV from the backedge taken count.
1078 ///
1079 /// This also maps the SCEV into the provided type and tries to handle the
1080 /// computation in a way that will fold cleanly.
getNumBytes(const SCEV * BECount,Type * IntPtr,const SCEV * StoreSizeSCEV,Loop * CurLoop,const DataLayout * DL,ScalarEvolution * SE)1081 static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr,
1082 const SCEV *StoreSizeSCEV, Loop *CurLoop,
1083 const DataLayout *DL, ScalarEvolution *SE) {
1084 const SCEV *TripCountSCEV = getTripCount(BECount, IntPtr, CurLoop, DL, SE);
1085
1086 return SE->getMulExpr(TripCountSCEV,
1087 SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr),
1088 SCEV::FlagNUW);
1089 }
1090
1091 /// processLoopStridedStore - We see a strided store of some value. If we can
1092 /// transform this into a memset or memset_pattern in the loop preheader, do so.
processLoopStridedStore(Value * DestPtr,const SCEV * StoreSizeSCEV,MaybeAlign StoreAlignment,Value * StoredVal,Instruction * TheStore,SmallPtrSetImpl<Instruction * > & Stores,const SCEVAddRecExpr * Ev,const SCEV * BECount,bool IsNegStride,bool IsLoopMemset)1093 bool LoopIdiomRecognize::processLoopStridedStore(
1094 Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment,
1095 Value *StoredVal, Instruction *TheStore,
1096 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
1097 const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) {
1098 Value *SplatValue = isBytewiseValue(StoredVal, *DL);
1099 Constant *PatternValue = nullptr;
1100
1101 if (!SplatValue)
1102 PatternValue = getMemSetPatternValue(StoredVal, DL);
1103
1104 assert((SplatValue || PatternValue) &&
1105 "Expected either splat value or pattern value.");
1106
1107 // The trip count of the loop and the base pointer of the addrec SCEV is
1108 // guaranteed to be loop invariant, which means that it should dominate the
1109 // header. This allows us to insert code for it in the preheader.
1110 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
1111 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1112 IRBuilder<> Builder(Preheader->getTerminator());
1113 SCEVExpander Expander(*SE, *DL, "loop-idiom");
1114 SCEVExpanderCleaner ExpCleaner(Expander, *DT);
1115
1116 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
1117 Type *IntIdxTy = DL->getIndexType(DestPtr->getType());
1118
1119 bool Changed = false;
1120 const SCEV *Start = Ev->getStart();
1121 // Handle negative strided loops.
1122 if (IsNegStride)
1123 Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSizeSCEV, SE);
1124
1125 // TODO: ideally we should still be able to generate memset if SCEV expander
1126 // is taught to generate the dependencies at the latest point.
1127 if (!isSafeToExpand(Start, *SE))
1128 return Changed;
1129
1130 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
1131 // this into a memset in the loop preheader now if we want. However, this
1132 // would be unsafe to do if there is anything else in the loop that may read
1133 // or write to the aliased location. Check for any overlap by generating the
1134 // base pointer and checking the region.
1135 Value *BasePtr =
1136 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
1137
1138 // From here on out, conservatively report to the pass manager that we've
1139 // changed the IR, even if we later clean up these added instructions. There
1140 // may be structural differences e.g. in the order of use lists not accounted
1141 // for in just a textual dump of the IR. This is written as a variable, even
1142 // though statically all the places this dominates could be replaced with
1143 // 'true', with the hope that anyone trying to be clever / "more precise" with
1144 // the return value will read this comment, and leave them alone.
1145 Changed = true;
1146
1147 if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1148 StoreSizeSCEV, *AA, Stores))
1149 return Changed;
1150
1151 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
1152 return Changed;
1153
1154 // Okay, everything looks good, insert the memset.
1155
1156 const SCEV *NumBytesS =
1157 getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1158
1159 // TODO: ideally we should still be able to generate memset if SCEV expander
1160 // is taught to generate the dependencies at the latest point.
1161 if (!isSafeToExpand(NumBytesS, *SE))
1162 return Changed;
1163
1164 Value *NumBytes =
1165 Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1166
1167 CallInst *NewCall;
1168 if (SplatValue) {
1169 NewCall = Builder.CreateMemSet(BasePtr, SplatValue, NumBytes,
1170 MaybeAlign(StoreAlignment));
1171 } else {
1172 // Everything is emitted in default address space
1173 Type *Int8PtrTy = DestInt8PtrTy;
1174
1175 Module *M = TheStore->getModule();
1176 StringRef FuncName = "memset_pattern16";
1177 FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(),
1178 Int8PtrTy, Int8PtrTy, IntIdxTy);
1179 inferLibFuncAttributes(M, FuncName, *TLI);
1180
1181 // Otherwise we should form a memset_pattern16. PatternValue is known to be
1182 // an constant array of 16-bytes. Plop the value into a mergable global.
1183 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1184 GlobalValue::PrivateLinkage,
1185 PatternValue, ".memset_pattern");
1186 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
1187 GV->setAlignment(Align(16));
1188 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
1189 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
1190 }
1191 NewCall->setDebugLoc(TheStore->getDebugLoc());
1192
1193 if (MSSAU) {
1194 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1195 NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1196 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1197 }
1198
1199 LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1200 << " from store to: " << *Ev << " at: " << *TheStore
1201 << "\n");
1202
1203 ORE.emit([&]() {
1204 OptimizationRemark R(DEBUG_TYPE, "ProcessLoopStridedStore",
1205 NewCall->getDebugLoc(), Preheader);
1206 R << "Transformed loop-strided store in "
1207 << ore::NV("Function", TheStore->getFunction())
1208 << " function into a call to "
1209 << ore::NV("NewFunction", NewCall->getCalledFunction())
1210 << "() intrinsic";
1211 if (!Stores.empty())
1212 R << ore::setExtraArgs();
1213 for (auto *I : Stores) {
1214 R << ore::NV("FromBlock", I->getParent()->getName())
1215 << ore::NV("ToBlock", Preheader->getName());
1216 }
1217 return R;
1218 });
1219
1220 // Okay, the memset has been formed. Zap the original store and anything that
1221 // feeds into it.
1222 for (auto *I : Stores) {
1223 if (MSSAU)
1224 MSSAU->removeMemoryAccess(I, true);
1225 deleteDeadInstruction(I);
1226 }
1227 if (MSSAU && VerifyMemorySSA)
1228 MSSAU->getMemorySSA()->verifyMemorySSA();
1229 ++NumMemSet;
1230 ExpCleaner.markResultUsed();
1231 return true;
1232 }
1233
1234 /// If the stored value is a strided load in the same loop with the same stride
1235 /// this may be transformable into a memcpy. This kicks in for stuff like
1236 /// for (i) A[i] = B[i];
processLoopStoreOfLoopLoad(StoreInst * SI,const SCEV * BECount)1237 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
1238 const SCEV *BECount) {
1239 assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.");
1240
1241 Value *StorePtr = SI->getPointerOperand();
1242 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
1243 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
1244
1245 // The store must be feeding a non-volatile load.
1246 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1247 assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.");
1248
1249 // See if the pointer expression is an AddRec like {base,+,1} on the current
1250 // loop, which indicates a strided load. If we have something else, it's a
1251 // random load we can't handle.
1252 Value *LoadPtr = LI->getPointerOperand();
1253 const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr));
1254
1255 const SCEV *StoreSizeSCEV = SE->getConstant(StorePtr->getType(), StoreSize);
1256 return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSizeSCEV,
1257 SI->getAlign(), LI->getAlign(), SI, LI,
1258 StoreEv, LoadEv, BECount);
1259 }
1260
1261 class MemmoveVerifier {
1262 public:
MemmoveVerifier(const Value & LoadBasePtr,const Value & StoreBasePtr,const DataLayout & DL)1263 explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr,
1264 const DataLayout &DL)
1265 : DL(DL), LoadOff(0), StoreOff(0),
1266 BP1(llvm::GetPointerBaseWithConstantOffset(
1267 LoadBasePtr.stripPointerCasts(), LoadOff, DL)),
1268 BP2(llvm::GetPointerBaseWithConstantOffset(
1269 StoreBasePtr.stripPointerCasts(), StoreOff, DL)),
1270 IsSameObject(BP1 == BP2) {}
1271
loadAndStoreMayFormMemmove(unsigned StoreSize,bool IsNegStride,const Instruction & TheLoad,bool IsMemCpy) const1272 bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride,
1273 const Instruction &TheLoad,
1274 bool IsMemCpy) const {
1275 if (IsMemCpy) {
1276 // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1277 // for negative stride.
1278 if ((!IsNegStride && LoadOff <= StoreOff) ||
1279 (IsNegStride && LoadOff >= StoreOff))
1280 return false;
1281 } else {
1282 // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1283 // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr.
1284 int64_t LoadSize =
1285 DL.getTypeSizeInBits(TheLoad.getType()).getFixedSize() / 8;
1286 if (BP1 != BP2 || LoadSize != int64_t(StoreSize))
1287 return false;
1288 if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) ||
1289 (IsNegStride && LoadOff + LoadSize > StoreOff))
1290 return false;
1291 }
1292 return true;
1293 }
1294
1295 private:
1296 const DataLayout &DL;
1297 int64_t LoadOff;
1298 int64_t StoreOff;
1299 const Value *BP1;
1300 const Value *BP2;
1301
1302 public:
1303 const bool IsSameObject;
1304 };
1305
processLoopStoreOfLoopLoad(Value * DestPtr,Value * SourcePtr,const SCEV * StoreSizeSCEV,MaybeAlign StoreAlign,MaybeAlign LoadAlign,Instruction * TheStore,Instruction * TheLoad,const SCEVAddRecExpr * StoreEv,const SCEVAddRecExpr * LoadEv,const SCEV * BECount)1306 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
1307 Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV,
1308 MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore,
1309 Instruction *TheLoad, const SCEVAddRecExpr *StoreEv,
1310 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
1311
1312 // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to
1313 // conservatively bail here, since otherwise we may have to transform
1314 // llvm.memcpy.inline into llvm.memcpy which is illegal.
1315 if (isa<MemCpyInlineInst>(TheStore))
1316 return false;
1317
1318 // The trip count of the loop and the base pointer of the addrec SCEV is
1319 // guaranteed to be loop invariant, which means that it should dominate the
1320 // header. This allows us to insert code for it in the preheader.
1321 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1322 IRBuilder<> Builder(Preheader->getTerminator());
1323 SCEVExpander Expander(*SE, *DL, "loop-idiom");
1324
1325 SCEVExpanderCleaner ExpCleaner(Expander, *DT);
1326
1327 bool Changed = false;
1328 const SCEV *StrStart = StoreEv->getStart();
1329 unsigned StrAS = DestPtr->getType()->getPointerAddressSpace();
1330 Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS));
1331
1332 APInt Stride = getStoreStride(StoreEv);
1333 const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(StoreSizeSCEV);
1334
1335 // TODO: Deal with non-constant size; Currently expect constant store size
1336 assert(ConstStoreSize && "store size is expected to be a constant");
1337
1338 int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue();
1339 bool IsNegStride = StoreSize == -Stride;
1340
1341 // Handle negative strided loops.
1342 if (IsNegStride)
1343 StrStart =
1344 getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSizeSCEV, SE);
1345
1346 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1347 // this into a memcpy in the loop preheader now if we want. However, this
1348 // would be unsafe to do if there is anything else in the loop that may read
1349 // or write the memory region we're storing to. This includes the load that
1350 // feeds the stores. Check for an alias by generating the base address and
1351 // checking everything.
1352 Value *StoreBasePtr = Expander.expandCodeFor(
1353 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
1354
1355 // From here on out, conservatively report to the pass manager that we've
1356 // changed the IR, even if we later clean up these added instructions. There
1357 // may be structural differences e.g. in the order of use lists not accounted
1358 // for in just a textual dump of the IR. This is written as a variable, even
1359 // though statically all the places this dominates could be replaced with
1360 // 'true', with the hope that anyone trying to be clever / "more precise" with
1361 // the return value will read this comment, and leave them alone.
1362 Changed = true;
1363
1364 SmallPtrSet<Instruction *, 2> IgnoredInsts;
1365 IgnoredInsts.insert(TheStore);
1366
1367 bool IsMemCpy = isa<MemCpyInst>(TheStore);
1368 const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store";
1369
1370 bool LoopAccessStore =
1371 mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1372 StoreSizeSCEV, *AA, IgnoredInsts);
1373 if (LoopAccessStore) {
1374 // For memmove case it's not enough to guarantee that loop doesn't access
1375 // TheStore and TheLoad. Additionally we need to make sure that TheStore is
1376 // the only user of TheLoad.
1377 if (!TheLoad->hasOneUse())
1378 return Changed;
1379 IgnoredInsts.insert(TheLoad);
1380 if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
1381 BECount, StoreSizeSCEV, *AA, IgnoredInsts)) {
1382 ORE.emit([&]() {
1383 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessStore",
1384 TheStore)
1385 << ore::NV("Inst", InstRemark) << " in "
1386 << ore::NV("Function", TheStore->getFunction())
1387 << " function will not be hoisted: "
1388 << ore::NV("Reason", "The loop may access store location");
1389 });
1390 return Changed;
1391 }
1392 IgnoredInsts.erase(TheLoad);
1393 }
1394
1395 const SCEV *LdStart = LoadEv->getStart();
1396 unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace();
1397
1398 // Handle negative strided loops.
1399 if (IsNegStride)
1400 LdStart =
1401 getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSizeSCEV, SE);
1402
1403 // For a memcpy, we have to make sure that the input array is not being
1404 // mutated by the loop.
1405 Value *LoadBasePtr = Expander.expandCodeFor(
1406 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
1407
1408 // If the store is a memcpy instruction, we must check if it will write to
1409 // the load memory locations. So remove it from the ignored stores.
1410 if (IsMemCpy)
1411 IgnoredInsts.erase(TheStore);
1412 MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL);
1413 if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
1414 StoreSizeSCEV, *AA, IgnoredInsts)) {
1415 if (!IsMemCpy) {
1416 ORE.emit([&]() {
1417 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessLoad",
1418 TheLoad)
1419 << ore::NV("Inst", InstRemark) << " in "
1420 << ore::NV("Function", TheStore->getFunction())
1421 << " function will not be hoisted: "
1422 << ore::NV("Reason", "The loop may access load location");
1423 });
1424 return Changed;
1425 }
1426 // At this point loop may access load only for memcpy in same underlying
1427 // object. If that's not the case bail out.
1428 if (!Verifier.IsSameObject)
1429 return Changed;
1430 }
1431
1432 bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore;
1433 if (UseMemMove)
1434 if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, *TheLoad,
1435 IsMemCpy))
1436 return Changed;
1437
1438 if (avoidLIRForMultiBlockLoop())
1439 return Changed;
1440
1441 // Okay, everything is safe, we can transform this!
1442
1443 const SCEV *NumBytesS =
1444 getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1445
1446 Value *NumBytes =
1447 Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1448
1449 CallInst *NewCall = nullptr;
1450 // Check whether to generate an unordered atomic memcpy:
1451 // If the load or store are atomic, then they must necessarily be unordered
1452 // by previous checks.
1453 if (!TheStore->isAtomic() && !TheLoad->isAtomic()) {
1454 if (UseMemMove)
1455 NewCall = Builder.CreateMemMove(StoreBasePtr, StoreAlign, LoadBasePtr,
1456 LoadAlign, NumBytes);
1457 else
1458 NewCall = Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr,
1459 LoadAlign, NumBytes);
1460 } else {
1461 // For now don't support unordered atomic memmove.
1462 if (UseMemMove)
1463 return Changed;
1464 // We cannot allow unaligned ops for unordered load/store, so reject
1465 // anything where the alignment isn't at least the element size.
1466 assert((StoreAlign.hasValue() && LoadAlign.hasValue()) &&
1467 "Expect unordered load/store to have align.");
1468 if (StoreAlign.getValue() < StoreSize || LoadAlign.getValue() < StoreSize)
1469 return Changed;
1470
1471 // If the element.atomic memcpy is not lowered into explicit
1472 // loads/stores later, then it will be lowered into an element-size
1473 // specific lib call. If the lib call doesn't exist for our store size, then
1474 // we shouldn't generate the memcpy.
1475 if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize())
1476 return Changed;
1477
1478 // Create the call.
1479 // Note that unordered atomic loads/stores are *required* by the spec to
1480 // have an alignment but non-atomic loads/stores may not.
1481 NewCall = Builder.CreateElementUnorderedAtomicMemCpy(
1482 StoreBasePtr, StoreAlign.getValue(), LoadBasePtr, LoadAlign.getValue(),
1483 NumBytes, StoreSize);
1484 }
1485 NewCall->setDebugLoc(TheStore->getDebugLoc());
1486
1487 if (MSSAU) {
1488 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1489 NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1490 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1491 }
1492
1493 LLVM_DEBUG(dbgs() << " Formed new call: " << *NewCall << "\n"
1494 << " from load ptr=" << *LoadEv << " at: " << *TheLoad
1495 << "\n"
1496 << " from store ptr=" << *StoreEv << " at: " << *TheStore
1497 << "\n");
1498
1499 ORE.emit([&]() {
1500 return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStoreOfLoopLoad",
1501 NewCall->getDebugLoc(), Preheader)
1502 << "Formed a call to "
1503 << ore::NV("NewFunction", NewCall->getCalledFunction())
1504 << "() intrinsic from " << ore::NV("Inst", InstRemark)
1505 << " instruction in " << ore::NV("Function", TheStore->getFunction())
1506 << " function"
1507 << ore::setExtraArgs()
1508 << ore::NV("FromBlock", TheStore->getParent()->getName())
1509 << ore::NV("ToBlock", Preheader->getName());
1510 });
1511
1512 // Okay, a new call to memcpy/memmove has been formed. Zap the original store
1513 // and anything that feeds into it.
1514 if (MSSAU)
1515 MSSAU->removeMemoryAccess(TheStore, true);
1516 deleteDeadInstruction(TheStore);
1517 if (MSSAU && VerifyMemorySSA)
1518 MSSAU->getMemorySSA()->verifyMemorySSA();
1519 if (UseMemMove)
1520 ++NumMemMove;
1521 else
1522 ++NumMemCpy;
1523 ExpCleaner.markResultUsed();
1524 return true;
1525 }
1526
1527 // When compiling for codesize we avoid idiom recognition for a multi-block loop
1528 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
1529 //
avoidLIRForMultiBlockLoop(bool IsMemset,bool IsLoopMemset)1530 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
1531 bool IsLoopMemset) {
1532 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
1533 if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) {
1534 LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
1535 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1536 << " avoided: multi-block top-level loop\n");
1537 return true;
1538 }
1539 }
1540
1541 return false;
1542 }
1543
runOnNoncountableLoop()1544 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1545 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
1546 << CurLoop->getHeader()->getParent()->getName()
1547 << "] Noncountable Loop %"
1548 << CurLoop->getHeader()->getName() << "\n");
1549
1550 return recognizePopcount() || recognizeAndInsertFFS() ||
1551 recognizeShiftUntilBitTest() || recognizeShiftUntilZero();
1552 }
1553
1554 /// Check if the given conditional branch is based on the comparison between
1555 /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is
1556 /// true), the control yields to the loop entry. If the branch matches the
1557 /// behavior, the variable involved in the comparison is returned. This function
1558 /// will be called to see if the precondition and postcondition of the loop are
1559 /// in desirable form.
matchCondition(BranchInst * BI,BasicBlock * LoopEntry,bool JmpOnZero=false)1560 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
1561 bool JmpOnZero = false) {
1562 if (!BI || !BI->isConditional())
1563 return nullptr;
1564
1565 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1566 if (!Cond)
1567 return nullptr;
1568
1569 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1570 if (!CmpZero || !CmpZero->isZero())
1571 return nullptr;
1572
1573 BasicBlock *TrueSucc = BI->getSuccessor(0);
1574 BasicBlock *FalseSucc = BI->getSuccessor(1);
1575 if (JmpOnZero)
1576 std::swap(TrueSucc, FalseSucc);
1577
1578 ICmpInst::Predicate Pred = Cond->getPredicate();
1579 if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) ||
1580 (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry))
1581 return Cond->getOperand(0);
1582
1583 return nullptr;
1584 }
1585
1586 // Check if the recurrence variable `VarX` is in the right form to create
1587 // the idiom. Returns the value coerced to a PHINode if so.
getRecurrenceVar(Value * VarX,Instruction * DefX,BasicBlock * LoopEntry)1588 static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
1589 BasicBlock *LoopEntry) {
1590 auto *PhiX = dyn_cast<PHINode>(VarX);
1591 if (PhiX && PhiX->getParent() == LoopEntry &&
1592 (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX))
1593 return PhiX;
1594 return nullptr;
1595 }
1596
1597 /// Return true iff the idiom is detected in the loop.
1598 ///
1599 /// Additionally:
1600 /// 1) \p CntInst is set to the instruction counting the population bit.
1601 /// 2) \p CntPhi is set to the corresponding phi node.
1602 /// 3) \p Var is set to the value whose population bits are being counted.
1603 ///
1604 /// The core idiom we are trying to detect is:
1605 /// \code
1606 /// if (x0 != 0)
1607 /// goto loop-exit // the precondition of the loop
1608 /// cnt0 = init-val;
1609 /// do {
1610 /// x1 = phi (x0, x2);
1611 /// cnt1 = phi(cnt0, cnt2);
1612 ///
1613 /// cnt2 = cnt1 + 1;
1614 /// ...
1615 /// x2 = x1 & (x1 - 1);
1616 /// ...
1617 /// } while(x != 0);
1618 ///
1619 /// loop-exit:
1620 /// \endcode
detectPopcountIdiom(Loop * CurLoop,BasicBlock * PreCondBB,Instruction * & CntInst,PHINode * & CntPhi,Value * & Var)1621 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1622 Instruction *&CntInst, PHINode *&CntPhi,
1623 Value *&Var) {
1624 // step 1: Check to see if the look-back branch match this pattern:
1625 // "if (a!=0) goto loop-entry".
1626 BasicBlock *LoopEntry;
1627 Instruction *DefX2, *CountInst;
1628 Value *VarX1, *VarX0;
1629 PHINode *PhiX, *CountPhi;
1630
1631 DefX2 = CountInst = nullptr;
1632 VarX1 = VarX0 = nullptr;
1633 PhiX = CountPhi = nullptr;
1634 LoopEntry = *(CurLoop->block_begin());
1635
1636 // step 1: Check if the loop-back branch is in desirable form.
1637 {
1638 if (Value *T = matchCondition(
1639 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1640 DefX2 = dyn_cast<Instruction>(T);
1641 else
1642 return false;
1643 }
1644
1645 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1646 {
1647 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1648 return false;
1649
1650 BinaryOperator *SubOneOp;
1651
1652 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1653 VarX1 = DefX2->getOperand(1);
1654 else {
1655 VarX1 = DefX2->getOperand(0);
1656 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1657 }
1658 if (!SubOneOp || SubOneOp->getOperand(0) != VarX1)
1659 return false;
1660
1661 ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1));
1662 if (!Dec ||
1663 !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1664 (SubOneOp->getOpcode() == Instruction::Add &&
1665 Dec->isMinusOne()))) {
1666 return false;
1667 }
1668 }
1669
1670 // step 3: Check the recurrence of variable X
1671 PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry);
1672 if (!PhiX)
1673 return false;
1674
1675 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1676 {
1677 CountInst = nullptr;
1678 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1679 IterE = LoopEntry->end();
1680 Iter != IterE; Iter++) {
1681 Instruction *Inst = &*Iter;
1682 if (Inst->getOpcode() != Instruction::Add)
1683 continue;
1684
1685 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1686 if (!Inc || !Inc->isOne())
1687 continue;
1688
1689 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1690 if (!Phi)
1691 continue;
1692
1693 // Check if the result of the instruction is live of the loop.
1694 bool LiveOutLoop = false;
1695 for (User *U : Inst->users()) {
1696 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1697 LiveOutLoop = true;
1698 break;
1699 }
1700 }
1701
1702 if (LiveOutLoop) {
1703 CountInst = Inst;
1704 CountPhi = Phi;
1705 break;
1706 }
1707 }
1708
1709 if (!CountInst)
1710 return false;
1711 }
1712
1713 // step 5: check if the precondition is in this form:
1714 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1715 {
1716 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1717 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1718 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1719 return false;
1720
1721 CntInst = CountInst;
1722 CntPhi = CountPhi;
1723 Var = T;
1724 }
1725
1726 return true;
1727 }
1728
1729 /// Return true if the idiom is detected in the loop.
1730 ///
1731 /// Additionally:
1732 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1733 /// or nullptr if there is no such.
1734 /// 2) \p CntPhi is set to the corresponding phi node
1735 /// or nullptr if there is no such.
1736 /// 3) \p Var is set to the value whose CTLZ could be used.
1737 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1738 ///
1739 /// The core idiom we are trying to detect is:
1740 /// \code
1741 /// if (x0 == 0)
1742 /// goto loop-exit // the precondition of the loop
1743 /// cnt0 = init-val;
1744 /// do {
1745 /// x = phi (x0, x.next); //PhiX
1746 /// cnt = phi(cnt0, cnt.next);
1747 ///
1748 /// cnt.next = cnt + 1;
1749 /// ...
1750 /// x.next = x >> 1; // DefX
1751 /// ...
1752 /// } while(x.next != 0);
1753 ///
1754 /// loop-exit:
1755 /// \endcode
detectShiftUntilZeroIdiom(Loop * CurLoop,const DataLayout & DL,Intrinsic::ID & IntrinID,Value * & InitX,Instruction * & CntInst,PHINode * & CntPhi,Instruction * & DefX)1756 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL,
1757 Intrinsic::ID &IntrinID, Value *&InitX,
1758 Instruction *&CntInst, PHINode *&CntPhi,
1759 Instruction *&DefX) {
1760 BasicBlock *LoopEntry;
1761 Value *VarX = nullptr;
1762
1763 DefX = nullptr;
1764 CntInst = nullptr;
1765 CntPhi = nullptr;
1766 LoopEntry = *(CurLoop->block_begin());
1767
1768 // step 1: Check if the loop-back branch is in desirable form.
1769 if (Value *T = matchCondition(
1770 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1771 DefX = dyn_cast<Instruction>(T);
1772 else
1773 return false;
1774
1775 // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1"
1776 if (!DefX || !DefX->isShift())
1777 return false;
1778 IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz :
1779 Intrinsic::ctlz;
1780 ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1));
1781 if (!Shft || !Shft->isOne())
1782 return false;
1783 VarX = DefX->getOperand(0);
1784
1785 // step 3: Check the recurrence of variable X
1786 PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
1787 if (!PhiX)
1788 return false;
1789
1790 InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader());
1791
1792 // Make sure the initial value can't be negative otherwise the ashr in the
1793 // loop might never reach zero which would make the loop infinite.
1794 if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL))
1795 return false;
1796
1797 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1798 // or cnt.next = cnt + -1.
1799 // TODO: We can skip the step. If loop trip count is known (CTLZ),
1800 // then all uses of "cnt.next" could be optimized to the trip count
1801 // plus "cnt0". Currently it is not optimized.
1802 // This step could be used to detect POPCNT instruction:
1803 // cnt.next = cnt + (x.next & 1)
1804 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1805 IterE = LoopEntry->end();
1806 Iter != IterE; Iter++) {
1807 Instruction *Inst = &*Iter;
1808 if (Inst->getOpcode() != Instruction::Add)
1809 continue;
1810
1811 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1812 if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
1813 continue;
1814
1815 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1816 if (!Phi)
1817 continue;
1818
1819 CntInst = Inst;
1820 CntPhi = Phi;
1821 break;
1822 }
1823 if (!CntInst)
1824 return false;
1825
1826 return true;
1827 }
1828
1829 /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop
1830 /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new
1831 /// trip count returns true; otherwise, returns false.
recognizeAndInsertFFS()1832 bool LoopIdiomRecognize::recognizeAndInsertFFS() {
1833 // Give up if the loop has multiple blocks or multiple backedges.
1834 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1835 return false;
1836
1837 Intrinsic::ID IntrinID;
1838 Value *InitX;
1839 Instruction *DefX = nullptr;
1840 PHINode *CntPhi = nullptr;
1841 Instruction *CntInst = nullptr;
1842 // Help decide if transformation is profitable. For ShiftUntilZero idiom,
1843 // this is always 6.
1844 size_t IdiomCanonicalSize = 6;
1845
1846 if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX,
1847 CntInst, CntPhi, DefX))
1848 return false;
1849
1850 bool IsCntPhiUsedOutsideLoop = false;
1851 for (User *U : CntPhi->users())
1852 if (!CurLoop->contains(cast<Instruction>(U))) {
1853 IsCntPhiUsedOutsideLoop = true;
1854 break;
1855 }
1856 bool IsCntInstUsedOutsideLoop = false;
1857 for (User *U : CntInst->users())
1858 if (!CurLoop->contains(cast<Instruction>(U))) {
1859 IsCntInstUsedOutsideLoop = true;
1860 break;
1861 }
1862 // If both CntInst and CntPhi are used outside the loop the profitability
1863 // is questionable.
1864 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1865 return false;
1866
1867 // For some CPUs result of CTLZ(X) intrinsic is undefined
1868 // when X is 0. If we can not guarantee X != 0, we need to check this
1869 // when expand.
1870 bool ZeroCheck = false;
1871 // It is safe to assume Preheader exist as it was checked in
1872 // parent function RunOnLoop.
1873 BasicBlock *PH = CurLoop->getLoopPreheader();
1874
1875 // If we are using the count instruction outside the loop, make sure we
1876 // have a zero check as a precondition. Without the check the loop would run
1877 // one iteration for before any check of the input value. This means 0 and 1
1878 // would have identical behavior in the original loop and thus
1879 if (!IsCntPhiUsedOutsideLoop) {
1880 auto *PreCondBB = PH->getSinglePredecessor();
1881 if (!PreCondBB)
1882 return false;
1883 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1884 if (!PreCondBI)
1885 return false;
1886 if (matchCondition(PreCondBI, PH) != InitX)
1887 return false;
1888 ZeroCheck = true;
1889 }
1890
1891 // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always
1892 // profitable if we delete the loop.
1893
1894 // the loop has only 6 instructions:
1895 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1896 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1897 // %shr = ashr %n.addr.0, 1
1898 // %tobool = icmp eq %shr, 0
1899 // %inc = add nsw %i.0, 1
1900 // br i1 %tobool
1901
1902 const Value *Args[] = {InitX,
1903 ConstantInt::getBool(InitX->getContext(), ZeroCheck)};
1904
1905 // @llvm.dbg doesn't count as they have no semantic effect.
1906 auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug();
1907 uint32_t HeaderSize =
1908 std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end());
1909
1910 IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args);
1911 InstructionCost Cost =
1912 TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency);
1913 if (HeaderSize != IdiomCanonicalSize &&
1914 Cost > TargetTransformInfo::TCC_Basic)
1915 return false;
1916
1917 transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX,
1918 DefX->getDebugLoc(), ZeroCheck,
1919 IsCntPhiUsedOutsideLoop);
1920 return true;
1921 }
1922
1923 /// Recognizes a population count idiom in a non-countable loop.
1924 ///
1925 /// If detected, transforms the relevant code to issue the popcount intrinsic
1926 /// function call, and returns true; otherwise, returns false.
recognizePopcount()1927 bool LoopIdiomRecognize::recognizePopcount() {
1928 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1929 return false;
1930
1931 // Counting population are usually conducted by few arithmetic instructions.
1932 // Such instructions can be easily "absorbed" by vacant slots in a
1933 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1934 // in a compact loop.
1935
1936 // Give up if the loop has multiple blocks or multiple backedges.
1937 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1938 return false;
1939
1940 BasicBlock *LoopBody = *(CurLoop->block_begin());
1941 if (LoopBody->size() >= 20) {
1942 // The loop is too big, bail out.
1943 return false;
1944 }
1945
1946 // It should have a preheader containing nothing but an unconditional branch.
1947 BasicBlock *PH = CurLoop->getLoopPreheader();
1948 if (!PH || &PH->front() != PH->getTerminator())
1949 return false;
1950 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1951 if (!EntryBI || EntryBI->isConditional())
1952 return false;
1953
1954 // It should have a precondition block where the generated popcount intrinsic
1955 // function can be inserted.
1956 auto *PreCondBB = PH->getSinglePredecessor();
1957 if (!PreCondBB)
1958 return false;
1959 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1960 if (!PreCondBI || PreCondBI->isUnconditional())
1961 return false;
1962
1963 Instruction *CntInst;
1964 PHINode *CntPhi;
1965 Value *Val;
1966 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1967 return false;
1968
1969 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1970 return true;
1971 }
1972
createPopcntIntrinsic(IRBuilder<> & IRBuilder,Value * Val,const DebugLoc & DL)1973 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1974 const DebugLoc &DL) {
1975 Value *Ops[] = {Val};
1976 Type *Tys[] = {Val->getType()};
1977
1978 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1979 Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1980 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1981 CI->setDebugLoc(DL);
1982
1983 return CI;
1984 }
1985
createFFSIntrinsic(IRBuilder<> & IRBuilder,Value * Val,const DebugLoc & DL,bool ZeroCheck,Intrinsic::ID IID)1986 static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1987 const DebugLoc &DL, bool ZeroCheck,
1988 Intrinsic::ID IID) {
1989 Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)};
1990 Type *Tys[] = {Val->getType()};
1991
1992 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1993 Function *Func = Intrinsic::getDeclaration(M, IID, Tys);
1994 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1995 CI->setDebugLoc(DL);
1996
1997 return CI;
1998 }
1999
2000 /// Transform the following loop (Using CTLZ, CTTZ is similar):
2001 /// loop:
2002 /// CntPhi = PHI [Cnt0, CntInst]
2003 /// PhiX = PHI [InitX, DefX]
2004 /// CntInst = CntPhi + 1
2005 /// DefX = PhiX >> 1
2006 /// LOOP_BODY
2007 /// Br: loop if (DefX != 0)
2008 /// Use(CntPhi) or Use(CntInst)
2009 ///
2010 /// Into:
2011 /// If CntPhi used outside the loop:
2012 /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
2013 /// Count = CountPrev + 1
2014 /// else
2015 /// Count = BitWidth(InitX) - CTLZ(InitX)
2016 /// loop:
2017 /// CntPhi = PHI [Cnt0, CntInst]
2018 /// PhiX = PHI [InitX, DefX]
2019 /// PhiCount = PHI [Count, Dec]
2020 /// CntInst = CntPhi + 1
2021 /// DefX = PhiX >> 1
2022 /// Dec = PhiCount - 1
2023 /// LOOP_BODY
2024 /// Br: loop if (Dec != 0)
2025 /// Use(CountPrev + Cnt0) // Use(CntPhi)
2026 /// or
2027 /// Use(Count + Cnt0) // Use(CntInst)
2028 ///
2029 /// If LOOP_BODY is empty the loop will be deleted.
2030 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
transformLoopToCountable(Intrinsic::ID IntrinID,BasicBlock * Preheader,Instruction * CntInst,PHINode * CntPhi,Value * InitX,Instruction * DefX,const DebugLoc & DL,bool ZeroCheck,bool IsCntPhiUsedOutsideLoop)2031 void LoopIdiomRecognize::transformLoopToCountable(
2032 Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst,
2033 PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL,
2034 bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) {
2035 BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator());
2036
2037 // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block
2038 IRBuilder<> Builder(PreheaderBr);
2039 Builder.SetCurrentDebugLocation(DL);
2040
2041 // If there are no uses of CntPhi crate:
2042 // Count = BitWidth - CTLZ(InitX);
2043 // NewCount = Count;
2044 // If there are uses of CntPhi create:
2045 // NewCount = BitWidth - CTLZ(InitX >> 1);
2046 // Count = NewCount + 1;
2047 Value *InitXNext;
2048 if (IsCntPhiUsedOutsideLoop) {
2049 if (DefX->getOpcode() == Instruction::AShr)
2050 InitXNext = Builder.CreateAShr(InitX, 1);
2051 else if (DefX->getOpcode() == Instruction::LShr)
2052 InitXNext = Builder.CreateLShr(InitX, 1);
2053 else if (DefX->getOpcode() == Instruction::Shl) // cttz
2054 InitXNext = Builder.CreateShl(InitX, 1);
2055 else
2056 llvm_unreachable("Unexpected opcode!");
2057 } else
2058 InitXNext = InitX;
2059 Value *Count =
2060 createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID);
2061 Type *CountTy = Count->getType();
2062 Count = Builder.CreateSub(
2063 ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count);
2064 Value *NewCount = Count;
2065 if (IsCntPhiUsedOutsideLoop)
2066 Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1));
2067
2068 NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType());
2069
2070 Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
2071 if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) {
2072 // If the counter was being incremented in the loop, add NewCount to the
2073 // counter's initial value, but only if the initial value is not zero.
2074 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
2075 if (!InitConst || !InitConst->isZero())
2076 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
2077 } else {
2078 // If the count was being decremented in the loop, subtract NewCount from
2079 // the counter's initial value.
2080 NewCount = Builder.CreateSub(CntInitVal, NewCount);
2081 }
2082
2083 // Step 2: Insert new IV and loop condition:
2084 // loop:
2085 // ...
2086 // PhiCount = PHI [Count, Dec]
2087 // ...
2088 // Dec = PhiCount - 1
2089 // ...
2090 // Br: loop if (Dec != 0)
2091 BasicBlock *Body = *(CurLoop->block_begin());
2092 auto *LbBr = cast<BranchInst>(Body->getTerminator());
2093 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
2094
2095 PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front());
2096
2097 Builder.SetInsertPoint(LbCond);
2098 Instruction *TcDec = cast<Instruction>(Builder.CreateSub(
2099 TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true));
2100
2101 TcPhi->addIncoming(Count, Preheader);
2102 TcPhi->addIncoming(TcDec, Body);
2103
2104 CmpInst::Predicate Pred =
2105 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
2106 LbCond->setPredicate(Pred);
2107 LbCond->setOperand(0, TcDec);
2108 LbCond->setOperand(1, ConstantInt::get(CountTy, 0));
2109
2110 // Step 3: All the references to the original counter outside
2111 // the loop are replaced with the NewCount
2112 if (IsCntPhiUsedOutsideLoop)
2113 CntPhi->replaceUsesOutsideBlock(NewCount, Body);
2114 else
2115 CntInst->replaceUsesOutsideBlock(NewCount, Body);
2116
2117 // step 4: Forget the "non-computable" trip-count SCEV associated with the
2118 // loop. The loop would otherwise not be deleted even if it becomes empty.
2119 SE->forgetLoop(CurLoop);
2120 }
2121
transformLoopToPopcount(BasicBlock * PreCondBB,Instruction * CntInst,PHINode * CntPhi,Value * Var)2122 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
2123 Instruction *CntInst,
2124 PHINode *CntPhi, Value *Var) {
2125 BasicBlock *PreHead = CurLoop->getLoopPreheader();
2126 auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator());
2127 const DebugLoc &DL = CntInst->getDebugLoc();
2128
2129 // Assuming before transformation, the loop is following:
2130 // if (x) // the precondition
2131 // do { cnt++; x &= x - 1; } while(x);
2132
2133 // Step 1: Insert the ctpop instruction at the end of the precondition block
2134 IRBuilder<> Builder(PreCondBr);
2135 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
2136 {
2137 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
2138 NewCount = PopCntZext =
2139 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
2140
2141 if (NewCount != PopCnt)
2142 (cast<Instruction>(NewCount))->setDebugLoc(DL);
2143
2144 // TripCnt is exactly the number of iterations the loop has
2145 TripCnt = NewCount;
2146
2147 // If the population counter's initial value is not zero, insert Add Inst.
2148 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
2149 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
2150 if (!InitConst || !InitConst->isZero()) {
2151 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
2152 (cast<Instruction>(NewCount))->setDebugLoc(DL);
2153 }
2154 }
2155
2156 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
2157 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
2158 // function would be partial dead code, and downstream passes will drag
2159 // it back from the precondition block to the preheader.
2160 {
2161 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
2162
2163 Value *Opnd0 = PopCntZext;
2164 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
2165 if (PreCond->getOperand(0) != Var)
2166 std::swap(Opnd0, Opnd1);
2167
2168 ICmpInst *NewPreCond = cast<ICmpInst>(
2169 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
2170 PreCondBr->setCondition(NewPreCond);
2171
2172 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
2173 }
2174
2175 // Step 3: Note that the population count is exactly the trip count of the
2176 // loop in question, which enable us to convert the loop from noncountable
2177 // loop into a countable one. The benefit is twofold:
2178 //
2179 // - If the loop only counts population, the entire loop becomes dead after
2180 // the transformation. It is a lot easier to prove a countable loop dead
2181 // than to prove a noncountable one. (In some C dialects, an infinite loop
2182 // isn't dead even if it computes nothing useful. In general, DCE needs
2183 // to prove a noncountable loop finite before safely delete it.)
2184 //
2185 // - If the loop also performs something else, it remains alive.
2186 // Since it is transformed to countable form, it can be aggressively
2187 // optimized by some optimizations which are in general not applicable
2188 // to a noncountable loop.
2189 //
2190 // After this step, this loop (conceptually) would look like following:
2191 // newcnt = __builtin_ctpop(x);
2192 // t = newcnt;
2193 // if (x)
2194 // do { cnt++; x &= x-1; t--) } while (t > 0);
2195 BasicBlock *Body = *(CurLoop->block_begin());
2196 {
2197 auto *LbBr = cast<BranchInst>(Body->getTerminator());
2198 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
2199 Type *Ty = TripCnt->getType();
2200
2201 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
2202
2203 Builder.SetInsertPoint(LbCond);
2204 Instruction *TcDec = cast<Instruction>(
2205 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
2206 "tcdec", false, true));
2207
2208 TcPhi->addIncoming(TripCnt, PreHead);
2209 TcPhi->addIncoming(TcDec, Body);
2210
2211 CmpInst::Predicate Pred =
2212 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
2213 LbCond->setPredicate(Pred);
2214 LbCond->setOperand(0, TcDec);
2215 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
2216 }
2217
2218 // Step 4: All the references to the original population counter outside
2219 // the loop are replaced with the NewCount -- the value returned from
2220 // __builtin_ctpop().
2221 CntInst->replaceUsesOutsideBlock(NewCount, Body);
2222
2223 // step 5: Forget the "non-computable" trip-count SCEV associated with the
2224 // loop. The loop would otherwise not be deleted even if it becomes empty.
2225 SE->forgetLoop(CurLoop);
2226 }
2227
2228 /// Match loop-invariant value.
2229 template <typename SubPattern_t> struct match_LoopInvariant {
2230 SubPattern_t SubPattern;
2231 const Loop *L;
2232
match_LoopInvariantmatch_LoopInvariant2233 match_LoopInvariant(const SubPattern_t &SP, const Loop *L)
2234 : SubPattern(SP), L(L) {}
2235
matchmatch_LoopInvariant2236 template <typename ITy> bool match(ITy *V) {
2237 return L->isLoopInvariant(V) && SubPattern.match(V);
2238 }
2239 };
2240
2241 /// Matches if the value is loop-invariant.
2242 template <typename Ty>
m_LoopInvariant(const Ty & M,const Loop * L)2243 inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) {
2244 return match_LoopInvariant<Ty>(M, L);
2245 }
2246
2247 /// Return true if the idiom is detected in the loop.
2248 ///
2249 /// The core idiom we are trying to detect is:
2250 /// \code
2251 /// entry:
2252 /// <...>
2253 /// %bitmask = shl i32 1, %bitpos
2254 /// br label %loop
2255 ///
2256 /// loop:
2257 /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2258 /// %x.curr.bitmasked = and i32 %x.curr, %bitmask
2259 /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2260 /// %x.next = shl i32 %x.curr, 1
2261 /// <...>
2262 /// br i1 %x.curr.isbitunset, label %loop, label %end
2263 ///
2264 /// end:
2265 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2266 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2267 /// <...>
2268 /// \endcode
detectShiftUntilBitTestIdiom(Loop * CurLoop,Value * & BaseX,Value * & BitMask,Value * & BitPos,Value * & CurrX,Instruction * & NextX)2269 static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX,
2270 Value *&BitMask, Value *&BitPos,
2271 Value *&CurrX, Instruction *&NextX) {
2272 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2273 " Performing shift-until-bittest idiom detection.\n");
2274
2275 // Give up if the loop has multiple blocks or multiple backedges.
2276 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2277 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2278 return false;
2279 }
2280
2281 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2282 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2283 assert(LoopPreheaderBB && "There is always a loop preheader.");
2284
2285 using namespace PatternMatch;
2286
2287 // Step 1: Check if the loop backedge is in desirable form.
2288
2289 ICmpInst::Predicate Pred;
2290 Value *CmpLHS, *CmpRHS;
2291 BasicBlock *TrueBB, *FalseBB;
2292 if (!match(LoopHeaderBB->getTerminator(),
2293 m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)),
2294 m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) {
2295 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2296 return false;
2297 }
2298
2299 // Step 2: Check if the backedge's condition is in desirable form.
2300
2301 auto MatchVariableBitMask = [&]() {
2302 return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2303 match(CmpLHS,
2304 m_c_And(m_Value(CurrX),
2305 m_CombineAnd(
2306 m_Value(BitMask),
2307 m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)),
2308 CurLoop))));
2309 };
2310 auto MatchConstantBitMask = [&]() {
2311 return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2312 match(CmpLHS, m_And(m_Value(CurrX),
2313 m_CombineAnd(m_Value(BitMask), m_Power2()))) &&
2314 (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask)));
2315 };
2316 auto MatchDecomposableConstantBitMask = [&]() {
2317 APInt Mask;
2318 return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) &&
2319 ICmpInst::isEquality(Pred) && Mask.isPowerOf2() &&
2320 (BitMask = ConstantInt::get(CurrX->getType(), Mask)) &&
2321 (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2()));
2322 };
2323
2324 if (!MatchVariableBitMask() && !MatchConstantBitMask() &&
2325 !MatchDecomposableConstantBitMask()) {
2326 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n");
2327 return false;
2328 }
2329
2330 // Step 3: Check if the recurrence is in desirable form.
2331 auto *CurrXPN = dyn_cast<PHINode>(CurrX);
2332 if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) {
2333 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2334 return false;
2335 }
2336
2337 BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB);
2338 NextX =
2339 dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB));
2340
2341 assert(CurLoop->isLoopInvariant(BaseX) &&
2342 "Expected BaseX to be avaliable in the preheader!");
2343
2344 if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) {
2345 // FIXME: support right-shift?
2346 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2347 return false;
2348 }
2349
2350 // Step 4: Check if the backedge's destinations are in desirable form.
2351
2352 assert(ICmpInst::isEquality(Pred) &&
2353 "Should only get equality predicates here.");
2354
2355 // cmp-br is commutative, so canonicalize to a single variant.
2356 if (Pred != ICmpInst::Predicate::ICMP_EQ) {
2357 Pred = ICmpInst::getInversePredicate(Pred);
2358 std::swap(TrueBB, FalseBB);
2359 }
2360
2361 // We expect to exit loop when comparison yields false,
2362 // so when it yields true we should branch back to loop header.
2363 if (TrueBB != LoopHeaderBB) {
2364 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2365 return false;
2366 }
2367
2368 // Okay, idiom checks out.
2369 return true;
2370 }
2371
2372 /// Look for the following loop:
2373 /// \code
2374 /// entry:
2375 /// <...>
2376 /// %bitmask = shl i32 1, %bitpos
2377 /// br label %loop
2378 ///
2379 /// loop:
2380 /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2381 /// %x.curr.bitmasked = and i32 %x.curr, %bitmask
2382 /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2383 /// %x.next = shl i32 %x.curr, 1
2384 /// <...>
2385 /// br i1 %x.curr.isbitunset, label %loop, label %end
2386 ///
2387 /// end:
2388 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2389 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2390 /// <...>
2391 /// \endcode
2392 ///
2393 /// And transform it into:
2394 /// \code
2395 /// entry:
2396 /// %bitmask = shl i32 1, %bitpos
2397 /// %lowbitmask = add i32 %bitmask, -1
2398 /// %mask = or i32 %lowbitmask, %bitmask
2399 /// %x.masked = and i32 %x, %mask
2400 /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked,
2401 /// i1 true)
2402 /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros
2403 /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1
2404 /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos
2405 /// %tripcount = add i32 %backedgetakencount, 1
2406 /// %x.curr = shl i32 %x, %backedgetakencount
2407 /// %x.next = shl i32 %x, %tripcount
2408 /// br label %loop
2409 ///
2410 /// loop:
2411 /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ]
2412 /// %loop.iv.next = add nuw i32 %loop.iv, 1
2413 /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount
2414 /// <...>
2415 /// br i1 %loop.ivcheck, label %end, label %loop
2416 ///
2417 /// end:
2418 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2419 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2420 /// <...>
2421 /// \endcode
recognizeShiftUntilBitTest()2422 bool LoopIdiomRecognize::recognizeShiftUntilBitTest() {
2423 bool MadeChange = false;
2424
2425 Value *X, *BitMask, *BitPos, *XCurr;
2426 Instruction *XNext;
2427 if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr,
2428 XNext)) {
2429 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2430 " shift-until-bittest idiom detection failed.\n");
2431 return MadeChange;
2432 }
2433 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n");
2434
2435 // Ok, it is the idiom we were looking for, we *could* transform this loop,
2436 // but is it profitable to transform?
2437
2438 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2439 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2440 assert(LoopPreheaderBB && "There is always a loop preheader.");
2441
2442 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2443 assert(SuccessorBB && "There is only a single successor.");
2444
2445 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2446 Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc());
2447
2448 Intrinsic::ID IntrID = Intrinsic::ctlz;
2449 Type *Ty = X->getType();
2450 unsigned Bitwidth = Ty->getScalarSizeInBits();
2451
2452 TargetTransformInfo::TargetCostKind CostKind =
2453 TargetTransformInfo::TCK_SizeAndLatency;
2454
2455 // The rewrite is considered to be unprofitable iff and only iff the
2456 // intrinsic/shift we'll use are not cheap. Note that we are okay with *just*
2457 // making the loop countable, even if nothing else changes.
2458 IntrinsicCostAttributes Attrs(
2459 IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getTrue()});
2460 InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2461 if (Cost > TargetTransformInfo::TCC_Basic) {
2462 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2463 " Intrinsic is too costly, not beneficial\n");
2464 return MadeChange;
2465 }
2466 if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) >
2467 TargetTransformInfo::TCC_Basic) {
2468 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n");
2469 return MadeChange;
2470 }
2471
2472 // Ok, transform appears worthwhile.
2473 MadeChange = true;
2474
2475 // Step 1: Compute the loop trip count.
2476
2477 Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty),
2478 BitPos->getName() + ".lowbitmask");
2479 Value *Mask =
2480 Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask");
2481 Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked");
2482 CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic(
2483 IntrID, Ty, {XMasked, /*is_zero_undef=*/Builder.getTrue()},
2484 /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros");
2485 Value *XMaskedNumActiveBits = Builder.CreateSub(
2486 ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros,
2487 XMasked->getName() + ".numactivebits", /*HasNUW=*/true,
2488 /*HasNSW=*/Bitwidth != 2);
2489 Value *XMaskedLeadingOnePos =
2490 Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty),
2491 XMasked->getName() + ".leadingonepos", /*HasNUW=*/false,
2492 /*HasNSW=*/Bitwidth > 2);
2493
2494 Value *LoopBackedgeTakenCount = Builder.CreateSub(
2495 BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount",
2496 /*HasNUW=*/true, /*HasNSW=*/true);
2497 // We know loop's backedge-taken count, but what's loop's trip count?
2498 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2499 Value *LoopTripCount =
2500 Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
2501 CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
2502 /*HasNSW=*/Bitwidth != 2);
2503
2504 // Step 2: Compute the recurrence's final value without a loop.
2505
2506 // NewX is always safe to compute, because `LoopBackedgeTakenCount`
2507 // will always be smaller than `bitwidth(X)`, i.e. we never get poison.
2508 Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount);
2509 NewX->takeName(XCurr);
2510 if (auto *I = dyn_cast<Instruction>(NewX))
2511 I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2512
2513 Value *NewXNext;
2514 // Rewriting XNext is more complicated, however, because `X << LoopTripCount`
2515 // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen
2516 // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know
2517 // that isn't the case, we'll need to emit an alternative, safe IR.
2518 if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() ||
2519 PatternMatch::match(
2520 BitPos, PatternMatch::m_SpecificInt_ICMP(
2521 ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(),
2522 Ty->getScalarSizeInBits() - 1))))
2523 NewXNext = Builder.CreateShl(X, LoopTripCount);
2524 else {
2525 // Otherwise, just additionally shift by one. It's the smallest solution,
2526 // alternatively, we could check that NewX is INT_MIN (or BitPos is )
2527 // and select 0 instead.
2528 NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1));
2529 }
2530
2531 NewXNext->takeName(XNext);
2532 if (auto *I = dyn_cast<Instruction>(NewXNext))
2533 I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2534
2535 // Step 3: Adjust the successor basic block to recieve the computed
2536 // recurrence's final value instead of the recurrence itself.
2537
2538 XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB);
2539 XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB);
2540
2541 // Step 4: Rewrite the loop into a countable form, with canonical IV.
2542
2543 // The new canonical induction variable.
2544 Builder.SetInsertPoint(&LoopHeaderBB->front());
2545 auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
2546
2547 // The induction itself.
2548 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2549 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
2550 auto *IVNext =
2551 Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next",
2552 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
2553
2554 // The loop trip count check.
2555 auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount,
2556 CurLoop->getName() + ".ivcheck");
2557 Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB);
2558 LoopHeaderBB->getTerminator()->eraseFromParent();
2559
2560 // Populate the IV PHI.
2561 IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
2562 IV->addIncoming(IVNext, LoopHeaderBB);
2563
2564 // Step 5: Forget the "non-computable" trip-count SCEV associated with the
2565 // loop. The loop would otherwise not be deleted even if it becomes empty.
2566
2567 SE->forgetLoop(CurLoop);
2568
2569 // Other passes will take care of actually deleting the loop if possible.
2570
2571 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n");
2572
2573 ++NumShiftUntilBitTest;
2574 return MadeChange;
2575 }
2576
2577 /// Return true if the idiom is detected in the loop.
2578 ///
2579 /// The core idiom we are trying to detect is:
2580 /// \code
2581 /// entry:
2582 /// <...>
2583 /// %start = <...>
2584 /// %extraoffset = <...>
2585 /// <...>
2586 /// br label %for.cond
2587 ///
2588 /// loop:
2589 /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2590 /// %nbits = add nsw i8 %iv, %extraoffset
2591 /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2592 /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2593 /// %iv.next = add i8 %iv, 1
2594 /// <...>
2595 /// br i1 %val.shifted.iszero, label %end, label %loop
2596 ///
2597 /// end:
2598 /// %iv.res = phi i8 [ %iv, %loop ] <...>
2599 /// %nbits.res = phi i8 [ %nbits, %loop ] <...>
2600 /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2601 /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2602 /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2603 /// <...>
2604 /// \endcode
detectShiftUntilZeroIdiom(Loop * CurLoop,ScalarEvolution * SE,Instruction * & ValShiftedIsZero,Intrinsic::ID & IntrinID,Instruction * & IV,Value * & Start,Value * & Val,const SCEV * & ExtraOffsetExpr,bool & InvertedCond)2605 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE,
2606 Instruction *&ValShiftedIsZero,
2607 Intrinsic::ID &IntrinID, Instruction *&IV,
2608 Value *&Start, Value *&Val,
2609 const SCEV *&ExtraOffsetExpr,
2610 bool &InvertedCond) {
2611 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2612 " Performing shift-until-zero idiom detection.\n");
2613
2614 // Give up if the loop has multiple blocks or multiple backedges.
2615 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2616 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2617 return false;
2618 }
2619
2620 Instruction *ValShifted, *NBits, *IVNext;
2621 Value *ExtraOffset;
2622
2623 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2624 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2625 assert(LoopPreheaderBB && "There is always a loop preheader.");
2626
2627 using namespace PatternMatch;
2628
2629 // Step 1: Check if the loop backedge, condition is in desirable form.
2630
2631 ICmpInst::Predicate Pred;
2632 BasicBlock *TrueBB, *FalseBB;
2633 if (!match(LoopHeaderBB->getTerminator(),
2634 m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB),
2635 m_BasicBlock(FalseBB))) ||
2636 !match(ValShiftedIsZero,
2637 m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) ||
2638 !ICmpInst::isEquality(Pred)) {
2639 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2640 return false;
2641 }
2642
2643 // Step 2: Check if the comparison's operand is in desirable form.
2644 // FIXME: Val could be a one-input PHI node, which we should look past.
2645 if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop),
2646 m_Instruction(NBits)))) {
2647 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n");
2648 return false;
2649 }
2650 IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz
2651 : Intrinsic::ctlz;
2652
2653 // Step 3: Check if the shift amount is in desirable form.
2654
2655 if (match(NBits, m_c_Add(m_Instruction(IV),
2656 m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2657 (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap()))
2658 ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset));
2659 else if (match(NBits,
2660 m_Sub(m_Instruction(IV),
2661 m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2662 NBits->hasNoSignedWrap())
2663 ExtraOffsetExpr = SE->getSCEV(ExtraOffset);
2664 else {
2665 IV = NBits;
2666 ExtraOffsetExpr = SE->getZero(NBits->getType());
2667 }
2668
2669 // Step 4: Check if the recurrence is in desirable form.
2670 auto *IVPN = dyn_cast<PHINode>(IV);
2671 if (!IVPN || IVPN->getParent() != LoopHeaderBB) {
2672 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2673 return false;
2674 }
2675
2676 Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB);
2677 IVNext = dyn_cast<Instruction>(IVPN->getIncomingValueForBlock(LoopHeaderBB));
2678
2679 if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) {
2680 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2681 return false;
2682 }
2683
2684 // Step 4: Check if the backedge's destinations are in desirable form.
2685
2686 assert(ICmpInst::isEquality(Pred) &&
2687 "Should only get equality predicates here.");
2688
2689 // cmp-br is commutative, so canonicalize to a single variant.
2690 InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ;
2691 if (InvertedCond) {
2692 Pred = ICmpInst::getInversePredicate(Pred);
2693 std::swap(TrueBB, FalseBB);
2694 }
2695
2696 // We expect to exit loop when comparison yields true,
2697 // so when it yields false we should branch back to loop header.
2698 if (FalseBB != LoopHeaderBB) {
2699 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2700 return false;
2701 }
2702
2703 // The new, countable, loop will certainly only run a known number of
2704 // iterations, It won't be infinite. But the old loop might be infinite
2705 // under certain conditions. For logical shifts, the value will become zero
2706 // after at most bitwidth(%Val) loop iterations. However, for arithmetic
2707 // right-shift, iff the sign bit was set, the value will never become zero,
2708 // and the loop may never finish.
2709 if (ValShifted->getOpcode() == Instruction::AShr &&
2710 !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) {
2711 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n");
2712 return false;
2713 }
2714
2715 // Okay, idiom checks out.
2716 return true;
2717 }
2718
2719 /// Look for the following loop:
2720 /// \code
2721 /// entry:
2722 /// <...>
2723 /// %start = <...>
2724 /// %extraoffset = <...>
2725 /// <...>
2726 /// br label %for.cond
2727 ///
2728 /// loop:
2729 /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2730 /// %nbits = add nsw i8 %iv, %extraoffset
2731 /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2732 /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2733 /// %iv.next = add i8 %iv, 1
2734 /// <...>
2735 /// br i1 %val.shifted.iszero, label %end, label %loop
2736 ///
2737 /// end:
2738 /// %iv.res = phi i8 [ %iv, %loop ] <...>
2739 /// %nbits.res = phi i8 [ %nbits, %loop ] <...>
2740 /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2741 /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2742 /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2743 /// <...>
2744 /// \endcode
2745 ///
2746 /// And transform it into:
2747 /// \code
2748 /// entry:
2749 /// <...>
2750 /// %start = <...>
2751 /// %extraoffset = <...>
2752 /// <...>
2753 /// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0)
2754 /// %val.numactivebits = sub i8 8, %val.numleadingzeros
2755 /// %extraoffset.neg = sub i8 0, %extraoffset
2756 /// %tmp = add i8 %val.numactivebits, %extraoffset.neg
2757 /// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start)
2758 /// %loop.tripcount = sub i8 %iv.final, %start
2759 /// br label %loop
2760 ///
2761 /// loop:
2762 /// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ]
2763 /// %loop.iv.next = add i8 %loop.iv, 1
2764 /// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount
2765 /// %iv = add i8 %loop.iv, %start
2766 /// <...>
2767 /// br i1 %loop.ivcheck, label %end, label %loop
2768 ///
2769 /// end:
2770 /// %iv.res = phi i8 [ %iv.final, %loop ] <...>
2771 /// <...>
2772 /// \endcode
recognizeShiftUntilZero()2773 bool LoopIdiomRecognize::recognizeShiftUntilZero() {
2774 bool MadeChange = false;
2775
2776 Instruction *ValShiftedIsZero;
2777 Intrinsic::ID IntrID;
2778 Instruction *IV;
2779 Value *Start, *Val;
2780 const SCEV *ExtraOffsetExpr;
2781 bool InvertedCond;
2782 if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV,
2783 Start, Val, ExtraOffsetExpr, InvertedCond)) {
2784 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2785 " shift-until-zero idiom detection failed.\n");
2786 return MadeChange;
2787 }
2788 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n");
2789
2790 // Ok, it is the idiom we were looking for, we *could* transform this loop,
2791 // but is it profitable to transform?
2792
2793 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2794 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2795 assert(LoopPreheaderBB && "There is always a loop preheader.");
2796
2797 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2798 assert(SuccessorBB && "There is only a single successor.");
2799
2800 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2801 Builder.SetCurrentDebugLocation(IV->getDebugLoc());
2802
2803 Type *Ty = Val->getType();
2804 unsigned Bitwidth = Ty->getScalarSizeInBits();
2805
2806 TargetTransformInfo::TargetCostKind CostKind =
2807 TargetTransformInfo::TCK_SizeAndLatency;
2808
2809 // The rewrite is considered to be unprofitable iff and only iff the
2810 // intrinsic we'll use are not cheap. Note that we are okay with *just*
2811 // making the loop countable, even if nothing else changes.
2812 IntrinsicCostAttributes Attrs(
2813 IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getFalse()});
2814 InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2815 if (Cost > TargetTransformInfo::TCC_Basic) {
2816 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2817 " Intrinsic is too costly, not beneficial\n");
2818 return MadeChange;
2819 }
2820
2821 // Ok, transform appears worthwhile.
2822 MadeChange = true;
2823
2824 bool OffsetIsZero = false;
2825 if (auto *ExtraOffsetExprC = dyn_cast<SCEVConstant>(ExtraOffsetExpr))
2826 OffsetIsZero = ExtraOffsetExprC->isZero();
2827
2828 // Step 1: Compute the loop's final IV value / trip count.
2829
2830 CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic(
2831 IntrID, Ty, {Val, /*is_zero_undef=*/Builder.getFalse()},
2832 /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros");
2833 Value *ValNumActiveBits = Builder.CreateSub(
2834 ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros,
2835 Val->getName() + ".numactivebits", /*HasNUW=*/true,
2836 /*HasNSW=*/Bitwidth != 2);
2837
2838 SCEVExpander Expander(*SE, *DL, "loop-idiom");
2839 Expander.setInsertPoint(&*Builder.GetInsertPoint());
2840 Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr);
2841
2842 Value *ValNumActiveBitsOffset = Builder.CreateAdd(
2843 ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset",
2844 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true);
2845 Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty},
2846 {ValNumActiveBitsOffset, Start},
2847 /*FMFSource=*/nullptr, "iv.final");
2848
2849 auto *LoopBackedgeTakenCount = cast<Instruction>(Builder.CreateSub(
2850 IVFinal, Start, CurLoop->getName() + ".backedgetakencount",
2851 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true));
2852 // FIXME: or when the offset was `add nuw`
2853
2854 // We know loop's backedge-taken count, but what's loop's trip count?
2855 Value *LoopTripCount =
2856 Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
2857 CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
2858 /*HasNSW=*/Bitwidth != 2);
2859
2860 // Step 2: Adjust the successor basic block to recieve the original
2861 // induction variable's final value instead of the orig. IV itself.
2862
2863 IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB);
2864
2865 // Step 3: Rewrite the loop into a countable form, with canonical IV.
2866
2867 // The new canonical induction variable.
2868 Builder.SetInsertPoint(&LoopHeaderBB->front());
2869 auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
2870
2871 // The induction itself.
2872 Builder.SetInsertPoint(LoopHeaderBB->getFirstNonPHI());
2873 auto *CIVNext =
2874 Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next",
2875 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
2876
2877 // The loop trip count check.
2878 auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount,
2879 CurLoop->getName() + ".ivcheck");
2880 auto *NewIVCheck = CIVCheck;
2881 if (InvertedCond) {
2882 NewIVCheck = Builder.CreateNot(CIVCheck);
2883 NewIVCheck->takeName(ValShiftedIsZero);
2884 }
2885
2886 // The original IV, but rebased to be an offset to the CIV.
2887 auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false,
2888 /*HasNSW=*/true); // FIXME: what about NUW?
2889 IVDePHId->takeName(IV);
2890
2891 // The loop terminator.
2892 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
2893 Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB);
2894 LoopHeaderBB->getTerminator()->eraseFromParent();
2895
2896 // Populate the IV PHI.
2897 CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
2898 CIV->addIncoming(CIVNext, LoopHeaderBB);
2899
2900 // Step 4: Forget the "non-computable" trip-count SCEV associated with the
2901 // loop. The loop would otherwise not be deleted even if it becomes empty.
2902
2903 SE->forgetLoop(CurLoop);
2904
2905 // Step 5: Try to cleanup the loop's body somewhat.
2906 IV->replaceAllUsesWith(IVDePHId);
2907 IV->eraseFromParent();
2908
2909 ValShiftedIsZero->replaceAllUsesWith(NewIVCheck);
2910 ValShiftedIsZero->eraseFromParent();
2911
2912 // Other passes will take care of actually deleting the loop if possible.
2913
2914 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n");
2915
2916 ++NumShiftUntilZero;
2917 return MadeChange;
2918 }
2919