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: memcmp, etc.
24//
25// This could recognize common matrix multiplies and dot product idioms and
26// replace them with calls to BLAS (if linked in??).
27//
28//===----------------------------------------------------------------------===//
29
30#include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
31#include "llvm/ADT/APInt.h"
32#include "llvm/ADT/ArrayRef.h"
33#include "llvm/ADT/DenseMap.h"
34#include "llvm/ADT/MapVector.h"
35#include "llvm/ADT/SetVector.h"
36#include "llvm/ADT/SmallPtrSet.h"
37#include "llvm/ADT/SmallVector.h"
38#include "llvm/ADT/Statistic.h"
39#include "llvm/ADT/StringRef.h"
40#include "llvm/Analysis/AliasAnalysis.h"
41#include "llvm/Analysis/CmpInstAnalysis.h"
42#include "llvm/Analysis/HashRecognize.h"
43#include "llvm/Analysis/LoopAccessAnalysis.h"
44#include "llvm/Analysis/LoopInfo.h"
45#include "llvm/Analysis/LoopPass.h"
46#include "llvm/Analysis/MemoryLocation.h"
47#include "llvm/Analysis/MemorySSA.h"
48#include "llvm/Analysis/MemorySSAUpdater.h"
49#include "llvm/Analysis/MustExecute.h"
50#include "llvm/Analysis/OptimizationRemarkEmitter.h"
51#include "llvm/Analysis/ScalarEvolution.h"
52#include "llvm/Analysis/ScalarEvolutionExpressions.h"
53#include "llvm/Analysis/ScalarEvolutionPatternMatch.h"
54#include "llvm/Analysis/TargetLibraryInfo.h"
55#include "llvm/Analysis/TargetTransformInfo.h"
56#include "llvm/Analysis/ValueTracking.h"
57#include "llvm/IR/BasicBlock.h"
58#include "llvm/IR/Constant.h"
59#include "llvm/IR/Constants.h"
60#include "llvm/IR/DataLayout.h"
61#include "llvm/IR/DebugLoc.h"
62#include "llvm/IR/DerivedTypes.h"
63#include "llvm/IR/Dominators.h"
64#include "llvm/IR/GlobalValue.h"
65#include "llvm/IR/GlobalVariable.h"
66#include "llvm/IR/IRBuilder.h"
67#include "llvm/IR/InstrTypes.h"
68#include "llvm/IR/Instruction.h"
69#include "llvm/IR/Instructions.h"
70#include "llvm/IR/IntrinsicInst.h"
71#include "llvm/IR/Intrinsics.h"
72#include "llvm/IR/LLVMContext.h"
73#include "llvm/IR/Module.h"
74#include "llvm/IR/PassManager.h"
75#include "llvm/IR/PatternMatch.h"
76#include "llvm/IR/ProfDataUtils.h"
77#include "llvm/IR/Type.h"
78#include "llvm/IR/User.h"
79#include "llvm/IR/Value.h"
80#include "llvm/IR/ValueHandle.h"
81#include "llvm/Support/Casting.h"
82#include "llvm/Support/CommandLine.h"
83#include "llvm/Support/Debug.h"
84#include "llvm/Support/InstructionCost.h"
85#include "llvm/Support/raw_ostream.h"
86#include "llvm/Transforms/Utils/BuildLibCalls.h"
87#include "llvm/Transforms/Utils/Local.h"
88#include "llvm/Transforms/Utils/LoopUtils.h"
89#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
90#include <algorithm>
91#include <cassert>
92#include <cstdint>
93#include <utility>
94
95using namespace llvm;
96using namespace SCEVPatternMatch;
97
98#define DEBUG_TYPE "loop-idiom"
99
100STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
101STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
102STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores");
103STATISTIC(NumStrLen, "Number of strlen's and wcslen's formed from loop loads");
104STATISTIC(
105 NumShiftUntilBitTest,
106 "Number of uncountable loops recognized as 'shift until bitttest' idiom");
107STATISTIC(NumShiftUntilZero,
108 "Number of uncountable loops recognized as 'shift until zero' idiom");
109
110namespace llvm {
111bool DisableLIRP::All;
112static cl::opt<bool, true>
113 DisableLIRPAll("disable-" DEBUG_TYPE "-all",
114 cl::desc("Options to disable Loop Idiom Recognize Pass."),
115 cl::location(L&: DisableLIRP::All), cl::init(Val: false),
116 cl::ReallyHidden);
117
118bool DisableLIRP::Memset;
119static cl::opt<bool, true>
120 DisableLIRPMemset("disable-" DEBUG_TYPE "-memset",
121 cl::desc("Proceed with loop idiom recognize pass, but do "
122 "not convert loop(s) to memset."),
123 cl::location(L&: DisableLIRP::Memset), cl::init(Val: false),
124 cl::ReallyHidden);
125
126bool DisableLIRP::Memcpy;
127static cl::opt<bool, true>
128 DisableLIRPMemcpy("disable-" DEBUG_TYPE "-memcpy",
129 cl::desc("Proceed with loop idiom recognize pass, but do "
130 "not convert loop(s) to memcpy."),
131 cl::location(L&: DisableLIRP::Memcpy), cl::init(Val: false),
132 cl::ReallyHidden);
133
134bool DisableLIRP::Strlen;
135static cl::opt<bool, true>
136 DisableLIRPStrlen("disable-loop-idiom-strlen",
137 cl::desc("Proceed with loop idiom recognize pass, but do "
138 "not convert loop(s) to strlen."),
139 cl::location(L&: DisableLIRP::Strlen), cl::init(Val: false),
140 cl::ReallyHidden);
141
142bool DisableLIRP::Wcslen;
143static cl::opt<bool, true>
144 EnableLIRPWcslen("disable-loop-idiom-wcslen",
145 cl::desc("Proceed with loop idiom recognize pass, "
146 "enable conversion of loop(s) to wcslen."),
147 cl::location(L&: DisableLIRP::Wcslen), cl::init(Val: false),
148 cl::ReallyHidden);
149
150bool DisableLIRP::HashRecognize;
151static cl::opt<bool, true>
152 DisableLIRPHashRecognize("disable-" DEBUG_TYPE "-hashrecognize",
153 cl::desc("Proceed with loop idiom recognize pass, "
154 "but do not optimize CRC loops."),
155 cl::location(L&: DisableLIRP::HashRecognize),
156 cl::init(Val: false), cl::ReallyHidden);
157
158static cl::opt<bool> UseLIRCodeSizeHeurs(
159 "use-lir-code-size-heurs",
160 cl::desc("Use loop idiom recognition code size heuristics when compiling "
161 "with -Os/-Oz"),
162 cl::init(Val: true), cl::Hidden);
163
164static cl::opt<bool> ForceMemsetPatternIntrinsic(
165 "loop-idiom-force-memset-pattern-intrinsic",
166 cl::desc("Use memset.pattern intrinsic whenever possible"), cl::init(Val: false),
167 cl::Hidden);
168
169extern cl::opt<bool> ProfcheckDisableMetadataFixes;
170
171} // namespace llvm
172
173namespace {
174
175class LoopIdiomRecognize {
176 Loop *CurLoop = nullptr;
177 AliasAnalysis *AA;
178 DominatorTree *DT;
179 LoopInfo *LI;
180 ScalarEvolution *SE;
181 TargetLibraryInfo *TLI;
182 const TargetTransformInfo *TTI;
183 const DataLayout *DL;
184 OptimizationRemarkEmitter &ORE;
185 bool ApplyCodeSizeHeuristics;
186 std::unique_ptr<MemorySSAUpdater> MSSAU;
187
188public:
189 explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
190 LoopInfo *LI, ScalarEvolution *SE,
191 TargetLibraryInfo *TLI,
192 const TargetTransformInfo *TTI, MemorySSA *MSSA,
193 const DataLayout *DL,
194 OptimizationRemarkEmitter &ORE)
195 : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) {
196 if (MSSA)
197 MSSAU = std::make_unique<MemorySSAUpdater>(args&: MSSA);
198 }
199
200 bool runOnLoop(Loop *L);
201
202private:
203 using StoreList = SmallVector<StoreInst *, 8>;
204 using StoreListMap = MapVector<Value *, StoreList>;
205
206 StoreListMap StoreRefsForMemset;
207 StoreListMap StoreRefsForMemsetPattern;
208 StoreList StoreRefsForMemcpy;
209 bool HasMemset;
210 bool HasMemsetPattern;
211 bool HasMemcpy;
212
213 /// Return code for isLegalStore()
214 enum LegalStoreKind {
215 None = 0,
216 Memset,
217 MemsetPattern,
218 Memcpy,
219 UnorderedAtomicMemcpy,
220 DontUse // Dummy retval never to be used. Allows catching errors in retval
221 // handling.
222 };
223
224 /// \name Countable Loop Idiom Handling
225 /// @{
226
227 bool runOnCountableLoop();
228 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
229 SmallVectorImpl<BasicBlock *> &ExitBlocks);
230
231 void collectStores(BasicBlock *BB);
232 LegalStoreKind isLegalStore(StoreInst *SI);
233 enum class ForMemset { No, Yes };
234 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
235 ForMemset For);
236
237 template <typename MemInst>
238 bool processLoopMemIntrinsic(
239 BasicBlock *BB,
240 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
241 const SCEV *BECount);
242 bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount);
243 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
244
245 bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV,
246 MaybeAlign StoreAlignment, Value *StoredVal,
247 Instruction *TheStore,
248 SmallPtrSetImpl<Instruction *> &Stores,
249 const SCEVAddRecExpr *Ev, const SCEV *BECount,
250 bool IsNegStride, bool IsLoopMemset = false);
251 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
252 bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr,
253 const SCEV *StoreSize, MaybeAlign StoreAlign,
254 MaybeAlign LoadAlign, Instruction *TheStore,
255 Instruction *TheLoad,
256 const SCEVAddRecExpr *StoreEv,
257 const SCEVAddRecExpr *LoadEv,
258 const SCEV *BECount);
259 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
260 bool IsLoopMemset = false);
261 bool optimizeCRCLoop(const PolynomialInfo &Info);
262
263 /// @}
264 /// \name Noncountable Loop Idiom Handling
265 /// @{
266
267 bool runOnNoncountableLoop();
268
269 bool recognizePopcount();
270 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
271 PHINode *CntPhi, Value *Var);
272 bool isProfitableToInsertFFS(Intrinsic::ID IntrinID, Value *InitX,
273 bool ZeroCheck, size_t CanonicalSize);
274 bool insertFFSIfProfitable(Intrinsic::ID IntrinID, Value *InitX,
275 Instruction *DefX, PHINode *CntPhi,
276 Instruction *CntInst);
277 bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz
278 bool recognizeShiftUntilLessThan();
279 void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB,
280 Instruction *CntInst, PHINode *CntPhi,
281 Value *Var, Instruction *DefX,
282 const DebugLoc &DL, bool ZeroCheck,
283 bool IsCntPhiUsedOutsideLoop,
284 bool InsertSub = false);
285
286 bool recognizeShiftUntilBitTest();
287 bool recognizeShiftUntilZero();
288 bool recognizeAndInsertStrLen();
289
290 /// @}
291};
292} // end anonymous namespace
293
294PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
295 LoopStandardAnalysisResults &AR,
296 LPMUpdater &) {
297 if (DisableLIRP::All)
298 return PreservedAnalyses::all();
299
300 const auto *DL = &L.getHeader()->getDataLayout();
301
302 // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
303 // pass. Function analyses need to be preserved across loop transformations
304 // but ORE cannot be preserved (see comment before the pass definition).
305 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
306
307 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI,
308 AR.MSSA, DL, ORE);
309 if (!LIR.runOnLoop(L: &L))
310 return PreservedAnalyses::all();
311
312 auto PA = getLoopPassPreservedAnalyses();
313 if (AR.MSSA)
314 PA.preserve<MemorySSAAnalysis>();
315 return PA;
316}
317
318static void deleteDeadInstruction(Instruction *I) {
319 I->replaceAllUsesWith(V: PoisonValue::get(T: I->getType()));
320 I->eraseFromParent();
321}
322
323//===----------------------------------------------------------------------===//
324//
325// Implementation of LoopIdiomRecognize
326//
327//===----------------------------------------------------------------------===//
328
329bool LoopIdiomRecognize::runOnLoop(Loop *L) {
330 CurLoop = L;
331 // If the loop could not be converted to canonical form, it must have an
332 // indirectbr in it, just give up.
333 if (!L->getLoopPreheader())
334 return false;
335
336 // Disable loop idiom recognition if the function's name is a common idiom.
337 StringRef Name = L->getHeader()->getParent()->getName();
338 if (Name == "memset" || Name == "memcpy" || Name == "strlen" ||
339 Name == "wcslen")
340 return false;
341
342 // Determine if code size heuristics need to be applied.
343 ApplyCodeSizeHeuristics =
344 L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs;
345
346 HasMemset = TLI->has(F: LibFunc_memset);
347 // TODO: Unconditionally enable use of the memset pattern intrinsic (or at
348 // least, opt-in via target hook) once we are confident it will never result
349 // in worse codegen than without. For now, use it only when the target
350 // supports memset_pattern16 libcall (or unless this is overridden by
351 // command line option).
352 HasMemsetPattern = TLI->has(F: LibFunc_memset_pattern16);
353 HasMemcpy = TLI->has(F: LibFunc_memcpy);
354
355 if (HasMemset || HasMemsetPattern || ForceMemsetPatternIntrinsic ||
356 HasMemcpy || !DisableLIRP::HashRecognize)
357 if (SE->hasLoopInvariantBackedgeTakenCount(L))
358 return runOnCountableLoop();
359
360 return runOnNoncountableLoop();
361}
362
363bool LoopIdiomRecognize::runOnCountableLoop() {
364 const SCEV *BECount = SE->getBackedgeTakenCount(L: CurLoop);
365 assert(!isa<SCEVCouldNotCompute>(BECount) &&
366 "runOnCountableLoop() called on a loop without a predictable"
367 "backedge-taken count");
368
369 // If this loop executes exactly one time, then it should be peeled, not
370 // optimized by this pass.
371 if (BECount->isZero())
372 return false;
373
374 SmallVector<BasicBlock *, 8> ExitBlocks;
375 CurLoop->getUniqueExitBlocks(ExitBlocks);
376
377 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
378 << CurLoop->getHeader()->getParent()->getName()
379 << "] Countable Loop %" << CurLoop->getHeader()->getName()
380 << "\n");
381
382 // The following transforms hoist stores/memsets into the loop pre-header.
383 // Give up if the loop has instructions that may throw.
384 SimpleLoopSafetyInfo SafetyInfo;
385 SafetyInfo.computeLoopSafetyInfo(CurLoop);
386 if (SafetyInfo.anyBlockMayThrow())
387 return false;
388
389 bool MadeChange = false;
390
391 // Scan all the blocks in the loop that are not in subloops.
392 for (auto *BB : CurLoop->getBlocks()) {
393 // Ignore blocks in subloops.
394 if (LI->getLoopFor(BB) != CurLoop)
395 continue;
396
397 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
398 }
399
400 // Optimize a CRC loop if HashRecognize found one, provided we're not
401 // optimizing for size.
402 if (!DisableLIRP::HashRecognize && !ApplyCodeSizeHeuristics)
403 if (auto Res = HashRecognize(*CurLoop, *SE).getResult())
404 optimizeCRCLoop(Info: *Res);
405
406 return MadeChange;
407}
408
409static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
410 const SCEVConstant *ConstStride = cast<SCEVConstant>(Val: StoreEv->getOperand(i: 1));
411 return ConstStride->getAPInt();
412}
413
414/// getMemSetPatternValue - If a strided store of the specified value is safe to
415/// turn into a memset.patternn intrinsic, return the Constant that should
416/// be passed in. Otherwise, return null.
417///
418/// TODO this function could allow more constants than it does today (e.g.
419/// those over 16 bytes) now it has transitioned to being used for the
420/// memset.pattern intrinsic rather than directly the memset_pattern16
421/// libcall.
422static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
423 // FIXME: This could check for UndefValue because it can be merged into any
424 // other valid pattern.
425
426 // If the value isn't a constant, we can't promote it to being in a constant
427 // array. We could theoretically do a store to an alloca or something, but
428 // that doesn't seem worthwhile.
429 Constant *C = dyn_cast<Constant>(Val: V);
430 if (!C || isa<ConstantExpr>(Val: C))
431 return nullptr;
432
433 // Only handle simple values that are a power of two bytes in size.
434 uint64_t Size = DL->getTypeSizeInBits(Ty: V->getType());
435 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
436 return nullptr;
437
438 // Don't care enough about darwin/ppc to implement this.
439 if (DL->isBigEndian())
440 return nullptr;
441
442 // Convert to size in bytes.
443 Size /= 8;
444
445 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
446 // if the top and bottom are the same (e.g. for vectors and large integers).
447 if (Size > 16)
448 return nullptr;
449
450 // For now, don't handle types that aren't int, floats, or pointers.
451 Type *CTy = C->getType();
452 if (!CTy->isIntOrPtrTy() && !CTy->isFloatingPointTy())
453 return nullptr;
454
455 return C;
456}
457
458LoopIdiomRecognize::LegalStoreKind
459LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
460 // Don't touch volatile stores.
461 if (SI->isVolatile())
462 return LegalStoreKind::None;
463 // We only want simple or unordered-atomic stores.
464 if (!SI->isUnordered())
465 return LegalStoreKind::None;
466
467 // Avoid merging nontemporal stores.
468 if (SI->getMetadata(KindID: LLVMContext::MD_nontemporal))
469 return LegalStoreKind::None;
470
471 Value *StoredVal = SI->getValueOperand();
472 Value *StorePtr = SI->getPointerOperand();
473
474 // Don't convert stores of non-integral pointer types to memsets (which stores
475 // integers).
476 if (DL->isNonIntegralPointerType(Ty: StoredVal->getType()->getScalarType()))
477 return LegalStoreKind::None;
478
479 // Reject stores that are so large that they overflow an unsigned.
480 // When storing out scalable vectors we bail out for now, since the code
481 // below currently only works for constant strides.
482 TypeSize SizeInBits = DL->getTypeSizeInBits(Ty: StoredVal->getType());
483 if (SizeInBits.isScalable() || (SizeInBits.getFixedValue() & 7) ||
484 (SizeInBits.getFixedValue() >> 32) != 0)
485 return LegalStoreKind::None;
486
487 // See if the pointer expression is an AddRec like {base,+,1} on the current
488 // loop, which indicates a strided store. If we have something else, it's a
489 // random store we can't handle.
490 const SCEV *StoreEv = SE->getSCEV(V: StorePtr);
491 const SCEVConstant *Stride;
492 if (!match(S: StoreEv, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_SCEVConstant(V&: Stride),
493 L: m_SpecificLoop(L: CurLoop))))
494 return LegalStoreKind::None;
495
496 // See if the store can be turned into a memset.
497
498 // If the stored value is a byte-wise value (like i32 -1), then it may be
499 // turned into a memset of i8 -1, assuming that all the consecutive bytes
500 // are stored. A store of i32 0x01020304 can never be turned into a memset,
501 // but it can be turned into memset_pattern if the target supports it.
502 Value *SplatValue = isBytewiseValue(V: StoredVal, DL: *DL);
503
504 // Note: memset and memset_pattern on unordered-atomic is yet not supported
505 bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple();
506
507 // If we're allowed to form a memset, and the stored value would be
508 // acceptable for memset, use it.
509 if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset &&
510 // Verify that the stored value is loop invariant. If not, we can't
511 // promote the memset.
512 CurLoop->isLoopInvariant(V: SplatValue)) {
513 // It looks like we can use SplatValue.
514 return LegalStoreKind::Memset;
515 }
516 if (!UnorderedAtomic && (HasMemsetPattern || ForceMemsetPatternIntrinsic) &&
517 !DisableLIRP::Memset &&
518 // Don't create memset_pattern16s with address spaces.
519 StorePtr->getType()->getPointerAddressSpace() == 0 &&
520 getMemSetPatternValue(V: StoredVal, DL)) {
521 // It looks like we can use PatternValue!
522 return LegalStoreKind::MemsetPattern;
523 }
524
525 // Otherwise, see if the store can be turned into a memcpy.
526 if (HasMemcpy && !DisableLIRP::Memcpy) {
527 // Check to see if the stride matches the size of the store. If so, then we
528 // know that every byte is touched in the loop.
529 unsigned StoreSize = DL->getTypeStoreSize(Ty: SI->getValueOperand()->getType());
530 APInt StrideAP = Stride->getAPInt();
531 if (StoreSize != StrideAP && StoreSize != -StrideAP)
532 return LegalStoreKind::None;
533
534 // The store must be feeding a non-volatile load.
535 LoadInst *LI = dyn_cast<LoadInst>(Val: SI->getValueOperand());
536
537 // Only allow non-volatile loads
538 if (!LI || LI->isVolatile())
539 return LegalStoreKind::None;
540 // Only allow simple or unordered-atomic loads
541 if (!LI->isUnordered())
542 return LegalStoreKind::None;
543
544 // See if the pointer expression is an AddRec like {base,+,1} on the current
545 // loop, which indicates a strided load. If we have something else, it's a
546 // random load we can't handle.
547 const SCEV *LoadEv = SE->getSCEV(V: LI->getPointerOperand());
548
549 // The store and load must share the same stride.
550 if (!match(S: LoadEv, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_scev_Specific(S: Stride),
551 L: m_SpecificLoop(L: CurLoop))))
552 return LegalStoreKind::None;
553
554 // Success. This store can be converted into a memcpy.
555 UnorderedAtomic = UnorderedAtomic || LI->isAtomic();
556 return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy
557 : LegalStoreKind::Memcpy;
558 }
559 // This store can't be transformed into a memset/memcpy.
560 return LegalStoreKind::None;
561}
562
563void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
564 StoreRefsForMemset.clear();
565 StoreRefsForMemsetPattern.clear();
566 StoreRefsForMemcpy.clear();
567 for (Instruction &I : *BB) {
568 StoreInst *SI = dyn_cast<StoreInst>(Val: &I);
569 if (!SI)
570 continue;
571
572 // Make sure this is a strided store with a constant stride.
573 switch (isLegalStore(SI)) {
574 case LegalStoreKind::None:
575 // Nothing to do
576 break;
577 case LegalStoreKind::Memset: {
578 // Find the base pointer.
579 Value *Ptr = getUnderlyingObject(V: SI->getPointerOperand());
580 StoreRefsForMemset[Ptr].push_back(Elt: SI);
581 } break;
582 case LegalStoreKind::MemsetPattern: {
583 // Find the base pointer.
584 Value *Ptr = getUnderlyingObject(V: SI->getPointerOperand());
585 StoreRefsForMemsetPattern[Ptr].push_back(Elt: SI);
586 } break;
587 case LegalStoreKind::Memcpy:
588 case LegalStoreKind::UnorderedAtomicMemcpy:
589 StoreRefsForMemcpy.push_back(Elt: SI);
590 break;
591 default:
592 assert(false && "unhandled return value");
593 break;
594 }
595 }
596}
597
598/// runOnLoopBlock - Process the specified block, which lives in a counted loop
599/// with the specified backedge count. This block is known to be in the current
600/// loop and not in any subloops.
601bool LoopIdiomRecognize::runOnLoopBlock(
602 BasicBlock *BB, const SCEV *BECount,
603 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
604 // We can only promote stores in this block if they are unconditionally
605 // executed in the loop. For a block to be unconditionally executed, it has
606 // to dominate all the exit blocks of the loop. Verify this now.
607 for (BasicBlock *ExitBlock : ExitBlocks)
608 if (!DT->dominates(A: BB, B: ExitBlock))
609 return false;
610
611 bool MadeChange = false;
612 // Look for store instructions, which may be optimized to memset/memcpy.
613 collectStores(BB);
614
615 // Look for a single store or sets of stores with a common base, which can be
616 // optimized into a memset (memset_pattern). The latter most commonly happens
617 // with structs and handunrolled loops.
618 for (auto &SL : StoreRefsForMemset)
619 MadeChange |= processLoopStores(SL&: SL.second, BECount, For: ForMemset::Yes);
620
621 for (auto &SL : StoreRefsForMemsetPattern)
622 MadeChange |= processLoopStores(SL&: SL.second, BECount, For: ForMemset::No);
623
624 // Optimize the store into a memcpy, if it feeds an similarly strided load.
625 for (auto &SI : StoreRefsForMemcpy)
626 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
627
628 MadeChange |= processLoopMemIntrinsic<MemCpyInst>(
629 BB, Processor: &LoopIdiomRecognize::processLoopMemCpy, BECount);
630 MadeChange |= processLoopMemIntrinsic<MemSetInst>(
631 BB, Processor: &LoopIdiomRecognize::processLoopMemSet, BECount);
632
633 return MadeChange;
634}
635
636/// See if this store(s) can be promoted to a memset.
637bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
638 const SCEV *BECount, ForMemset For) {
639 // Try to find consecutive stores that can be transformed into memsets.
640 SetVector<StoreInst *> Heads, Tails;
641 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
642
643 // Do a quadratic search on all of the given stores and find
644 // all of the pairs of stores that follow each other.
645 SmallVector<unsigned, 16> IndexQueue;
646 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
647 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
648
649 Value *FirstStoredVal = SL[i]->getValueOperand();
650 Value *FirstStorePtr = SL[i]->getPointerOperand();
651 const SCEVAddRecExpr *FirstStoreEv =
652 cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: FirstStorePtr));
653 APInt FirstStride = getStoreStride(StoreEv: FirstStoreEv);
654 unsigned FirstStoreSize = DL->getTypeStoreSize(Ty: SL[i]->getValueOperand()->getType());
655
656 // See if we can optimize just this store in isolation.
657 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
658 Heads.insert(X: SL[i]);
659 continue;
660 }
661
662 Value *FirstSplatValue = nullptr;
663 Constant *FirstPatternValue = nullptr;
664
665 if (For == ForMemset::Yes)
666 FirstSplatValue = isBytewiseValue(V: FirstStoredVal, DL: *DL);
667 else
668 FirstPatternValue = getMemSetPatternValue(V: FirstStoredVal, DL);
669
670 assert((FirstSplatValue || FirstPatternValue) &&
671 "Expected either splat value or pattern value.");
672
673 IndexQueue.clear();
674 // If a store has multiple consecutive store candidates, search Stores
675 // array according to the sequence: from i+1 to e, then from i-1 to 0.
676 // This is because usually pairing with immediate succeeding or preceding
677 // candidate create the best chance to find memset opportunity.
678 unsigned j = 0;
679 for (j = i + 1; j < e; ++j)
680 IndexQueue.push_back(Elt: j);
681 for (j = i; j > 0; --j)
682 IndexQueue.push_back(Elt: j - 1);
683
684 for (auto &k : IndexQueue) {
685 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
686 Value *SecondStorePtr = SL[k]->getPointerOperand();
687 const SCEVAddRecExpr *SecondStoreEv =
688 cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: SecondStorePtr));
689 APInt SecondStride = getStoreStride(StoreEv: SecondStoreEv);
690
691 if (FirstStride != SecondStride)
692 continue;
693
694 Value *SecondStoredVal = SL[k]->getValueOperand();
695 Value *SecondSplatValue = nullptr;
696 Constant *SecondPatternValue = nullptr;
697
698 if (For == ForMemset::Yes)
699 SecondSplatValue = isBytewiseValue(V: SecondStoredVal, DL: *DL);
700 else
701 SecondPatternValue = getMemSetPatternValue(V: SecondStoredVal, DL);
702
703 assert((SecondSplatValue || SecondPatternValue) &&
704 "Expected either splat value or pattern value.");
705
706 if (isConsecutiveAccess(A: SL[i], B: SL[k], DL: *DL, SE&: *SE, CheckType: false)) {
707 if (For == ForMemset::Yes) {
708 if (isa<UndefValue>(Val: FirstSplatValue))
709 FirstSplatValue = SecondSplatValue;
710 if (FirstSplatValue != SecondSplatValue)
711 continue;
712 } else {
713 if (isa<UndefValue>(Val: FirstPatternValue))
714 FirstPatternValue = SecondPatternValue;
715 if (FirstPatternValue != SecondPatternValue)
716 continue;
717 }
718 Tails.insert(X: SL[k]);
719 Heads.insert(X: SL[i]);
720 ConsecutiveChain[SL[i]] = SL[k];
721 break;
722 }
723 }
724 }
725
726 // We may run into multiple chains that merge into a single chain. We mark the
727 // stores that we transformed so that we don't visit the same store twice.
728 SmallPtrSet<Value *, 16> TransformedStores;
729 bool Changed = false;
730
731 // For stores that start but don't end a link in the chain:
732 for (StoreInst *I : Heads) {
733 if (Tails.count(key: I))
734 continue;
735
736 // We found a store instr that starts a chain. Now follow the chain and try
737 // to transform it.
738 SmallPtrSet<Instruction *, 8> AdjacentStores;
739 StoreInst *HeadStore = I;
740 unsigned StoreSize = 0;
741
742 // Collect the chain into a list.
743 while (Tails.count(key: I) || Heads.count(key: I)) {
744 if (TransformedStores.count(Ptr: I))
745 break;
746 AdjacentStores.insert(Ptr: I);
747
748 StoreSize += DL->getTypeStoreSize(Ty: I->getValueOperand()->getType());
749 // Move to the next value in the chain.
750 I = ConsecutiveChain[I];
751 }
752
753 Value *StoredVal = HeadStore->getValueOperand();
754 Value *StorePtr = HeadStore->getPointerOperand();
755 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: StorePtr));
756 APInt Stride = getStoreStride(StoreEv);
757
758 // Check to see if the stride matches the size of the stores. If so, then
759 // we know that every byte is touched in the loop.
760 if (StoreSize != Stride && StoreSize != -Stride)
761 continue;
762
763 bool IsNegStride = StoreSize == -Stride;
764
765 Type *IntIdxTy = DL->getIndexType(PtrTy: StorePtr->getType());
766 const SCEV *StoreSizeSCEV = SE->getConstant(Ty: IntIdxTy, V: StoreSize);
767 if (processLoopStridedStore(DestPtr: StorePtr, StoreSizeSCEV,
768 StoreAlignment: MaybeAlign(HeadStore->getAlign()), StoredVal,
769 TheStore: HeadStore, Stores&: AdjacentStores, Ev: StoreEv, BECount,
770 IsNegStride)) {
771 TransformedStores.insert_range(R&: AdjacentStores);
772 Changed = true;
773 }
774 }
775
776 return Changed;
777}
778
779/// processLoopMemIntrinsic - Template function for calling different processor
780/// functions based on mem intrinsic type.
781template <typename MemInst>
782bool LoopIdiomRecognize::processLoopMemIntrinsic(
783 BasicBlock *BB,
784 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
785 const SCEV *BECount) {
786 bool MadeChange = false;
787 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
788 Instruction *Inst = &*I++;
789 // Look for memory instructions, which may be optimized to a larger one.
790 if (MemInst *MI = dyn_cast<MemInst>(Inst)) {
791 WeakTrackingVH InstPtr(&*I);
792 if (!(this->*Processor)(MI, BECount))
793 continue;
794 MadeChange = true;
795
796 // If processing the instruction invalidated our iterator, start over from
797 // the top of the block.
798 if (!InstPtr)
799 I = BB->begin();
800 }
801 }
802 return MadeChange;
803}
804
805/// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy
806bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
807 const SCEV *BECount) {
808 // We can only handle non-volatile memcpys with a constant size.
809 if (MCI->isVolatile() || !isa<ConstantInt>(Val: MCI->getLength()))
810 return false;
811
812 // If we're not allowed to hack on memcpy, we fail.
813 if ((!HasMemcpy && !MCI->isForceInlined()) || DisableLIRP::Memcpy)
814 return false;
815
816 Value *Dest = MCI->getDest();
817 Value *Source = MCI->getSource();
818 if (!Dest || !Source)
819 return false;
820
821 // See if the load and store pointer expressions are AddRec like {base,+,1} on
822 // the current loop, which indicates a strided load and store. If we have
823 // something else, it's a random load or store we can't handle.
824 const SCEV *StoreEv = SE->getSCEV(V: Dest);
825 const SCEV *LoadEv = SE->getSCEV(V: Source);
826 const APInt *StoreStrideValue, *LoadStrideValue;
827 if (!match(S: StoreEv,
828 P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_scev_APInt(C&: StoreStrideValue),
829 L: m_SpecificLoop(L: CurLoop))) ||
830 !match(S: LoadEv,
831 P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_scev_APInt(C&: LoadStrideValue),
832 L: m_SpecificLoop(L: CurLoop))))
833 return false;
834
835 // Reject memcpys that are so large that they overflow an unsigned.
836 uint64_t SizeInBytes = cast<ConstantInt>(Val: MCI->getLength())->getZExtValue();
837 if ((SizeInBytes >> 32) != 0)
838 return false;
839
840 // Huge stride value - give up
841 if (StoreStrideValue->getBitWidth() > 64 ||
842 LoadStrideValue->getBitWidth() > 64)
843 return false;
844
845 if (SizeInBytes != *StoreStrideValue && SizeInBytes != -*StoreStrideValue) {
846 ORE.emit(RemarkBuilder: [&]() {
847 return OptimizationRemarkMissed(DEBUG_TYPE, "SizeStrideUnequal", MCI)
848 << ore::NV("Inst", "memcpy") << " in "
849 << ore::NV("Function", MCI->getFunction())
850 << " function will not be hoisted: "
851 << ore::NV("Reason", "memcpy size is not equal to stride");
852 });
853 return false;
854 }
855
856 int64_t StoreStrideInt = StoreStrideValue->getSExtValue();
857 int64_t LoadStrideInt = LoadStrideValue->getSExtValue();
858 // Check if the load stride matches the store stride.
859 if (StoreStrideInt != LoadStrideInt)
860 return false;
861
862 return processLoopStoreOfLoopLoad(
863 DestPtr: Dest, SourcePtr: Source, StoreSize: SE->getConstant(Ty: Dest->getType(), V: SizeInBytes),
864 StoreAlign: MCI->getDestAlign(), LoadAlign: MCI->getSourceAlign(), TheStore: MCI, TheLoad: MCI,
865 StoreEv: cast<SCEVAddRecExpr>(Val: StoreEv), LoadEv: cast<SCEVAddRecExpr>(Val: LoadEv), BECount);
866}
867
868/// processLoopMemSet - See if this memset can be promoted to a large memset.
869bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
870 const SCEV *BECount) {
871 // We can only handle non-volatile memsets.
872 if (MSI->isVolatile())
873 return false;
874
875 // If we're not allowed to hack on memset, we fail.
876 if (!HasMemset || DisableLIRP::Memset)
877 return false;
878
879 Value *Pointer = MSI->getDest();
880
881 // See if the pointer expression is an AddRec like {base,+,1} on the current
882 // loop, which indicates a strided store. If we have something else, it's a
883 // random store we can't handle.
884 const SCEV *Ev = SE->getSCEV(V: Pointer);
885 const SCEV *PointerStrideSCEV;
886 if (!match(S: Ev, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_SCEV(V&: PointerStrideSCEV),
887 L: m_SpecificLoop(L: CurLoop)))) {
888 LLVM_DEBUG(dbgs() << " Pointer is not affine, abort\n");
889 return false;
890 }
891
892 SCEVUse MemsetSizeSCEV = SE->getSCEV(V: MSI->getLength());
893
894 bool IsNegStride = false;
895 const bool IsConstantSize = isa<ConstantInt>(Val: MSI->getLength());
896
897 if (IsConstantSize) {
898 // Memset size is constant.
899 // Check if the pointer stride matches the memset size. If so, then
900 // we know that every byte is touched in the loop.
901 LLVM_DEBUG(dbgs() << " memset size is constant\n");
902 uint64_t SizeInBytes = cast<ConstantInt>(Val: MSI->getLength())->getZExtValue();
903 const APInt *Stride;
904 if (!match(S: PointerStrideSCEV, P: m_scev_APInt(C&: Stride)))
905 return false;
906
907 if (SizeInBytes != *Stride && SizeInBytes != -*Stride)
908 return false;
909
910 IsNegStride = SizeInBytes == -*Stride;
911 } else {
912 // Memset size is non-constant.
913 // Check if the pointer stride matches the memset size.
914 // To be conservative, the pass would not promote pointers that aren't in
915 // address space zero. Also, the pass only handles memset length and stride
916 // that are invariant for the top level loop.
917 LLVM_DEBUG(dbgs() << " memset size is non-constant\n");
918 if (Pointer->getType()->getPointerAddressSpace() != 0) {
919 LLVM_DEBUG(dbgs() << " pointer is not in address space zero, "
920 << "abort\n");
921 return false;
922 }
923 if (!SE->isLoopInvariant(S: MemsetSizeSCEV, L: CurLoop)) {
924 LLVM_DEBUG(dbgs() << " memset size is not a loop-invariant, "
925 << "abort\n");
926 return false;
927 }
928
929 // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV
930 IsNegStride = PointerStrideSCEV->isNonConstantNegative();
931 SCEVUse PositiveStrideSCEV =
932 IsNegStride ? SCEVUse(SE->getNegativeSCEV(V: PointerStrideSCEV))
933 : SCEVUse(PointerStrideSCEV);
934 LLVM_DEBUG(dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"
935 << " PositiveStrideSCEV: " << *PositiveStrideSCEV
936 << "\n");
937
938 if (PositiveStrideSCEV != MemsetSizeSCEV) {
939 // If an expression is covered by the loop guard, compare again and
940 // proceed with optimization if equal.
941 const SCEV *FoldedPositiveStride =
942 SE->applyLoopGuards(Expr: PositiveStrideSCEV, L: CurLoop);
943 const SCEV *FoldedMemsetSize =
944 SE->applyLoopGuards(Expr: MemsetSizeSCEV, L: CurLoop);
945
946 LLVM_DEBUG(dbgs() << " Try to fold SCEV based on loop guard\n"
947 << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n"
948 << " FoldedPositiveStride: " << *FoldedPositiveStride
949 << "\n");
950
951 if (FoldedPositiveStride != FoldedMemsetSize) {
952 LLVM_DEBUG(dbgs() << " SCEV don't match, abort\n");
953 return false;
954 }
955 }
956 }
957
958 // Verify that the memset value is loop invariant. If not, we can't promote
959 // the memset.
960 Value *SplatValue = MSI->getValue();
961 if (!SplatValue || !CurLoop->isLoopInvariant(V: SplatValue))
962 return false;
963
964 SmallPtrSet<Instruction *, 1> MSIs;
965 MSIs.insert(Ptr: MSI);
966 return processLoopStridedStore(DestPtr: Pointer, StoreSizeSCEV: SE->getSCEV(V: MSI->getLength()),
967 StoreAlignment: MSI->getDestAlign(), StoredVal: SplatValue, TheStore: MSI, Stores&: MSIs,
968 Ev: cast<SCEVAddRecExpr>(Val: Ev), BECount, IsNegStride,
969 /*IsLoopMemset=*/true);
970}
971
972/// mayLoopAccessLocation - Return true if the specified loop might access the
973/// specified pointer location, which is a loop-strided access. The 'Access'
974/// argument specifies what the verboten forms of access are (read or write).
975static bool
976mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
977 const SCEV *BECount, const SCEV *StoreSizeSCEV,
978 AliasAnalysis &AA,
979 SmallPtrSetImpl<Instruction *> &IgnoredInsts) {
980 // Get the location that may be stored across the loop. Since the access is
981 // strided positively through memory, we say that the modified location starts
982 // at the pointer and has infinite size.
983 LocationSize AccessSize = LocationSize::afterPointer();
984
985 // If the loop iterates a fixed number of times, we can refine the access size
986 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
987 const APInt *BECst, *ConstSize;
988 if (match(S: BECount, P: m_scev_APInt(C&: BECst)) &&
989 match(S: StoreSizeSCEV, P: m_scev_APInt(C&: ConstSize))) {
990 std::optional<uint64_t> BEInt = BECst->tryZExtValue();
991 std::optional<uint64_t> SizeInt = ConstSize->tryZExtValue();
992 // FIXME: Should this check for overflow?
993 if (BEInt && SizeInt)
994 AccessSize = LocationSize::precise(Value: (*BEInt + 1) * *SizeInt);
995 }
996
997 // TODO: For this to be really effective, we have to dive into the pointer
998 // operand in the store. Store to &A[i] of 100 will always return may alias
999 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
1000 // which will then no-alias a store to &A[100].
1001 MemoryLocation StoreLoc(Ptr, AccessSize);
1002
1003 for (BasicBlock *B : L->blocks())
1004 for (Instruction &I : *B)
1005 if (!IgnoredInsts.contains(Ptr: &I) &&
1006 isModOrRefSet(MRI: AA.getModRefInfo(I: &I, OptLoc: StoreLoc) & Access))
1007 return true;
1008 return false;
1009}
1010
1011// If we have a negative stride, Start refers to the end of the memory location
1012// we're trying to memset. Therefore, we need to recompute the base pointer,
1013// which is just Start - BECount*Size.
1014static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
1015 Type *IntPtr, const SCEV *StoreSizeSCEV,
1016 ScalarEvolution *SE) {
1017 const SCEV *Index = SE->getTruncateOrZeroExtend(V: BECount, Ty: IntPtr);
1018 if (!StoreSizeSCEV->isOne()) {
1019 // index = back edge count * store size
1020 Index = SE->getMulExpr(LHS: Index,
1021 RHS: SE->getTruncateOrZeroExtend(V: StoreSizeSCEV, Ty: IntPtr),
1022 Flags: SCEV::FlagNUW);
1023 }
1024 // base pointer = start - index * store size
1025 return SE->getMinusSCEV(LHS: Start, RHS: Index);
1026}
1027
1028/// Compute the number of bytes as a SCEV from the backedge taken count.
1029///
1030/// This also maps the SCEV into the provided type and tries to handle the
1031/// computation in a way that will fold cleanly.
1032static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr,
1033 const SCEV *StoreSizeSCEV, Loop *CurLoop,
1034 const DataLayout *DL, ScalarEvolution *SE) {
1035 const SCEV *TripCountSCEV =
1036 SE->getTripCountFromExitCount(ExitCount: BECount, EvalTy: IntPtr, L: CurLoop);
1037 return SE->getMulExpr(LHS: TripCountSCEV,
1038 RHS: SE->getTruncateOrZeroExtend(V: StoreSizeSCEV, Ty: IntPtr),
1039 Flags: SCEV::FlagNUW);
1040}
1041
1042/// processLoopStridedStore - We see a strided store of some value. If we can
1043/// transform this into a memset or memset_pattern in the loop preheader, do so.
1044bool LoopIdiomRecognize::processLoopStridedStore(
1045 Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment,
1046 Value *StoredVal, Instruction *TheStore,
1047 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
1048 const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) {
1049 Module *M = TheStore->getModule();
1050
1051 // The trip count of the loop and the base pointer of the addrec SCEV is
1052 // guaranteed to be loop invariant, which means that it should dominate the
1053 // header. This allows us to insert code for it in the preheader.
1054 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
1055 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1056 IRBuilder<> Builder(Preheader->getTerminator());
1057 SCEVExpander Expander(*SE, "loop-idiom");
1058 SCEVExpanderCleaner ExpCleaner(Expander);
1059
1060 Type *DestInt8PtrTy = Builder.getPtrTy(AddrSpace: DestAS);
1061 Type *IntIdxTy = DL->getIndexType(PtrTy: DestPtr->getType());
1062
1063 bool Changed = false;
1064 const SCEV *Start = Ev->getStart();
1065 // Handle negative strided loops.
1066 if (IsNegStride)
1067 Start = getStartForNegStride(Start, BECount, IntPtr: IntIdxTy, StoreSizeSCEV, SE);
1068
1069 // TODO: ideally we should still be able to generate memset if SCEV expander
1070 // is taught to generate the dependencies at the latest point.
1071 if (!Expander.isSafeToExpand(S: Start))
1072 return Changed;
1073
1074 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
1075 // this into a memset in the loop preheader now if we want. However, this
1076 // would be unsafe to do if there is anything else in the loop that may read
1077 // or write to the aliased location. Check for any overlap by generating the
1078 // base pointer and checking the region.
1079 Value *BasePtr =
1080 Expander.expandCodeFor(SH: Start, Ty: DestInt8PtrTy, I: Preheader->getTerminator());
1081
1082 // From here on out, conservatively report to the pass manager that we've
1083 // changed the IR, even if we later clean up these added instructions. There
1084 // may be structural differences e.g. in the order of use lists not accounted
1085 // for in just a textual dump of the IR. This is written as a variable, even
1086 // though statically all the places this dominates could be replaced with
1087 // 'true', with the hope that anyone trying to be clever / "more precise" with
1088 // the return value will read this comment, and leave them alone.
1089 Changed = true;
1090
1091 if (mayLoopAccessLocation(Ptr: BasePtr, Access: ModRefInfo::ModRef, L: CurLoop, BECount,
1092 StoreSizeSCEV, AA&: *AA, IgnoredInsts&: Stores))
1093 return Changed;
1094
1095 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
1096 return Changed;
1097
1098 // Okay, everything looks good, insert the memset.
1099 Value *SplatValue = isBytewiseValue(V: StoredVal, DL: *DL);
1100 Constant *PatternValue = nullptr;
1101 if (!SplatValue)
1102 PatternValue = getMemSetPatternValue(V: StoredVal, DL);
1103
1104 // MemsetArg is the number of bytes for the memset libcall, and the number
1105 // of pattern repetitions if the memset.pattern intrinsic is being used.
1106 Value *MemsetArg;
1107 std::optional<int64_t> BytesWritten;
1108
1109 if (PatternValue && (HasMemsetPattern || ForceMemsetPatternIntrinsic)) {
1110 const SCEV *TripCountS =
1111 SE->getTripCountFromExitCount(ExitCount: BECount, EvalTy: IntIdxTy, L: CurLoop);
1112 if (!Expander.isSafeToExpand(S: TripCountS))
1113 return Changed;
1114 const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(Val: StoreSizeSCEV);
1115 if (!ConstStoreSize)
1116 return Changed;
1117 Value *TripCount = Expander.expandCodeFor(SH: TripCountS, Ty: IntIdxTy,
1118 I: Preheader->getTerminator());
1119 uint64_t PatternRepsPerTrip =
1120 (ConstStoreSize->getValue()->getZExtValue() * 8) /
1121 DL->getTypeSizeInBits(Ty: PatternValue->getType());
1122 // If ConstStoreSize is not equal to the width of PatternValue, then
1123 // MemsetArg is TripCount * (ConstStoreSize/PatternValueWidth). Else
1124 // MemSetArg is just TripCount.
1125 MemsetArg =
1126 PatternRepsPerTrip == 1
1127 ? TripCount
1128 : Builder.CreateMul(LHS: TripCount,
1129 RHS: Builder.getIntN(N: IntIdxTy->getIntegerBitWidth(),
1130 C: PatternRepsPerTrip));
1131 if (auto *CI = dyn_cast<ConstantInt>(Val: TripCount))
1132 BytesWritten =
1133 CI->getZExtValue() * ConstStoreSize->getValue()->getZExtValue();
1134
1135 } else {
1136 const SCEV *NumBytesS =
1137 getNumBytes(BECount, IntPtr: IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1138
1139 // TODO: ideally we should still be able to generate memset if SCEV expander
1140 // is taught to generate the dependencies at the latest point.
1141 if (!Expander.isSafeToExpand(S: NumBytesS))
1142 return Changed;
1143 MemsetArg =
1144 Expander.expandCodeFor(SH: NumBytesS, Ty: IntIdxTy, I: Preheader->getTerminator());
1145 if (auto *CI = dyn_cast<ConstantInt>(Val: MemsetArg))
1146 BytesWritten = CI->getZExtValue();
1147 }
1148 assert(MemsetArg && "MemsetArg should have been set");
1149
1150 AAMDNodes AATags = TheStore->getAAMetadata();
1151 for (Instruction *Store : Stores)
1152 AATags = AATags.merge(Other: Store->getAAMetadata());
1153 if (BytesWritten)
1154 AATags = AATags.extendTo(Len: BytesWritten.value());
1155 else
1156 AATags = AATags.extendTo(Len: -1);
1157
1158 CallInst *NewCall;
1159 if (SplatValue) {
1160 NewCall = Builder.CreateMemSet(Ptr: BasePtr, Val: SplatValue, Size: MemsetArg,
1161 Align: MaybeAlign(StoreAlignment),
1162 /*isVolatile=*/false, AAInfo: AATags);
1163 } else if (ForceMemsetPatternIntrinsic ||
1164 isLibFuncEmittable(M, TLI, TheLibFunc: LibFunc_memset_pattern16)) {
1165 assert(isa<SCEVConstant>(StoreSizeSCEV) && "Expected constant store size");
1166
1167 NewCall = Builder.CreateIntrinsic(
1168 ID: Intrinsic::experimental_memset_pattern,
1169 Types: {DestInt8PtrTy, PatternValue->getType(), IntIdxTy},
1170 Args: {BasePtr, PatternValue, MemsetArg,
1171 ConstantInt::getFalse(Context&: M->getContext())});
1172 if (StoreAlignment)
1173 cast<MemSetPatternInst>(Val: NewCall)->setDestAlignment(*StoreAlignment);
1174 NewCall->setAAMetadata(AATags);
1175 } else {
1176 // Neither a memset, nor memset_pattern16
1177 return Changed;
1178 }
1179
1180 NewCall->setDebugLoc(TheStore->getDebugLoc());
1181
1182 if (MSSAU) {
1183 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1184 I: NewCall, Definition: nullptr, BB: NewCall->getParent(), Point: MemorySSA::BeforeTerminator);
1185 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewMemAcc), RenameUses: true);
1186 }
1187
1188 LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1189 << " from store to: " << *Ev << " at: " << *TheStore
1190 << "\n");
1191
1192 ORE.emit(RemarkBuilder: [&]() {
1193 OptimizationRemark R(DEBUG_TYPE, "ProcessLoopStridedStore",
1194 NewCall->getDebugLoc(), Preheader);
1195 R << "Transformed loop-strided store in "
1196 << ore::NV("Function", TheStore->getFunction())
1197 << " function into a call to "
1198 << ore::NV("NewFunction", NewCall->getCalledFunction())
1199 << "() intrinsic";
1200 if (!Stores.empty())
1201 R << ore::setExtraArgs();
1202 for (auto *I : Stores) {
1203 R << ore::NV("FromBlock", I->getParent()->getName())
1204 << ore::NV("ToBlock", Preheader->getName());
1205 }
1206 return R;
1207 });
1208
1209 // Okay, the memset has been formed. Zap the original store and anything that
1210 // feeds into it.
1211 for (auto *I : Stores) {
1212 if (MSSAU)
1213 MSSAU->removeMemoryAccess(I, OptimizePhis: true);
1214 deleteDeadInstruction(I);
1215 }
1216 if (MSSAU && VerifyMemorySSA)
1217 MSSAU->getMemorySSA()->verifyMemorySSA();
1218 ++NumMemSet;
1219 ExpCleaner.markResultUsed();
1220 return true;
1221}
1222
1223/// If the stored value is a strided load in the same loop with the same stride
1224/// this may be transformable into a memcpy. This kicks in for stuff like
1225/// for (i) A[i] = B[i];
1226bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
1227 const SCEV *BECount) {
1228 assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.");
1229
1230 Value *StorePtr = SI->getPointerOperand();
1231 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: StorePtr));
1232 unsigned StoreSize = DL->getTypeStoreSize(Ty: SI->getValueOperand()->getType());
1233
1234 // The store must be feeding a non-volatile load.
1235 LoadInst *LI = cast<LoadInst>(Val: SI->getValueOperand());
1236 assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.");
1237
1238 // See if the pointer expression is an AddRec like {base,+,1} on the current
1239 // loop, which indicates a strided load. If we have something else, it's a
1240 // random load we can't handle.
1241 Value *LoadPtr = LI->getPointerOperand();
1242 const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(Val: SE->getSCEV(V: LoadPtr));
1243
1244 const SCEV *StoreSizeSCEV = SE->getConstant(Ty: StorePtr->getType(), V: StoreSize);
1245 return processLoopStoreOfLoopLoad(DestPtr: StorePtr, SourcePtr: LoadPtr, StoreSize: StoreSizeSCEV,
1246 StoreAlign: SI->getAlign(), LoadAlign: LI->getAlign(), TheStore: SI, TheLoad: LI,
1247 StoreEv, LoadEv, BECount);
1248}
1249
1250namespace {
1251class MemmoveVerifier {
1252public:
1253 explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr,
1254 const DataLayout &DL)
1255 : DL(DL), BP1(llvm::GetPointerBaseWithConstantOffset(
1256 Ptr: LoadBasePtr.stripPointerCasts(), Offset&: LoadOff, DL)),
1257 BP2(llvm::GetPointerBaseWithConstantOffset(
1258 Ptr: StoreBasePtr.stripPointerCasts(), Offset&: StoreOff, DL)),
1259 IsSameObject(BP1 == BP2) {}
1260
1261 bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride,
1262 const Instruction &TheLoad,
1263 bool IsMemCpy) const {
1264 if (IsMemCpy) {
1265 // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1266 // for negative stride.
1267 if ((!IsNegStride && LoadOff <= StoreOff) ||
1268 (IsNegStride && LoadOff >= StoreOff))
1269 return false;
1270 } else {
1271 // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1272 // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr.
1273 int64_t LoadSize =
1274 DL.getTypeSizeInBits(Ty: TheLoad.getType()).getFixedValue() / 8;
1275 if (BP1 != BP2 || LoadSize != int64_t(StoreSize))
1276 return false;
1277 if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) ||
1278 (IsNegStride && LoadOff + LoadSize > StoreOff))
1279 return false;
1280 }
1281 return true;
1282 }
1283
1284private:
1285 const DataLayout &DL;
1286 int64_t LoadOff = 0;
1287 int64_t StoreOff = 0;
1288 const Value *BP1;
1289 const Value *BP2;
1290
1291public:
1292 const bool IsSameObject;
1293};
1294} // namespace
1295
1296bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
1297 Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV,
1298 MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore,
1299 Instruction *TheLoad, const SCEVAddRecExpr *StoreEv,
1300 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
1301
1302 // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to
1303 // conservatively bail here, since otherwise we may have to transform
1304 // llvm.memcpy.inline into llvm.memcpy which is illegal.
1305 if (auto *MCI = dyn_cast<MemCpyInst>(Val: TheStore); MCI && MCI->isForceInlined())
1306 return false;
1307
1308 // The trip count of the loop and the base pointer of the addrec SCEV is
1309 // guaranteed to be loop invariant, which means that it should dominate the
1310 // header. This allows us to insert code for it in the preheader.
1311 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1312 IRBuilder<> Builder(Preheader->getTerminator());
1313 SCEVExpander Expander(*SE, "loop-idiom");
1314
1315 SCEVExpanderCleaner ExpCleaner(Expander);
1316
1317 bool Changed = false;
1318 const SCEV *StrStart = StoreEv->getStart();
1319 unsigned StrAS = DestPtr->getType()->getPointerAddressSpace();
1320 Type *IntIdxTy = Builder.getIntNTy(N: DL->getIndexSizeInBits(AS: StrAS));
1321
1322 APInt Stride = getStoreStride(StoreEv);
1323 const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(Val: StoreSizeSCEV);
1324
1325 // TODO: Deal with non-constant size; Currently expect constant store size
1326 assert(ConstStoreSize && "store size is expected to be a constant");
1327
1328 int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue();
1329 bool IsNegStride = StoreSize == -Stride;
1330
1331 // Handle negative strided loops.
1332 if (IsNegStride)
1333 StrStart =
1334 getStartForNegStride(Start: StrStart, BECount, IntPtr: IntIdxTy, StoreSizeSCEV, SE);
1335
1336 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1337 // this into a memcpy in the loop preheader now if we want. However, this
1338 // would be unsafe to do if there is anything else in the loop that may read
1339 // or write the memory region we're storing to. This includes the load that
1340 // feeds the stores. Check for an alias by generating the base address and
1341 // checking everything.
1342 Value *StoreBasePtr = Expander.expandCodeFor(
1343 SH: StrStart, Ty: Builder.getPtrTy(AddrSpace: StrAS), I: Preheader->getTerminator());
1344
1345 // From here on out, conservatively report to the pass manager that we've
1346 // changed the IR, even if we later clean up these added instructions. There
1347 // may be structural differences e.g. in the order of use lists not accounted
1348 // for in just a textual dump of the IR. This is written as a variable, even
1349 // though statically all the places this dominates could be replaced with
1350 // 'true', with the hope that anyone trying to be clever / "more precise" with
1351 // the return value will read this comment, and leave them alone.
1352 Changed = true;
1353
1354 SmallPtrSet<Instruction *, 2> IgnoredInsts;
1355 IgnoredInsts.insert(Ptr: TheStore);
1356
1357 bool IsMemCpy = isa<MemCpyInst>(Val: TheStore);
1358 const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store";
1359
1360 bool LoopAccessStore =
1361 mayLoopAccessLocation(Ptr: StoreBasePtr, Access: ModRefInfo::ModRef, L: CurLoop, BECount,
1362 StoreSizeSCEV, AA&: *AA, IgnoredInsts);
1363 if (LoopAccessStore) {
1364 // For memmove case it's not enough to guarantee that loop doesn't access
1365 // TheStore and TheLoad. Additionally we need to make sure that TheStore is
1366 // the only user of TheLoad.
1367 if (!TheLoad->hasOneUse())
1368 return Changed;
1369 IgnoredInsts.insert(Ptr: TheLoad);
1370 if (mayLoopAccessLocation(Ptr: StoreBasePtr, Access: ModRefInfo::ModRef, L: CurLoop,
1371 BECount, StoreSizeSCEV, AA&: *AA, IgnoredInsts)) {
1372 ORE.emit(RemarkBuilder: [&]() {
1373 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessStore",
1374 TheStore)
1375 << ore::NV("Inst", InstRemark) << " in "
1376 << ore::NV("Function", TheStore->getFunction())
1377 << " function will not be hoisted: "
1378 << ore::NV("Reason", "The loop may access store location");
1379 });
1380 return Changed;
1381 }
1382 IgnoredInsts.erase(Ptr: TheLoad);
1383 }
1384
1385 const SCEV *LdStart = LoadEv->getStart();
1386 unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace();
1387
1388 // Handle negative strided loops.
1389 if (IsNegStride)
1390 LdStart =
1391 getStartForNegStride(Start: LdStart, BECount, IntPtr: IntIdxTy, StoreSizeSCEV, SE);
1392
1393 // For a memcpy, we have to make sure that the input array is not being
1394 // mutated by the loop.
1395 Value *LoadBasePtr = Expander.expandCodeFor(SH: LdStart, Ty: Builder.getPtrTy(AddrSpace: LdAS),
1396 I: Preheader->getTerminator());
1397
1398 // If the store is a memcpy instruction, we must check if it will write to
1399 // the load memory locations. So remove it from the ignored stores.
1400 MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL);
1401 if (IsMemCpy && !Verifier.IsSameObject)
1402 IgnoredInsts.erase(Ptr: TheStore);
1403 if (mayLoopAccessLocation(Ptr: LoadBasePtr, Access: ModRefInfo::Mod, L: CurLoop, BECount,
1404 StoreSizeSCEV, AA&: *AA, IgnoredInsts)) {
1405 ORE.emit(RemarkBuilder: [&]() {
1406 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessLoad", TheLoad)
1407 << ore::NV("Inst", InstRemark) << " in "
1408 << ore::NV("Function", TheStore->getFunction())
1409 << " function will not be hoisted: "
1410 << ore::NV("Reason", "The loop may access load location");
1411 });
1412 return Changed;
1413 }
1414
1415 bool IsAtomic = TheStore->isAtomic() || TheLoad->isAtomic();
1416 bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore;
1417
1418 if (IsAtomic) {
1419 // For now don't support unordered atomic memmove.
1420 if (UseMemMove)
1421 return Changed;
1422
1423 // We cannot allow unaligned ops for unordered load/store, so reject
1424 // anything where the alignment isn't at least the element size.
1425 assert((StoreAlign && LoadAlign) &&
1426 "Expect unordered load/store to have align.");
1427 if (*StoreAlign < StoreSize || *LoadAlign < StoreSize)
1428 return Changed;
1429
1430 // If the element.atomic memcpy is not lowered into explicit
1431 // loads/stores later, then it will be lowered into an element-size
1432 // specific lib call. If the lib call doesn't exist for our store size, then
1433 // we shouldn't generate the memcpy.
1434 if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize())
1435 return Changed;
1436 }
1437
1438 if (UseMemMove)
1439 if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, TheLoad: *TheLoad,
1440 IsMemCpy))
1441 return Changed;
1442
1443 if (avoidLIRForMultiBlockLoop())
1444 return Changed;
1445
1446 // Okay, everything is safe, we can transform this!
1447
1448 const SCEV *NumBytesS =
1449 getNumBytes(BECount, IntPtr: IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1450
1451 Value *NumBytes =
1452 Expander.expandCodeFor(SH: NumBytesS, Ty: IntIdxTy, I: Preheader->getTerminator());
1453
1454 AAMDNodes AATags = TheLoad->getAAMetadata();
1455 AAMDNodes StoreAATags = TheStore->getAAMetadata();
1456 AATags = AATags.merge(Other: StoreAATags);
1457 if (auto CI = dyn_cast<ConstantInt>(Val: NumBytes))
1458 AATags = AATags.extendTo(Len: CI->getZExtValue());
1459 else
1460 AATags = AATags.extendTo(Len: -1);
1461
1462 CallInst *NewCall = nullptr;
1463 // Check whether to generate an unordered atomic memcpy:
1464 // If the load or store are atomic, then they must necessarily be unordered
1465 // by previous checks.
1466 if (!IsAtomic) {
1467 if (UseMemMove)
1468 NewCall = Builder.CreateMemMove(Dst: StoreBasePtr, DstAlign: StoreAlign, Src: LoadBasePtr,
1469 SrcAlign: LoadAlign, Size: NumBytes,
1470 /*isVolatile=*/false, AAInfo: AATags);
1471 else
1472 NewCall =
1473 Builder.CreateMemCpy(Dst: StoreBasePtr, DstAlign: StoreAlign, Src: LoadBasePtr, SrcAlign: LoadAlign,
1474 Size: NumBytes, /*isVolatile=*/false, AAInfo: AATags);
1475 } else {
1476 // Create the call.
1477 // Note that unordered atomic loads/stores are *required* by the spec to
1478 // have an alignment but non-atomic loads/stores may not.
1479 NewCall = Builder.CreateElementUnorderedAtomicMemCpy(
1480 Dst: StoreBasePtr, DstAlign: *StoreAlign, Src: LoadBasePtr, SrcAlign: *LoadAlign, Size: NumBytes, ElementSize: StoreSize,
1481 AAInfo: AATags);
1482 }
1483 NewCall->setDebugLoc(TheStore->getDebugLoc());
1484
1485 if (MSSAU) {
1486 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1487 I: NewCall, Definition: nullptr, BB: NewCall->getParent(), Point: MemorySSA::BeforeTerminator);
1488 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewMemAcc), RenameUses: true);
1489 }
1490
1491 LLVM_DEBUG(dbgs() << " Formed new call: " << *NewCall << "\n"
1492 << " from load ptr=" << *LoadEv << " at: " << *TheLoad
1493 << "\n"
1494 << " from store ptr=" << *StoreEv << " at: " << *TheStore
1495 << "\n");
1496
1497 ORE.emit(RemarkBuilder: [&]() {
1498 return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStoreOfLoopLoad",
1499 NewCall->getDebugLoc(), Preheader)
1500 << "Formed a call to "
1501 << ore::NV("NewFunction", NewCall->getCalledFunction())
1502 << "() intrinsic from " << ore::NV("Inst", InstRemark)
1503 << " instruction in " << ore::NV("Function", TheStore->getFunction())
1504 << " function"
1505 << ore::setExtraArgs()
1506 << ore::NV("FromBlock", TheStore->getParent()->getName())
1507 << ore::NV("ToBlock", Preheader->getName());
1508 });
1509
1510 // Okay, a new call to memcpy/memmove has been formed. Zap the original store
1511 // and anything that feeds into it.
1512 if (MSSAU)
1513 MSSAU->removeMemoryAccess(I: TheStore, OptimizePhis: true);
1514 deleteDeadInstruction(I: TheStore);
1515 if (MSSAU && VerifyMemorySSA)
1516 MSSAU->getMemorySSA()->verifyMemorySSA();
1517 if (UseMemMove)
1518 ++NumMemMove;
1519 else
1520 ++NumMemCpy;
1521 ExpCleaner.markResultUsed();
1522 return true;
1523}
1524
1525// When compiling for codesize we avoid idiom recognition for a multi-block loop
1526// unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
1527//
1528bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
1529 bool IsLoopMemset) {
1530 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
1531 if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) {
1532 LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
1533 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1534 << " avoided: multi-block top-level loop\n");
1535 return true;
1536 }
1537 }
1538
1539 return false;
1540}
1541
1542bool LoopIdiomRecognize::optimizeCRCLoop(const PolynomialInfo &Info) {
1543 // FIXME: Hexagon has a special HexagonLoopIdiom that optimizes CRC using
1544 // carry-less multiplication instructions, which is more efficient than our
1545 // Sarwate table-lookup optimization. Hence, until we're able to emit
1546 // target-specific instructions for Hexagon, subsuming HexagonLoopIdiom,
1547 // disable the optimization for Hexagon.
1548 Module &M = *CurLoop->getHeader()->getModule();
1549 Triple TT(M.getTargetTriple());
1550 if (TT.getArch() == Triple::hexagon)
1551 return false;
1552
1553 // First, create a new GlobalVariable corresponding to the
1554 // Sarwate-lookup-table.
1555 Type *CRCTy = Info.LHS->getType();
1556 unsigned CRCBW = CRCTy->getIntegerBitWidth();
1557 std::array<Constant *, 256> CRCConstants;
1558 transform(Range: HashRecognize::genSarwateTable(GenPoly: Info.RHS, ByteOrderSwapped: Info.ByteOrderSwapped),
1559 d_first: CRCConstants.begin(),
1560 F: [CRCTy](const APInt &E) { return ConstantInt::get(Ty: CRCTy, V: E); });
1561 Constant *ConstArray =
1562 ConstantArray::get(T: ArrayType::get(ElementType: CRCTy, NumElements: 256), V: CRCConstants);
1563 GlobalVariable *GV =
1564 new GlobalVariable(M, ConstArray->getType(), true,
1565 GlobalValue::PrivateLinkage, ConstArray, ".crctable");
1566
1567 PHINode *IV = CurLoop->getCanonicalInductionVariable();
1568 SmallVector<PHINode *, 2> Cleanup;
1569
1570 // Next, mark all PHIs for removal except IV.
1571 {
1572 for (PHINode &PN : CurLoop->getHeader()->phis()) {
1573 if (&PN == IV)
1574 continue;
1575 PN.replaceAllUsesWith(V: PoisonValue::get(T: PN.getType()));
1576 Cleanup.push_back(Elt: &PN);
1577 }
1578 }
1579
1580 // Next, fix up the trip count.
1581 {
1582 unsigned NewBTC = (Info.TripCount / 8) - 1;
1583 BasicBlock *LoopBlk = CurLoop->getLoopLatch();
1584 CondBrInst *BrInst = cast<CondBrInst>(Val: LoopBlk->getTerminator());
1585 CmpPredicate ExitPred = BrInst->getSuccessor(i: 0) == LoopBlk
1586 ? ICmpInst::Predicate::ICMP_NE
1587 : ICmpInst::Predicate::ICMP_EQ;
1588 Instruction *ExitCond = CurLoop->getLatchCmpInst();
1589 Value *ExitLimit = ConstantInt::get(Ty: IV->getType(), V: NewBTC);
1590 IRBuilder<> Builder(ExitCond);
1591 Value *NewExitCond =
1592 Builder.CreateICmp(P: ExitPred, LHS: IV, RHS: ExitLimit, Name: "exit.cond");
1593 ExitCond->replaceAllUsesWith(V: NewExitCond);
1594 deleteDeadInstruction(I: ExitCond);
1595 }
1596
1597 // Finally, fill the loop with the Sarwate-table-lookup logic, and replace all
1598 // uses of ComputedValue.
1599 //
1600 // Little-endian:
1601 // crc = (crc >> 8) ^ tbl[(iv'th byte of data) ^ (bottom byte of crc)]
1602 // Big-Endian:
1603 // crc = (crc << 8) ^ tbl[(iv'th byte of data) ^ (top byte of crc)]
1604 {
1605 auto LoByte = [](IRBuilderBase &Builder, Value *Op, const Twine &Name) {
1606 return Builder.CreateZExtOrTrunc(
1607 V: Op, DestTy: IntegerType::getInt8Ty(C&: Op->getContext()), Name);
1608 };
1609 auto HiIdx = [LoByte, CRCBW](IRBuilderBase &Builder, Value *Op,
1610 const Twine &Name) {
1611 Type *OpTy = Op->getType();
1612
1613 // When the bitwidth of the CRC mismatches the Op's bitwidth, we need to
1614 // use the CRC's bitwidth as the reference for shifting right.
1615 return LoByte(Builder,
1616 CRCBW > 8 ? Builder.CreateLShr(
1617 LHS: Op, RHS: ConstantInt::get(Ty: OpTy, V: CRCBW - 8), Name)
1618 : Op,
1619 Name + ".lo.byte");
1620 };
1621
1622 IRBuilder<> Builder(CurLoop->getHeader(),
1623 CurLoop->getHeader()->getFirstNonPHIIt());
1624
1625 // Create the CRC PHI, and initialize its incoming value to the initial
1626 // value of CRC.
1627 PHINode *CRCPhi = Builder.CreatePHI(Ty: CRCTy, NumReservedValues: 2, Name: "crc");
1628 CRCPhi->addIncoming(V: Info.LHS, BB: CurLoop->getLoopPreheader());
1629
1630 // CRC is now an evolving variable, initialized to the PHI.
1631 Value *CRC = CRCPhi;
1632
1633 // TableIndexer = ((top|bottom) byte of CRC). It is XOR'ed with (iv'th byte
1634 // of LHSAux), if LHSAux is non-nullptr.
1635 Value *Indexer = CRC;
1636 if (Value *Data = Info.LHSAux) {
1637 Type *DataTy = Data->getType();
1638
1639 // To index into the (iv'th byte of LHSAux), we multiply iv by 8, and we
1640 // shift right by that amount, and take the lo-byte (in the little-endian
1641 // case), or shift left by that amount, and take the hi-idx (in the
1642 // big-endian case).
1643 Value *IVBits = Builder.CreateZExtOrTrunc(
1644 V: Builder.CreateShl(LHS: IV, RHS: 3, Name: "iv.bits"), DestTy: DataTy, Name: "iv.indexer");
1645 Value *DataIndexer =
1646 Info.ByteOrderSwapped
1647 ? Builder.CreateShl(LHS: Data, RHS: IVBits, Name: "data.indexer")
1648 : Builder.CreateLShr(LHS: Data, RHS: IVBits, Name: "data.indexer");
1649 Indexer = Builder.CreateXor(
1650 LHS: DataIndexer,
1651 RHS: Builder.CreateZExtOrTrunc(V: Indexer, DestTy: DataTy, Name: "crc.indexer.cast"),
1652 Name: "crc.data.indexer");
1653 }
1654
1655 Indexer = Info.ByteOrderSwapped ? HiIdx(Builder, Indexer, "indexer.hi")
1656 : LoByte(Builder, Indexer, "indexer.lo");
1657
1658 // Always index into a GEP using the index type.
1659 Indexer = Builder.CreateZExt(
1660 V: Indexer, DestTy: SE->getDataLayout().getIndexType(PtrTy: GV->getType()),
1661 Name: "indexer.ext");
1662
1663 // CRCTableLd = CRCTable[(iv'th byte of data) ^ (top|bottom) byte of CRC].
1664 Value *CRCTableGEP =
1665 Builder.CreateInBoundsGEP(Ty: CRCTy, Ptr: GV, IdxList: Indexer, Name: "tbl.ptradd");
1666 Value *CRCTableLd = Builder.CreateLoad(Ty: CRCTy, Ptr: CRCTableGEP, Name: "tbl.ld");
1667
1668 // CRCNext = (CRC (<<|>>) 8) ^ CRCTableLd, or simply CRCTableLd in case of
1669 // CRC-8.
1670 Value *CRCNext = CRCTableLd;
1671 if (CRCBW > 8) {
1672 Value *CRCShift = Info.ByteOrderSwapped
1673 ? Builder.CreateShl(LHS: CRC, RHS: 8, Name: "crc.be.shift")
1674 : Builder.CreateLShr(LHS: CRC, RHS: 8, Name: "crc.le.shift");
1675 CRCNext = Builder.CreateXor(LHS: CRCShift, RHS: CRCTableLd, Name: "crc.next");
1676 }
1677
1678 // Connect the back-edge for the loop, and RAUW the ComputedValue.
1679 CRCPhi->addIncoming(V: CRCNext, BB: CurLoop->getLoopLatch());
1680 Info.ComputedValue->replaceUsesOutsideBlock(V: CRCNext,
1681 BB: CurLoop->getLoopLatch());
1682 }
1683
1684 // Cleanup.
1685 {
1686 for (PHINode *PN : Cleanup)
1687 RecursivelyDeleteDeadPHINode(PN);
1688 SE->forgetLoop(L: CurLoop);
1689 }
1690 return true;
1691}
1692
1693bool LoopIdiomRecognize::runOnNoncountableLoop() {
1694 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
1695 << CurLoop->getHeader()->getParent()->getName()
1696 << "] Noncountable Loop %"
1697 << CurLoop->getHeader()->getName() << "\n");
1698
1699 return recognizePopcount() || recognizeAndInsertFFS() ||
1700 recognizeShiftUntilBitTest() || recognizeShiftUntilZero() ||
1701 recognizeShiftUntilLessThan() || recognizeAndInsertStrLen();
1702}
1703
1704/// Check if the given conditional branch is based on the comparison between
1705/// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is
1706/// true), the control yields to the loop entry. If the branch matches the
1707/// behavior, the variable involved in the comparison is returned. This function
1708/// will be called to see if the precondition and postcondition of the loop are
1709/// in desirable form.
1710static Value *matchCondition(CondBrInst *BI, BasicBlock *LoopEntry,
1711 bool JmpOnZero = false) {
1712 ICmpInst *Cond = dyn_cast<ICmpInst>(Val: BI->getCondition());
1713 if (!Cond)
1714 return nullptr;
1715
1716 auto *CmpZero = dyn_cast<ConstantInt>(Val: Cond->getOperand(i_nocapture: 1));
1717 if (!CmpZero || !CmpZero->isZero())
1718 return nullptr;
1719
1720 BasicBlock *TrueSucc = BI->getSuccessor(i: 0);
1721 BasicBlock *FalseSucc = BI->getSuccessor(i: 1);
1722 if (JmpOnZero)
1723 std::swap(a&: TrueSucc, b&: FalseSucc);
1724
1725 ICmpInst::Predicate Pred = Cond->getPredicate();
1726 if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) ||
1727 (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry))
1728 return Cond->getOperand(i_nocapture: 0);
1729
1730 return nullptr;
1731}
1732
1733namespace {
1734
1735class StrlenVerifier {
1736public:
1737 explicit StrlenVerifier(const Loop *CurLoop, ScalarEvolution *SE,
1738 const TargetLibraryInfo *TLI)
1739 : CurLoop(CurLoop), SE(SE), TLI(TLI) {}
1740
1741 bool isValidStrlenIdiom() {
1742 // Give up if the loop has multiple blocks, multiple backedges, or
1743 // multiple exit blocks
1744 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1 ||
1745 !CurLoop->getUniqueExitBlock())
1746 return false;
1747
1748 // It should have a preheader and a branch instruction.
1749 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1750 if (!Preheader ||
1751 !isa<UncondBrInst, CondBrInst>(Val: Preheader->getTerminator()))
1752 return false;
1753
1754 // The loop exit must be conditioned on an icmp with 0 the null terminator.
1755 // The icmp operand has to be a load on some SSA reg that increments
1756 // by 1 in the loop.
1757 BasicBlock *LoopBody = *CurLoop->block_begin();
1758
1759 // Skip if the body is too big as it most likely is not a strlen idiom.
1760 if (!LoopBody || LoopBody->size() >= 15)
1761 return false;
1762
1763 CondBrInst *LoopTerm = dyn_cast<CondBrInst>(Val: LoopBody->getTerminator());
1764 if (!LoopTerm)
1765 return false;
1766 Value *LoopCond = matchCondition(BI: LoopTerm, LoopEntry: LoopBody);
1767 if (!LoopCond)
1768 return false;
1769
1770 LoadInst *LoopLoad = dyn_cast<LoadInst>(Val: LoopCond);
1771 if (!LoopLoad || LoopLoad->getPointerAddressSpace() != 0)
1772 return false;
1773
1774 OperandType = LoopLoad->getType();
1775 if (!OperandType || !OperandType->isIntegerTy())
1776 return false;
1777
1778 // See if the pointer expression is an AddRec with constant step a of form
1779 // ({n,+,a}) where a is the width of the char type.
1780 Value *IncPtr = LoopLoad->getPointerOperand();
1781 const SCEV *LoadEv = SE->getSCEV(V: IncPtr);
1782 const APInt *Step;
1783 if (!match(S: LoadEv,
1784 P: m_scev_AffineAddRec(Op0: m_SCEV(V&: LoadBaseEv), Op1: m_scev_APInt(C&: Step))))
1785 return false;
1786
1787 LLVM_DEBUG(dbgs() << "pointer load scev: " << *LoadEv << "\n");
1788
1789 unsigned StepSize = Step->getZExtValue();
1790
1791 // Verify that StepSize is consistent with platform char width.
1792 OpWidth = OperandType->getIntegerBitWidth();
1793 unsigned WcharSize = TLI->getWCharSize(M: *LoopLoad->getModule());
1794 if (OpWidth != StepSize * 8)
1795 return false;
1796 if (OpWidth != 8 && OpWidth != 16 && OpWidth != 32)
1797 return false;
1798 if (OpWidth >= 16)
1799 if (OpWidth != WcharSize * 8)
1800 return false;
1801
1802 // Scan every instruction in the loop to ensure there are no side effects.
1803 for (Instruction &I : *LoopBody)
1804 if (I.mayHaveSideEffects())
1805 return false;
1806
1807 BasicBlock *LoopExitBB = CurLoop->getExitBlock();
1808 if (!LoopExitBB)
1809 return false;
1810
1811 for (PHINode &PN : LoopExitBB->phis()) {
1812 if (!SE->isSCEVable(Ty: PN.getType()))
1813 return false;
1814
1815 const SCEV *Ev = SE->getSCEV(V: &PN);
1816 if (!Ev)
1817 return false;
1818
1819 LLVM_DEBUG(dbgs() << "loop exit phi scev: " << *Ev << "\n");
1820
1821 // Since we verified that the loop trip count will be a valid strlen
1822 // idiom, we can expand all lcssa phi with {n,+,1} as (n + strlen) and use
1823 // SCEVExpander materialize the loop output.
1824 const SCEVAddRecExpr *AddRecEv = dyn_cast<SCEVAddRecExpr>(Val: Ev);
1825 if (!AddRecEv || !AddRecEv->isAffine())
1826 return false;
1827
1828 // We only want RecAddExpr with recurrence step that is constant. This
1829 // is good enough for all the idioms we want to recognize. Later we expand
1830 // and materialize the recurrence as {base,+,a} -> (base + a * strlen)
1831 if (!isa<SCEVConstant>(Val: AddRecEv->getStepRecurrence(SE&: *SE)))
1832 return false;
1833 }
1834
1835 return true;
1836 }
1837
1838public:
1839 const Loop *CurLoop;
1840 ScalarEvolution *SE;
1841 const TargetLibraryInfo *TLI;
1842
1843 unsigned OpWidth;
1844 ConstantInt *StepSizeCI;
1845 const SCEV *LoadBaseEv;
1846 Type *OperandType;
1847};
1848
1849} // namespace
1850
1851/// The Strlen Idiom we are trying to detect has the following structure
1852///
1853/// preheader:
1854/// ...
1855/// br label %body, ...
1856///
1857/// body:
1858/// ... ; %0 is incremented by a gep
1859/// %1 = load i8, ptr %0, align 1
1860/// %2 = icmp eq i8 %1, 0
1861/// br i1 %2, label %exit, label %body
1862///
1863/// exit:
1864/// %lcssa = phi [%0, %body], ...
1865///
1866/// We expect the strlen idiom to have a load of a character type that
1867/// is compared against '\0', and such load pointer operand must have scev
1868/// expression of the form {%str,+,c} where c is a ConstantInt of the
1869/// appropiate character width for the idiom, and %str is the base of the string
1870/// And, that all lcssa phis have the form {...,+,n} where n is a constant,
1871///
1872/// When transforming the output of the strlen idiom, the lccsa phi are
1873/// expanded using SCEVExpander as {base scev,+,a} -> (base scev + a * strlen)
1874/// and all subsequent uses are replaced. For example,
1875///
1876/// \code{.c}
1877/// const char* base = str;
1878/// while (*str != '\0')
1879/// ++str;
1880/// size_t result = str - base;
1881/// \endcode
1882///
1883/// will be transformed as follows: The idiom will be replaced by a strlen
1884/// computation to compute the address of the null terminator of the string.
1885///
1886/// \code{.c}
1887/// const char* base = str;
1888/// const char* end = base + strlen(str);
1889/// size_t result = end - base;
1890/// \endcode
1891///
1892/// In the case we index by an induction variable, as long as the induction
1893/// variable has a constant int increment, we can replace all such indvars
1894/// with the closed form computation of strlen
1895///
1896/// \code{.c}
1897/// size_t i = 0;
1898/// while (str[i] != '\0')
1899/// ++i;
1900/// size_t result = i;
1901/// \endcode
1902///
1903/// Will be replaced by
1904///
1905/// \code{.c}
1906/// size_t i = 0 + strlen(str);
1907/// size_t result = i;
1908/// \endcode
1909///
1910bool LoopIdiomRecognize::recognizeAndInsertStrLen() {
1911 if (DisableLIRP::All)
1912 return false;
1913
1914 StrlenVerifier Verifier(CurLoop, SE, TLI);
1915
1916 if (!Verifier.isValidStrlenIdiom())
1917 return false;
1918
1919 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1920 BasicBlock *LoopBody = *CurLoop->block_begin();
1921 BasicBlock *LoopExitBB = CurLoop->getExitBlock();
1922 CondBrInst *LoopTerm = cast<CondBrInst>(Val: LoopBody->getTerminator());
1923 assert(Preheader && LoopBody && LoopExitBB &&
1924 "Should be verified to be valid by StrlenVerifier");
1925
1926 if (Verifier.OpWidth == 8) {
1927 if (DisableLIRP::Strlen)
1928 return false;
1929 if (!isLibFuncEmittable(M: Preheader->getModule(), TLI, TheLibFunc: LibFunc_strlen))
1930 return false;
1931 } else {
1932 if (DisableLIRP::Wcslen)
1933 return false;
1934 if (!isLibFuncEmittable(M: Preheader->getModule(), TLI, TheLibFunc: LibFunc_wcslen))
1935 return false;
1936 }
1937
1938 IRBuilder<> Builder(Preheader->getTerminator());
1939 Builder.SetCurrentDebugLocation(CurLoop->getStartLoc());
1940 SCEVExpander Expander(*SE, "strlen_idiom");
1941 Value *MaterialzedBase = Expander.expandCodeFor(
1942 SH: Verifier.LoadBaseEv, Ty: Verifier.LoadBaseEv->getType(),
1943 I: Builder.GetInsertPoint());
1944
1945 Value *StrLenFunc = nullptr;
1946 if (Verifier.OpWidth == 8) {
1947 StrLenFunc = emitStrLen(Ptr: MaterialzedBase, B&: Builder, DL: *DL, TLI);
1948 } else {
1949 StrLenFunc = emitWcsLen(Ptr: MaterialzedBase, B&: Builder, DL: *DL, TLI);
1950 }
1951 assert(StrLenFunc && "Failed to emit strlen function.");
1952
1953 const SCEV *StrlenEv = SE->getSCEV(V: StrLenFunc);
1954 SmallVector<PHINode *, 4> Cleanup;
1955 for (PHINode &PN : LoopExitBB->phis()) {
1956 // We can now materialize the loop output as all phi have scev {base,+,a}.
1957 // We expand the phi as:
1958 // %strlen = call i64 @strlen(%str)
1959 // %phi.new = base expression + step * %strlen
1960 const SCEV *Ev = SE->getSCEV(V: &PN);
1961 const SCEVAddRecExpr *AddRecEv = dyn_cast<SCEVAddRecExpr>(Val: Ev);
1962 const SCEVConstant *Step =
1963 dyn_cast<SCEVConstant>(Val: AddRecEv->getStepRecurrence(SE&: *SE));
1964 const SCEV *Base = AddRecEv->getStart();
1965
1966 // It is safe to truncate to base since if base is narrower than size_t
1967 // the equivalent user code will have to truncate anyways.
1968 const SCEV *NewEv = SE->getAddExpr(
1969 LHS: Base, RHS: SE->getMulExpr(LHS: Step, RHS: SE->getTruncateOrSignExtend(
1970 V: StrlenEv, Ty: Base->getType())));
1971
1972 Value *MaterializedPHI = Expander.expandCodeFor(SH: NewEv, Ty: NewEv->getType(),
1973 I: Builder.GetInsertPoint());
1974 Expander.clear();
1975 PN.replaceAllUsesWith(V: MaterializedPHI);
1976 Cleanup.push_back(Elt: &PN);
1977 }
1978
1979 // All LCSSA Loop Phi are dead, the left over dead loop body can be cleaned
1980 // up by later passes
1981 for (PHINode *PN : Cleanup)
1982 RecursivelyDeleteDeadPHINode(PN);
1983
1984 // LoopDeletion only delete invariant loops with known trip-count. We can
1985 // update the condition so it will reliablely delete the invariant loop
1986 assert((LoopTerm->getSuccessor(0) == LoopBody ||
1987 LoopTerm->getSuccessor(1) == LoopBody) &&
1988 "loop body must have a successor that is it self");
1989 ConstantInt *NewLoopCond = LoopTerm->getSuccessor(i: 0) == LoopBody
1990 ? Builder.getFalse()
1991 : Builder.getTrue();
1992 LoopTerm->setCondition(NewLoopCond);
1993 SE->forgetLoop(L: CurLoop);
1994
1995 ++NumStrLen;
1996 LLVM_DEBUG(dbgs() << " Formed strlen idiom: " << *StrLenFunc << "\n");
1997 ORE.emit(RemarkBuilder: [&]() {
1998 return OptimizationRemark(DEBUG_TYPE, "recognizeAndInsertStrLen",
1999 CurLoop->getStartLoc(), Preheader)
2000 << "Transformed " << StrLenFunc->getName() << " loop idiom";
2001 });
2002
2003 return true;
2004}
2005
2006/// Check if the given conditional branch is based on an unsigned less-than
2007/// comparison between a variable and a constant, and if the comparison is false
2008/// the control yields to the loop entry. If the branch matches the behaviour,
2009/// the variable involved in the comparison is returned.
2010static Value *matchShiftULTCondition(CondBrInst *BI, BasicBlock *LoopEntry,
2011 APInt &Threshold) {
2012 ICmpInst *Cond = dyn_cast<ICmpInst>(Val: BI->getCondition());
2013 if (!Cond)
2014 return nullptr;
2015
2016 ConstantInt *CmpConst = dyn_cast<ConstantInt>(Val: Cond->getOperand(i_nocapture: 1));
2017 if (!CmpConst)
2018 return nullptr;
2019
2020 BasicBlock *FalseSucc = BI->getSuccessor(i: 1);
2021 ICmpInst::Predicate Pred = Cond->getPredicate();
2022
2023 if (Pred == ICmpInst::ICMP_ULT && FalseSucc == LoopEntry) {
2024 Threshold = CmpConst->getValue();
2025 return Cond->getOperand(i_nocapture: 0);
2026 }
2027
2028 return nullptr;
2029}
2030
2031// Check if the recurrence variable `VarX` is in the right form to create
2032// the idiom. Returns the value coerced to a PHINode if so.
2033static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
2034 BasicBlock *LoopEntry) {
2035 auto *PhiX = dyn_cast<PHINode>(Val: VarX);
2036 if (PhiX && PhiX->getParent() == LoopEntry &&
2037 (PhiX->getOperand(i_nocapture: 0) == DefX || PhiX->getOperand(i_nocapture: 1) == DefX))
2038 return PhiX;
2039 return nullptr;
2040}
2041
2042/// Return true if the idiom is detected in the loop.
2043///
2044/// Additionally:
2045/// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
2046/// or nullptr if there is no such.
2047/// 2) \p CntPhi is set to the corresponding phi node
2048/// or nullptr if there is no such.
2049/// 3) \p InitX is set to the value whose CTLZ could be used.
2050/// 4) \p DefX is set to the instruction calculating Loop exit condition.
2051/// 5) \p Threshold is set to the constant involved in the unsigned less-than
2052/// comparison.
2053///
2054/// The core idiom we are trying to detect is:
2055/// \code
2056/// if (x0 < 2)
2057/// goto loop-exit // the precondition of the loop
2058/// cnt0 = init-val
2059/// do {
2060/// x = phi (x0, x.next); //PhiX
2061/// cnt = phi (cnt0, cnt.next)
2062///
2063/// cnt.next = cnt + 1;
2064/// ...
2065/// x.next = x >> 1; // DefX
2066/// } while (x >= 4)
2067/// loop-exit:
2068/// \endcode
2069static bool detectShiftUntilLessThanIdiom(Loop *CurLoop, const DataLayout &DL,
2070 Intrinsic::ID &IntrinID,
2071 Value *&InitX, Instruction *&CntInst,
2072 PHINode *&CntPhi, Instruction *&DefX,
2073 APInt &Threshold) {
2074 BasicBlock *LoopEntry;
2075
2076 DefX = nullptr;
2077 CntInst = nullptr;
2078 CntPhi = nullptr;
2079 LoopEntry = *(CurLoop->block_begin());
2080
2081 // step 1: Check if the loop-back branch is in desirable form.
2082 auto *EntryBI = dyn_cast<CondBrInst>(Val: LoopEntry->getTerminator());
2083 if (!EntryBI)
2084 return false;
2085 if (Value *T = matchShiftULTCondition(BI: EntryBI, LoopEntry, Threshold))
2086 DefX = dyn_cast<Instruction>(Val: T);
2087 else
2088 return false;
2089
2090 // step 2: Check the recurrence of variable X
2091 if (!DefX || !isa<PHINode>(Val: DefX))
2092 return false;
2093
2094 PHINode *VarPhi = cast<PHINode>(Val: DefX);
2095 int Idx = VarPhi->getBasicBlockIndex(BB: LoopEntry);
2096 if (Idx == -1)
2097 return false;
2098
2099 DefX = dyn_cast<Instruction>(Val: VarPhi->getIncomingValue(i: Idx));
2100 if (!DefX || DefX->getNumOperands() == 0 || DefX->getOperand(i: 0) != VarPhi)
2101 return false;
2102
2103 // step 3: detect instructions corresponding to "x.next = x >> 1"
2104 if (DefX->getOpcode() != Instruction::LShr)
2105 return false;
2106
2107 IntrinID = Intrinsic::ctlz;
2108 ConstantInt *Shft = dyn_cast<ConstantInt>(Val: DefX->getOperand(i: 1));
2109 if (!Shft || !Shft->isOne())
2110 return false;
2111
2112 InitX = VarPhi->getIncomingValueForBlock(BB: CurLoop->getLoopPreheader());
2113
2114 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
2115 // or cnt.next = cnt + -1.
2116 // TODO: We can skip the step. If loop trip count is known (CTLZ),
2117 // then all uses of "cnt.next" could be optimized to the trip count
2118 // plus "cnt0". Currently it is not optimized.
2119 // This step could be used to detect POPCNT instruction:
2120 // cnt.next = cnt + (x.next & 1)
2121 for (Instruction &Inst :
2122 llvm::make_range(x: LoopEntry->getFirstNonPHIIt(), y: LoopEntry->end())) {
2123 if (Inst.getOpcode() != Instruction::Add)
2124 continue;
2125
2126 ConstantInt *Inc = dyn_cast<ConstantInt>(Val: Inst.getOperand(i: 1));
2127 if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
2128 continue;
2129
2130 PHINode *Phi = getRecurrenceVar(VarX: Inst.getOperand(i: 0), DefX: &Inst, LoopEntry);
2131 if (!Phi)
2132 continue;
2133
2134 CntInst = &Inst;
2135 CntPhi = Phi;
2136 break;
2137 }
2138 if (!CntInst)
2139 return false;
2140
2141 return true;
2142}
2143
2144/// Return true iff the idiom is detected in the loop.
2145///
2146/// Additionally:
2147/// 1) \p CntInst is set to the instruction counting the population bit.
2148/// 2) \p CntPhi is set to the corresponding phi node.
2149/// 3) \p Var is set to the value whose population bits are being counted.
2150///
2151/// The core idiom we are trying to detect is:
2152/// \code
2153/// if (x0 != 0)
2154/// goto loop-exit // the precondition of the loop
2155/// cnt0 = init-val;
2156/// do {
2157/// x1 = phi (x0, x2);
2158/// cnt1 = phi(cnt0, cnt2);
2159///
2160/// cnt2 = cnt1 + 1;
2161/// ...
2162/// x2 = x1 & (x1 - 1);
2163/// ...
2164/// } while(x != 0);
2165///
2166/// loop-exit:
2167/// \endcode
2168static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
2169 Instruction *&CntInst, PHINode *&CntPhi,
2170 Value *&Var) {
2171 // step 1: Check to see if the look-back branch match this pattern:
2172 // "if (a!=0) goto loop-entry".
2173 BasicBlock *LoopEntry;
2174 Instruction *DefX2, *CountInst;
2175 Value *VarX1, *VarX0;
2176 PHINode *PhiX, *CountPhi;
2177
2178 DefX2 = CountInst = nullptr;
2179 VarX1 = VarX0 = nullptr;
2180 PhiX = CountPhi = nullptr;
2181 LoopEntry = *(CurLoop->block_begin());
2182
2183 // step 1: Check if the loop-back branch is in desirable form.
2184 {
2185 auto *LoopTerm = dyn_cast<CondBrInst>(Val: LoopEntry->getTerminator());
2186 if (!LoopTerm)
2187 return false;
2188 DefX2 = dyn_cast_or_null<Instruction>(Val: matchCondition(BI: LoopTerm, LoopEntry));
2189 }
2190
2191 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
2192 {
2193 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
2194 return false;
2195
2196 BinaryOperator *SubOneOp;
2197
2198 if ((SubOneOp = dyn_cast<BinaryOperator>(Val: DefX2->getOperand(i: 0))))
2199 VarX1 = DefX2->getOperand(i: 1);
2200 else {
2201 VarX1 = DefX2->getOperand(i: 0);
2202 SubOneOp = dyn_cast<BinaryOperator>(Val: DefX2->getOperand(i: 1));
2203 }
2204 if (!SubOneOp || SubOneOp->getOperand(i_nocapture: 0) != VarX1)
2205 return false;
2206
2207 ConstantInt *Dec = dyn_cast<ConstantInt>(Val: SubOneOp->getOperand(i_nocapture: 1));
2208 if (!Dec ||
2209 !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) ||
2210 (SubOneOp->getOpcode() == Instruction::Add &&
2211 Dec->isMinusOne()))) {
2212 return false;
2213 }
2214 }
2215
2216 // step 3: Check the recurrence of variable X
2217 PhiX = getRecurrenceVar(VarX: VarX1, DefX: DefX2, LoopEntry);
2218 if (!PhiX)
2219 return false;
2220
2221 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
2222 {
2223 CountInst = nullptr;
2224 for (Instruction &Inst :
2225 llvm::make_range(x: LoopEntry->getFirstNonPHIIt(), y: LoopEntry->end())) {
2226 if (Inst.getOpcode() != Instruction::Add)
2227 continue;
2228
2229 ConstantInt *Inc = dyn_cast<ConstantInt>(Val: Inst.getOperand(i: 1));
2230 if (!Inc || !Inc->isOne())
2231 continue;
2232
2233 PHINode *Phi = getRecurrenceVar(VarX: Inst.getOperand(i: 0), DefX: &Inst, LoopEntry);
2234 if (!Phi)
2235 continue;
2236
2237 // Check if the result of the instruction is live of the loop.
2238 bool LiveOutLoop = false;
2239 for (User *U : Inst.users()) {
2240 if ((cast<Instruction>(Val: U))->getParent() != LoopEntry) {
2241 LiveOutLoop = true;
2242 break;
2243 }
2244 }
2245
2246 if (LiveOutLoop) {
2247 CountInst = &Inst;
2248 CountPhi = Phi;
2249 break;
2250 }
2251 }
2252
2253 if (!CountInst)
2254 return false;
2255 }
2256
2257 // step 5: check if the precondition is in this form:
2258 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
2259 {
2260 auto *PreCondBr = dyn_cast<CondBrInst>(Val: PreCondBB->getTerminator());
2261 if (!PreCondBr)
2262 return false;
2263 Value *T = matchCondition(BI: PreCondBr, LoopEntry: CurLoop->getLoopPreheader());
2264 if (T != PhiX->getOperand(i_nocapture: 0) && T != PhiX->getOperand(i_nocapture: 1))
2265 return false;
2266
2267 CntInst = CountInst;
2268 CntPhi = CountPhi;
2269 Var = T;
2270 }
2271
2272 return true;
2273}
2274
2275/// Return true if the idiom is detected in the loop.
2276///
2277/// Additionally:
2278/// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
2279/// or nullptr if there is no such.
2280/// 2) \p CntPhi is set to the corresponding phi node
2281/// or nullptr if there is no such.
2282/// 3) \p Var is set to the value whose CTLZ could be used.
2283/// 4) \p DefX is set to the instruction calculating Loop exit condition.
2284///
2285/// The core idiom we are trying to detect is:
2286/// \code
2287/// if (x0 == 0)
2288/// goto loop-exit // the precondition of the loop
2289/// cnt0 = init-val;
2290/// do {
2291/// x = phi (x0, x.next); //PhiX
2292/// cnt = phi(cnt0, cnt.next);
2293///
2294/// cnt.next = cnt + 1;
2295/// ...
2296/// x.next = x >> 1; // DefX
2297/// ...
2298/// } while(x.next != 0);
2299///
2300/// loop-exit:
2301/// \endcode
2302static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL,
2303 Intrinsic::ID &IntrinID, Value *&InitX,
2304 Instruction *&CntInst, PHINode *&CntPhi,
2305 Instruction *&DefX) {
2306 BasicBlock *LoopEntry;
2307 Value *VarX = nullptr;
2308
2309 DefX = nullptr;
2310 CntInst = nullptr;
2311 CntPhi = nullptr;
2312 LoopEntry = *(CurLoop->block_begin());
2313
2314 // step 1: Check if the loop-back branch is in desirable form.
2315 auto *LoopTerm = dyn_cast<CondBrInst>(Val: LoopEntry->getTerminator());
2316 if (!LoopTerm)
2317 return false;
2318 DefX = dyn_cast_or_null<Instruction>(Val: matchCondition(BI: LoopTerm, LoopEntry));
2319
2320 // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1"
2321 if (!DefX || !DefX->isShift())
2322 return false;
2323 IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz :
2324 Intrinsic::ctlz;
2325 ConstantInt *Shft = dyn_cast<ConstantInt>(Val: DefX->getOperand(i: 1));
2326 if (!Shft || !Shft->isOne())
2327 return false;
2328 VarX = DefX->getOperand(i: 0);
2329
2330 // step 3: Check the recurrence of variable X
2331 PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
2332 if (!PhiX)
2333 return false;
2334
2335 InitX = PhiX->getIncomingValueForBlock(BB: CurLoop->getLoopPreheader());
2336
2337 // Make sure the initial value can't be negative otherwise the ashr in the
2338 // loop might never reach zero which would make the loop infinite.
2339 if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(V: InitX, SQ: DL))
2340 return false;
2341
2342 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
2343 // or cnt.next = cnt + -1.
2344 // TODO: We can skip the step. If loop trip count is known (CTLZ),
2345 // then all uses of "cnt.next" could be optimized to the trip count
2346 // plus "cnt0". Currently it is not optimized.
2347 // This step could be used to detect POPCNT instruction:
2348 // cnt.next = cnt + (x.next & 1)
2349 for (Instruction &Inst :
2350 llvm::make_range(x: LoopEntry->getFirstNonPHIIt(), y: LoopEntry->end())) {
2351 if (Inst.getOpcode() != Instruction::Add)
2352 continue;
2353
2354 ConstantInt *Inc = dyn_cast<ConstantInt>(Val: Inst.getOperand(i: 1));
2355 if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
2356 continue;
2357
2358 PHINode *Phi = getRecurrenceVar(VarX: Inst.getOperand(i: 0), DefX: &Inst, LoopEntry);
2359 if (!Phi)
2360 continue;
2361
2362 CntInst = &Inst;
2363 CntPhi = Phi;
2364 break;
2365 }
2366 if (!CntInst)
2367 return false;
2368
2369 return true;
2370}
2371
2372// Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always
2373// profitable if we delete the loop.
2374bool LoopIdiomRecognize::isProfitableToInsertFFS(Intrinsic::ID IntrinID,
2375 Value *InitX, bool ZeroCheck,
2376 size_t CanonicalSize) {
2377 const Value *Args[] = {InitX,
2378 ConstantInt::getBool(Context&: InitX->getContext(), V: ZeroCheck)};
2379
2380 // @llvm.dbg doesn't count as they have no semantic effect.
2381 auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug();
2382 uint32_t HeaderSize =
2383 std::distance(first: InstWithoutDebugIt.begin(), last: InstWithoutDebugIt.end());
2384
2385 IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args);
2386 InstructionCost Cost = TTI->getIntrinsicInstrCost(
2387 ICA: Attrs, CostKind: TargetTransformInfo::TCK_SizeAndLatency);
2388 if (HeaderSize != CanonicalSize && Cost > TargetTransformInfo::TCC_Basic)
2389 return false;
2390
2391 return true;
2392}
2393
2394/// Convert CTLZ / CTTZ idiom loop into countable loop.
2395/// If CTLZ / CTTZ inserted as a new trip count returns true; otherwise,
2396/// returns false.
2397bool LoopIdiomRecognize::insertFFSIfProfitable(Intrinsic::ID IntrinID,
2398 Value *InitX, Instruction *DefX,
2399 PHINode *CntPhi,
2400 Instruction *CntInst) {
2401 bool IsCntPhiUsedOutsideLoop = false;
2402 for (User *U : CntPhi->users())
2403 if (!CurLoop->contains(Inst: cast<Instruction>(Val: U))) {
2404 IsCntPhiUsedOutsideLoop = true;
2405 break;
2406 }
2407 bool IsCntInstUsedOutsideLoop = false;
2408 for (User *U : CntInst->users())
2409 if (!CurLoop->contains(Inst: cast<Instruction>(Val: U))) {
2410 IsCntInstUsedOutsideLoop = true;
2411 break;
2412 }
2413 // If both CntInst and CntPhi are used outside the loop the profitability
2414 // is questionable.
2415 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
2416 return false;
2417
2418 // For some CPUs result of CTLZ(X) intrinsic is undefined
2419 // when X is 0. If we can not guarantee X != 0, we need to check this
2420 // when expand.
2421 bool ZeroCheck = false;
2422 // It is safe to assume Preheader exist as it was checked in
2423 // parent function RunOnLoop.
2424 BasicBlock *PH = CurLoop->getLoopPreheader();
2425
2426 // If we are using the count instruction outside the loop, make sure we
2427 // have a zero check as a precondition. Without the check the loop would run
2428 // one iteration for before any check of the input value. This means 0 and 1
2429 // would have identical behavior in the original loop and thus
2430 if (!IsCntPhiUsedOutsideLoop) {
2431 auto *PreCondBB = PH->getSinglePredecessor();
2432 if (!PreCondBB)
2433 return false;
2434 auto *PreCondBI = dyn_cast<CondBrInst>(Val: PreCondBB->getTerminator());
2435 if (!PreCondBI)
2436 return false;
2437 if (matchCondition(BI: PreCondBI, LoopEntry: PH) != InitX)
2438 return false;
2439 ZeroCheck = true;
2440 }
2441
2442 // FFS idiom loop has only 6 instructions:
2443 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
2444 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
2445 // %shr = ashr %n.addr.0, 1
2446 // %tobool = icmp eq %shr, 0
2447 // %inc = add nsw %i.0, 1
2448 // br i1 %tobool
2449 size_t IdiomCanonicalSize = 6;
2450 if (!isProfitableToInsertFFS(IntrinID, InitX, ZeroCheck, CanonicalSize: IdiomCanonicalSize))
2451 return false;
2452
2453 transformLoopToCountable(IntrinID, PreCondBB: PH, CntInst, CntPhi, Var: InitX, DefX,
2454 DL: DefX->getDebugLoc(), ZeroCheck,
2455 IsCntPhiUsedOutsideLoop);
2456 return true;
2457}
2458
2459/// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop
2460/// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new
2461/// trip count returns true; otherwise, returns false.
2462bool LoopIdiomRecognize::recognizeAndInsertFFS() {
2463 // Give up if the loop has multiple blocks or multiple backedges.
2464 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
2465 return false;
2466
2467 Intrinsic::ID IntrinID;
2468 Value *InitX;
2469 Instruction *DefX = nullptr;
2470 PHINode *CntPhi = nullptr;
2471 Instruction *CntInst = nullptr;
2472
2473 if (!detectShiftUntilZeroIdiom(CurLoop, DL: *DL, IntrinID, InitX, CntInst, CntPhi,
2474 DefX))
2475 return false;
2476
2477 return insertFFSIfProfitable(IntrinID, InitX, DefX, CntPhi, CntInst);
2478}
2479
2480bool LoopIdiomRecognize::recognizeShiftUntilLessThan() {
2481 // Give up if the loop has multiple blocks or multiple backedges.
2482 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
2483 return false;
2484
2485 Intrinsic::ID IntrinID;
2486 Value *InitX;
2487 Instruction *DefX = nullptr;
2488 PHINode *CntPhi = nullptr;
2489 Instruction *CntInst = nullptr;
2490
2491 APInt LoopThreshold;
2492 if (!detectShiftUntilLessThanIdiom(CurLoop, DL: *DL, IntrinID, InitX, CntInst,
2493 CntPhi, DefX, Threshold&: LoopThreshold))
2494 return false;
2495
2496 if (LoopThreshold == 2) {
2497 // Treat as regular FFS.
2498 return insertFFSIfProfitable(IntrinID, InitX, DefX, CntPhi, CntInst);
2499 }
2500
2501 // Look for Floor Log2 Idiom.
2502 if (LoopThreshold != 4)
2503 return false;
2504
2505 // Abort if CntPhi is used outside of the loop.
2506 for (User *U : CntPhi->users())
2507 if (!CurLoop->contains(Inst: cast<Instruction>(Val: U)))
2508 return false;
2509
2510 // It is safe to assume Preheader exist as it was checked in
2511 // parent function RunOnLoop.
2512 BasicBlock *PH = CurLoop->getLoopPreheader();
2513 auto *PreCondBB = PH->getSinglePredecessor();
2514 if (!PreCondBB)
2515 return false;
2516 auto *PreCondBI = dyn_cast<CondBrInst>(Val: PreCondBB->getTerminator());
2517 if (!PreCondBI)
2518 return false;
2519
2520 APInt PreLoopThreshold;
2521 if (matchShiftULTCondition(BI: PreCondBI, LoopEntry: PH, Threshold&: PreLoopThreshold) != InitX ||
2522 PreLoopThreshold != 2)
2523 return false;
2524
2525 bool ZeroCheck = true;
2526
2527 // the loop has only 6 instructions:
2528 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
2529 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
2530 // %shr = ashr %n.addr.0, 1
2531 // %tobool = icmp ult %n.addr.0, C
2532 // %inc = add nsw %i.0, 1
2533 // br i1 %tobool
2534 size_t IdiomCanonicalSize = 6;
2535 if (!isProfitableToInsertFFS(IntrinID, InitX, ZeroCheck, CanonicalSize: IdiomCanonicalSize))
2536 return false;
2537
2538 // log2(x) = w − 1 − clz(x)
2539 transformLoopToCountable(IntrinID, PreCondBB: PH, CntInst, CntPhi, Var: InitX, DefX,
2540 DL: DefX->getDebugLoc(), ZeroCheck,
2541 /*IsCntPhiUsedOutsideLoop=*/false,
2542 /*InsertSub=*/true);
2543 return true;
2544}
2545
2546/// Recognizes a population count idiom in a non-countable loop.
2547///
2548/// If detected, transforms the relevant code to issue the popcount intrinsic
2549/// function call, and returns true; otherwise, returns false.
2550bool LoopIdiomRecognize::recognizePopcount() {
2551 if (TTI->getPopcntSupport(IntTyWidthInBit: 32) != TargetTransformInfo::PSK_FastHardware)
2552 return false;
2553
2554 // Counting population are usually conducted by few arithmetic instructions.
2555 // Such instructions can be easily "absorbed" by vacant slots in a
2556 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
2557 // in a compact loop.
2558
2559 // Give up if the loop has multiple blocks or multiple backedges.
2560 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
2561 return false;
2562
2563 BasicBlock *LoopBody = *(CurLoop->block_begin());
2564 if (LoopBody->size() >= 20) {
2565 // The loop is too big, bail out.
2566 return false;
2567 }
2568
2569 // It should have a preheader containing nothing but an unconditional branch.
2570 BasicBlock *PH = CurLoop->getLoopPreheader();
2571 if (!PH || &PH->front() != PH->getTerminator())
2572 return false;
2573 auto *EntryBI = dyn_cast<UncondBrInst>(Val: PH->getTerminator());
2574 if (!EntryBI)
2575 return false;
2576
2577 // It should have a precondition block where the generated popcount intrinsic
2578 // function can be inserted.
2579 auto *PreCondBB = PH->getSinglePredecessor();
2580 if (!PreCondBB)
2581 return false;
2582 auto *PreCondBI = dyn_cast<CondBrInst>(Val: PreCondBB->getTerminator());
2583 if (!PreCondBI)
2584 return false;
2585
2586 Instruction *CntInst;
2587 PHINode *CntPhi;
2588 Value *Val;
2589 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Var&: Val))
2590 return false;
2591
2592 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Var: Val);
2593 return true;
2594}
2595
2596static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
2597 const DebugLoc &DL) {
2598 Value *Ops[] = {Val};
2599 Type *Tys[] = {Val->getType()};
2600
2601 CallInst *CI = IRBuilder.CreateIntrinsic(ID: Intrinsic::ctpop, Types: Tys, Args: Ops);
2602 CI->setDebugLoc(DL);
2603
2604 return CI;
2605}
2606
2607static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
2608 const DebugLoc &DL, bool ZeroCheck,
2609 Intrinsic::ID IID) {
2610 Value *Ops[] = {Val, IRBuilder.getInt1(V: ZeroCheck)};
2611 Type *Tys[] = {Val->getType()};
2612
2613 CallInst *CI = IRBuilder.CreateIntrinsic(ID: IID, Types: Tys, Args: Ops);
2614 CI->setDebugLoc(DL);
2615
2616 return CI;
2617}
2618
2619/// Transform the following loop (Using CTLZ, CTTZ is similar):
2620/// loop:
2621/// CntPhi = PHI [Cnt0, CntInst]
2622/// PhiX = PHI [InitX, DefX]
2623/// CntInst = CntPhi + 1
2624/// DefX = PhiX >> 1
2625/// LOOP_BODY
2626/// Br: loop if (DefX != 0)
2627/// Use(CntPhi) or Use(CntInst)
2628///
2629/// Into:
2630/// If CntPhi used outside the loop:
2631/// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
2632/// Count = CountPrev + 1
2633/// else
2634/// Count = BitWidth(InitX) - CTLZ(InitX)
2635/// loop:
2636/// CntPhi = PHI [Cnt0, CntInst]
2637/// PhiX = PHI [InitX, DefX]
2638/// PhiCount = PHI [Count, Dec]
2639/// CntInst = CntPhi + 1
2640/// DefX = PhiX >> 1
2641/// Dec = PhiCount - 1
2642/// LOOP_BODY
2643/// Br: loop if (Dec != 0)
2644/// Use(CountPrev + Cnt0) // Use(CntPhi)
2645/// or
2646/// Use(Count + Cnt0) // Use(CntInst)
2647///
2648/// If LOOP_BODY is empty the loop will be deleted.
2649/// If CntInst and DefX are not used in LOOP_BODY they will be removed.
2650void LoopIdiomRecognize::transformLoopToCountable(
2651 Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst,
2652 PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL,
2653 bool ZeroCheck, bool IsCntPhiUsedOutsideLoop, bool InsertSub) {
2654 // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block
2655 IRBuilder<> Builder(Preheader->getTerminator());
2656 Builder.SetCurrentDebugLocation(DL);
2657
2658 // If there are no uses of CntPhi crate:
2659 // Count = BitWidth - CTLZ(InitX);
2660 // NewCount = Count;
2661 // If there are uses of CntPhi create:
2662 // NewCount = BitWidth - CTLZ(InitX >> 1);
2663 // Count = NewCount + 1;
2664 Value *InitXNext;
2665 if (IsCntPhiUsedOutsideLoop) {
2666 if (DefX->getOpcode() == Instruction::AShr)
2667 InitXNext = Builder.CreateAShr(LHS: InitX, RHS: 1);
2668 else if (DefX->getOpcode() == Instruction::LShr)
2669 InitXNext = Builder.CreateLShr(LHS: InitX, RHS: 1);
2670 else if (DefX->getOpcode() == Instruction::Shl) // cttz
2671 InitXNext = Builder.CreateShl(LHS: InitX, RHS: 1);
2672 else
2673 llvm_unreachable("Unexpected opcode!");
2674 } else
2675 InitXNext = InitX;
2676 Value *Count =
2677 createFFSIntrinsic(IRBuilder&: Builder, Val: InitXNext, DL, ZeroCheck, IID: IntrinID);
2678 Type *CountTy = Count->getType();
2679 Count = Builder.CreateSub(
2680 LHS: ConstantInt::get(Ty: CountTy, V: CountTy->getIntegerBitWidth()), RHS: Count);
2681 if (InsertSub)
2682 Count = Builder.CreateSub(LHS: Count, RHS: ConstantInt::get(Ty: CountTy, V: 1));
2683 Value *NewCount = Count;
2684 if (IsCntPhiUsedOutsideLoop)
2685 Count = Builder.CreateAdd(LHS: Count, RHS: ConstantInt::get(Ty: CountTy, V: 1));
2686
2687 NewCount = Builder.CreateZExtOrTrunc(V: NewCount, DestTy: CntInst->getType());
2688
2689 Value *CntInitVal = CntPhi->getIncomingValueForBlock(BB: Preheader);
2690 if (cast<ConstantInt>(Val: CntInst->getOperand(i: 1))->isOne()) {
2691 // If the counter was being incremented in the loop, add NewCount to the
2692 // counter's initial value, but only if the initial value is not zero.
2693 ConstantInt *InitConst = dyn_cast<ConstantInt>(Val: CntInitVal);
2694 if (!InitConst || !InitConst->isZero())
2695 NewCount = Builder.CreateAdd(LHS: NewCount, RHS: CntInitVal);
2696 } else {
2697 // If the count was being decremented in the loop, subtract NewCount from
2698 // the counter's initial value.
2699 NewCount = Builder.CreateSub(LHS: CntInitVal, RHS: NewCount);
2700 }
2701
2702 // Step 2: Insert new IV and loop condition:
2703 // loop:
2704 // ...
2705 // PhiCount = PHI [Count, Dec]
2706 // ...
2707 // Dec = PhiCount - 1
2708 // ...
2709 // Br: loop if (Dec != 0)
2710 BasicBlock *Body = *(CurLoop->block_begin());
2711 auto *LbBr = cast<CondBrInst>(Val: Body->getTerminator());
2712 ICmpInst *LbCond = cast<ICmpInst>(Val: LbBr->getCondition());
2713
2714 PHINode *TcPhi = PHINode::Create(Ty: CountTy, NumReservedValues: 2, NameStr: "tcphi");
2715 TcPhi->insertBefore(InsertPos: Body->begin());
2716
2717 Builder.SetInsertPoint(LbCond);
2718 Instruction *TcDec = cast<Instruction>(Val: Builder.CreateSub(
2719 LHS: TcPhi, RHS: ConstantInt::get(Ty: CountTy, V: 1), Name: "tcdec", HasNUW: false, HasNSW: true));
2720
2721 TcPhi->addIncoming(V: Count, BB: Preheader);
2722 TcPhi->addIncoming(V: TcDec, BB: Body);
2723
2724 CmpInst::Predicate Pred =
2725 (LbBr->getSuccessor(i: 0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
2726 LbCond->setPredicate(Pred);
2727 LbCond->setOperand(i_nocapture: 0, Val_nocapture: TcDec);
2728 LbCond->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: CountTy, V: 0));
2729
2730 // Step 3: All the references to the original counter outside
2731 // the loop are replaced with the NewCount
2732 if (IsCntPhiUsedOutsideLoop)
2733 CntPhi->replaceUsesOutsideBlock(V: NewCount, BB: Body);
2734 else
2735 CntInst->replaceUsesOutsideBlock(V: NewCount, BB: Body);
2736
2737 // step 4: Forget the "non-computable" trip-count SCEV associated with the
2738 // loop. The loop would otherwise not be deleted even if it becomes empty.
2739 SE->forgetLoop(L: CurLoop);
2740}
2741
2742void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
2743 Instruction *CntInst,
2744 PHINode *CntPhi, Value *Var) {
2745 BasicBlock *PreHead = CurLoop->getLoopPreheader();
2746 auto *PreCondBr = cast<CondBrInst>(Val: PreCondBB->getTerminator());
2747 const DebugLoc &DL = CntInst->getDebugLoc();
2748
2749 // Assuming before transformation, the loop is following:
2750 // if (x) // the precondition
2751 // do { cnt++; x &= x - 1; } while(x);
2752
2753 // Step 1: Insert the ctpop instruction at the end of the precondition block
2754 IRBuilder<> Builder(PreCondBr);
2755 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
2756 {
2757 PopCnt = createPopcntIntrinsic(IRBuilder&: Builder, Val: Var, DL);
2758 NewCount = PopCntZext =
2759 Builder.CreateZExtOrTrunc(V: PopCnt, DestTy: cast<IntegerType>(Val: CntPhi->getType()));
2760
2761 if (NewCount != PopCnt)
2762 (cast<Instruction>(Val: NewCount))->setDebugLoc(DL);
2763
2764 // TripCnt is exactly the number of iterations the loop has
2765 TripCnt = NewCount;
2766
2767 // If the population counter's initial value is not zero, insert Add Inst.
2768 Value *CntInitVal = CntPhi->getIncomingValueForBlock(BB: PreHead);
2769 ConstantInt *InitConst = dyn_cast<ConstantInt>(Val: CntInitVal);
2770 if (!InitConst || !InitConst->isZero()) {
2771 NewCount = Builder.CreateAdd(LHS: NewCount, RHS: CntInitVal);
2772 (cast<Instruction>(Val: NewCount))->setDebugLoc(DL);
2773 }
2774 }
2775
2776 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
2777 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
2778 // function would be partial dead code, and downstream passes will drag
2779 // it back from the precondition block to the preheader.
2780 {
2781 ICmpInst *PreCond = cast<ICmpInst>(Val: PreCondBr->getCondition());
2782
2783 Value *Opnd0 = PopCntZext;
2784 Value *Opnd1 = ConstantInt::get(Ty: PopCntZext->getType(), V: 0);
2785 if (PreCond->getOperand(i_nocapture: 0) != Var)
2786 std::swap(a&: Opnd0, b&: Opnd1);
2787
2788 ICmpInst *NewPreCond = cast<ICmpInst>(
2789 Val: Builder.CreateICmp(P: PreCond->getPredicate(), LHS: Opnd0, RHS: Opnd1));
2790 PreCondBr->setCondition(NewPreCond);
2791
2792 RecursivelyDeleteTriviallyDeadInstructions(V: PreCond, TLI);
2793 }
2794
2795 // Step 3: Note that the population count is exactly the trip count of the
2796 // loop in question, which enable us to convert the loop from noncountable
2797 // loop into a countable one. The benefit is twofold:
2798 //
2799 // - If the loop only counts population, the entire loop becomes dead after
2800 // the transformation. It is a lot easier to prove a countable loop dead
2801 // than to prove a noncountable one. (In some C dialects, an infinite loop
2802 // isn't dead even if it computes nothing useful. In general, DCE needs
2803 // to prove a noncountable loop finite before safely delete it.)
2804 //
2805 // - If the loop also performs something else, it remains alive.
2806 // Since it is transformed to countable form, it can be aggressively
2807 // optimized by some optimizations which are in general not applicable
2808 // to a noncountable loop.
2809 //
2810 // After this step, this loop (conceptually) would look like following:
2811 // newcnt = __builtin_ctpop(x);
2812 // t = newcnt;
2813 // if (x)
2814 // do { cnt++; x &= x-1; t--) } while (t > 0);
2815 BasicBlock *Body = *(CurLoop->block_begin());
2816 {
2817 auto *LbBr = cast<CondBrInst>(Val: Body->getTerminator());
2818 ICmpInst *LbCond = cast<ICmpInst>(Val: LbBr->getCondition());
2819 Type *Ty = TripCnt->getType();
2820
2821 PHINode *TcPhi = PHINode::Create(Ty, NumReservedValues: 2, NameStr: "tcphi");
2822 TcPhi->insertBefore(InsertPos: Body->begin());
2823
2824 Builder.SetInsertPoint(LbCond);
2825 Instruction *TcDec = cast<Instruction>(
2826 Val: Builder.CreateSub(LHS: TcPhi, RHS: ConstantInt::get(Ty, V: 1),
2827 Name: "tcdec", HasNUW: false, HasNSW: true));
2828
2829 TcPhi->addIncoming(V: TripCnt, BB: PreHead);
2830 TcPhi->addIncoming(V: TcDec, BB: Body);
2831
2832 CmpInst::Predicate Pred =
2833 (LbBr->getSuccessor(i: 0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
2834 LbCond->setPredicate(Pred);
2835 LbCond->setOperand(i_nocapture: 0, Val_nocapture: TcDec);
2836 LbCond->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty, V: 0));
2837 }
2838
2839 // Step 4: All the references to the original population counter outside
2840 // the loop are replaced with the NewCount -- the value returned from
2841 // __builtin_ctpop().
2842 CntInst->replaceUsesOutsideBlock(V: NewCount, BB: Body);
2843
2844 // step 5: Forget the "non-computable" trip-count SCEV associated with the
2845 // loop. The loop would otherwise not be deleted even if it becomes empty.
2846 SE->forgetLoop(L: CurLoop);
2847}
2848
2849/// Match loop-invariant value.
2850template <typename SubPattern_t> struct match_LoopInvariant {
2851 SubPattern_t SubPattern;
2852 const Loop *L;
2853
2854 match_LoopInvariant(const SubPattern_t &SP, const Loop *L)
2855 : SubPattern(SP), L(L) {}
2856
2857 template <typename ITy> bool match(ITy *V) const {
2858 return L->isLoopInvariant(V) && SubPattern.match(V);
2859 }
2860};
2861
2862/// Matches if the value is loop-invariant.
2863template <typename Ty>
2864inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) {
2865 return match_LoopInvariant<Ty>(M, L);
2866}
2867
2868/// Return true if the idiom is detected in the loop.
2869///
2870/// The core idiom we are trying to detect is:
2871/// \code
2872/// entry:
2873/// <...>
2874/// %bitmask = shl i32 1, %bitpos
2875/// br label %loop
2876///
2877/// loop:
2878/// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2879/// %x.curr.bitmasked = and i32 %x.curr, %bitmask
2880/// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2881/// %x.next = shl i32 %x.curr, 1
2882/// <...>
2883/// br i1 %x.curr.isbitunset, label %loop, label %end
2884///
2885/// end:
2886/// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2887/// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2888/// <...>
2889/// \endcode
2890static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX,
2891 Value *&BitMask, Value *&BitPos,
2892 Value *&CurrX, Instruction *&NextX) {
2893 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2894 " Performing shift-until-bittest idiom detection.\n");
2895
2896 // Give up if the loop has multiple blocks or multiple backedges.
2897 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2898 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2899 return false;
2900 }
2901
2902 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2903 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2904 assert(LoopPreheaderBB && "There is always a loop preheader.");
2905
2906 using namespace PatternMatch;
2907
2908 // Step 1: Check if the loop backedge is in desirable form.
2909
2910 CmpPredicate Pred;
2911 Value *CmpLHS, *CmpRHS;
2912 BasicBlock *TrueBB, *FalseBB;
2913 if (!match(V: LoopHeaderBB->getTerminator(),
2914 P: m_Br(C: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS)),
2915 T: m_BasicBlock(V&: TrueBB), F: m_BasicBlock(V&: FalseBB)))) {
2916 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2917 return false;
2918 }
2919
2920 // Step 2: Check if the backedge's condition is in desirable form.
2921
2922 auto MatchVariableBitMask = [&]() {
2923 return ICmpInst::isEquality(P: Pred) && match(V: CmpRHS, P: m_Zero()) &&
2924 match(V: CmpLHS,
2925 P: m_c_And(L: m_Value(V&: CurrX),
2926 R: m_CombineAnd(
2927 L: m_Value(V&: BitMask),
2928 R: m_LoopInvariant(M: m_Shl(L: m_One(), R: m_Value(V&: BitPos)),
2929 L: CurLoop))));
2930 };
2931
2932 auto MatchDecomposableConstantBitMask = [&]() {
2933 auto Res = llvm::decomposeBitTestICmp(
2934 LHS: CmpLHS, RHS: CmpRHS, Pred, /*LookThroughTrunc=*/true,
2935 /*AllowNonZeroC=*/false, /*DecomposeAnd=*/true);
2936 if (Res && Res->Mask.isPowerOf2()) {
2937 assert(ICmpInst::isEquality(Res->Pred));
2938 Pred = Res->Pred;
2939 CurrX = Res->X;
2940 BitMask = ConstantInt::get(Ty: CurrX->getType(), V: Res->Mask);
2941 BitPos = ConstantInt::get(Ty: CurrX->getType(), V: Res->Mask.logBase2());
2942 return true;
2943 }
2944 return false;
2945 };
2946
2947 if (!MatchVariableBitMask() && !MatchDecomposableConstantBitMask()) {
2948 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n");
2949 return false;
2950 }
2951
2952 // Step 3: Check if the recurrence is in desirable form.
2953 auto *CurrXPN = dyn_cast<PHINode>(Val: CurrX);
2954 if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) {
2955 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2956 return false;
2957 }
2958
2959 BaseX = CurrXPN->getIncomingValueForBlock(BB: LoopPreheaderBB);
2960 NextX =
2961 dyn_cast<Instruction>(Val: CurrXPN->getIncomingValueForBlock(BB: LoopHeaderBB));
2962
2963 assert(CurLoop->isLoopInvariant(BaseX) &&
2964 "Expected BaseX to be available in the preheader!");
2965
2966 if (!NextX || !match(V: NextX, P: m_Shl(L: m_Specific(V: CurrX), R: m_One()))) {
2967 // FIXME: support right-shift?
2968 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2969 return false;
2970 }
2971
2972 // Step 4: Check if the backedge's destinations are in desirable form.
2973
2974 assert(ICmpInst::isEquality(Pred) &&
2975 "Should only get equality predicates here.");
2976
2977 // cmp-br is commutative, so canonicalize to a single variant.
2978 if (Pred != ICmpInst::Predicate::ICMP_EQ) {
2979 Pred = ICmpInst::getInversePredicate(pred: Pred);
2980 std::swap(a&: TrueBB, b&: FalseBB);
2981 }
2982
2983 // We expect to exit loop when comparison yields false,
2984 // so when it yields true we should branch back to loop header.
2985 if (TrueBB != LoopHeaderBB) {
2986 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2987 return false;
2988 }
2989
2990 // Okay, idiom checks out.
2991 return true;
2992}
2993
2994/// Look for the following loop:
2995/// \code
2996/// entry:
2997/// <...>
2998/// %bitmask = shl i32 1, %bitpos
2999/// br label %loop
3000///
3001/// loop:
3002/// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
3003/// %x.curr.bitmasked = and i32 %x.curr, %bitmask
3004/// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
3005/// %x.next = shl i32 %x.curr, 1
3006/// <...>
3007/// br i1 %x.curr.isbitunset, label %loop, label %end
3008///
3009/// end:
3010/// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
3011/// %x.next.res = phi i32 [ %x.next, %loop ] <...>
3012/// <...>
3013/// \endcode
3014///
3015/// And transform it into:
3016/// \code
3017/// entry:
3018/// %bitmask = shl i32 1, %bitpos
3019/// %lowbitmask = add i32 %bitmask, -1
3020/// %mask = or i32 %lowbitmask, %bitmask
3021/// %x.masked = and i32 %x, %mask
3022/// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked,
3023/// i1 true)
3024/// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros
3025/// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1
3026/// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos
3027/// %tripcount = add i32 %backedgetakencount, 1
3028/// %x.curr = shl i32 %x, %backedgetakencount
3029/// %x.next = shl i32 %x, %tripcount
3030/// br label %loop
3031///
3032/// loop:
3033/// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ]
3034/// %loop.iv.next = add nuw i32 %loop.iv, 1
3035/// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount
3036/// <...>
3037/// br i1 %loop.ivcheck, label %end, label %loop
3038///
3039/// end:
3040/// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
3041/// %x.next.res = phi i32 [ %x.next, %loop ] <...>
3042/// <...>
3043/// \endcode
3044bool LoopIdiomRecognize::recognizeShiftUntilBitTest() {
3045 bool MadeChange = false;
3046
3047 Value *X, *BitMask, *BitPos, *XCurr;
3048 Instruction *XNext;
3049 if (!detectShiftUntilBitTestIdiom(CurLoop, BaseX&: X, BitMask, BitPos, CurrX&: XCurr,
3050 NextX&: XNext)) {
3051 LLVM_DEBUG(dbgs() << DEBUG_TYPE
3052 " shift-until-bittest idiom detection failed.\n");
3053 return MadeChange;
3054 }
3055 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n");
3056
3057 // Ok, it is the idiom we were looking for, we *could* transform this loop,
3058 // but is it profitable to transform?
3059
3060 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
3061 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
3062 assert(LoopPreheaderBB && "There is always a loop preheader.");
3063
3064 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
3065 assert(SuccessorBB && "There is only a single successor.");
3066
3067 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
3068 Builder.SetCurrentDebugLocation(cast<Instruction>(Val: XCurr)->getDebugLoc());
3069
3070 Intrinsic::ID IntrID = Intrinsic::ctlz;
3071 Type *Ty = X->getType();
3072 unsigned Bitwidth = Ty->getScalarSizeInBits();
3073
3074 TargetTransformInfo::TargetCostKind CostKind =
3075 TargetTransformInfo::TCK_SizeAndLatency;
3076
3077 // The rewrite is considered to be unprofitable iff and only iff the
3078 // intrinsic/shift we'll use are not cheap. Note that we are okay with *just*
3079 // making the loop countable, even if nothing else changes.
3080 IntrinsicCostAttributes Attrs(
3081 IntrID, Ty, {PoisonValue::get(T: Ty), /*is_zero_poison=*/Builder.getTrue()});
3082 InstructionCost Cost = TTI->getIntrinsicInstrCost(ICA: Attrs, CostKind);
3083 if (Cost > TargetTransformInfo::TCC_Basic) {
3084 LLVM_DEBUG(dbgs() << DEBUG_TYPE
3085 " Intrinsic is too costly, not beneficial\n");
3086 return MadeChange;
3087 }
3088 if (TTI->getArithmeticInstrCost(Opcode: Instruction::Shl, Ty, CostKind) >
3089 TargetTransformInfo::TCC_Basic) {
3090 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n");
3091 return MadeChange;
3092 }
3093
3094 // Ok, transform appears worthwhile.
3095 MadeChange = true;
3096
3097 if (!isGuaranteedNotToBeUndefOrPoison(V: BitPos)) {
3098 // BitMask may be computed from BitPos, Freeze BitPos so we can increase
3099 // it's use count.
3100 std::optional<BasicBlock::iterator> InsertPt = std::nullopt;
3101 if (auto *BitPosI = dyn_cast<Instruction>(Val: BitPos))
3102 InsertPt = BitPosI->getInsertionPointAfterDef();
3103 else
3104 InsertPt = DT->getRoot()->getFirstNonPHIOrDbgOrAlloca();
3105 if (!InsertPt)
3106 return false;
3107 FreezeInst *BitPosFrozen =
3108 new FreezeInst(BitPos, BitPos->getName() + ".fr", *InsertPt);
3109 BitPos->replaceUsesWithIf(New: BitPosFrozen, ShouldReplace: [BitPosFrozen](Use &U) {
3110 return U.getUser() != BitPosFrozen;
3111 });
3112 BitPos = BitPosFrozen;
3113 }
3114
3115 // Step 1: Compute the loop trip count.
3116
3117 Value *LowBitMask = Builder.CreateAdd(LHS: BitMask, RHS: Constant::getAllOnesValue(Ty),
3118 Name: BitPos->getName() + ".lowbitmask");
3119 Value *Mask =
3120 Builder.CreateOr(LHS: LowBitMask, RHS: BitMask, Name: BitPos->getName() + ".mask");
3121 Value *XMasked = Builder.CreateAnd(LHS: X, RHS: Mask, Name: X->getName() + ".masked");
3122 CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic(
3123 ID: IntrID, Types: Ty, Args: {XMasked, /*is_zero_poison=*/Builder.getTrue()},
3124 /*FMFSource=*/nullptr, Name: XMasked->getName() + ".numleadingzeros");
3125 Value *XMaskedNumActiveBits = Builder.CreateSub(
3126 LHS: ConstantInt::get(Ty, V: Ty->getScalarSizeInBits()), RHS: XMaskedNumLeadingZeros,
3127 Name: XMasked->getName() + ".numactivebits", /*HasNUW=*/true,
3128 /*HasNSW=*/Bitwidth != 2);
3129 Value *XMaskedLeadingOnePos =
3130 Builder.CreateAdd(LHS: XMaskedNumActiveBits, RHS: Constant::getAllOnesValue(Ty),
3131 Name: XMasked->getName() + ".leadingonepos", /*HasNUW=*/false,
3132 /*HasNSW=*/Bitwidth > 2);
3133
3134 Value *LoopBackedgeTakenCount = Builder.CreateSub(
3135 LHS: BitPos, RHS: XMaskedLeadingOnePos, Name: CurLoop->getName() + ".backedgetakencount",
3136 /*HasNUW=*/true, /*HasNSW=*/true);
3137 // We know loop's backedge-taken count, but what's loop's trip count?
3138 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
3139 Value *LoopTripCount =
3140 Builder.CreateAdd(LHS: LoopBackedgeTakenCount, RHS: ConstantInt::get(Ty, V: 1),
3141 Name: CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
3142 /*HasNSW=*/Bitwidth != 2);
3143
3144 // Step 2: Compute the recurrence's final value without a loop.
3145
3146 // NewX is always safe to compute, because `LoopBackedgeTakenCount`
3147 // will always be smaller than `bitwidth(X)`, i.e. we never get poison.
3148 Value *NewX = Builder.CreateShl(LHS: X, RHS: LoopBackedgeTakenCount);
3149 NewX->takeName(V: XCurr);
3150 if (auto *I = dyn_cast<Instruction>(Val: NewX))
3151 I->copyIRFlags(V: XNext, /*IncludeWrapFlags=*/true);
3152
3153 Value *NewXNext;
3154 // Rewriting XNext is more complicated, however, because `X << LoopTripCount`
3155 // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen
3156 // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know
3157 // that isn't the case, we'll need to emit an alternative, safe IR.
3158 if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() ||
3159 PatternMatch::match(
3160 V: BitPos, P: PatternMatch::m_SpecificInt_ICMP(
3161 Predicate: ICmpInst::ICMP_NE, Threshold: APInt(Ty->getScalarSizeInBits(),
3162 Ty->getScalarSizeInBits() - 1))))
3163 NewXNext = Builder.CreateShl(LHS: X, RHS: LoopTripCount);
3164 else {
3165 // Otherwise, just additionally shift by one. It's the smallest solution,
3166 // alternatively, we could check that NewX is INT_MIN (or BitPos is )
3167 // and select 0 instead.
3168 NewXNext = Builder.CreateShl(LHS: NewX, RHS: ConstantInt::get(Ty, V: 1));
3169 }
3170
3171 NewXNext->takeName(V: XNext);
3172 if (auto *I = dyn_cast<Instruction>(Val: NewXNext))
3173 I->copyIRFlags(V: XNext, /*IncludeWrapFlags=*/true);
3174
3175 // Step 3: Adjust the successor basic block to receive the computed
3176 // recurrence's final value instead of the recurrence itself.
3177
3178 XCurr->replaceUsesOutsideBlock(V: NewX, BB: LoopHeaderBB);
3179 XNext->replaceUsesOutsideBlock(V: NewXNext, BB: LoopHeaderBB);
3180
3181 // Step 4: Rewrite the loop into a countable form, with canonical IV.
3182
3183 // The new canonical induction variable.
3184 Builder.SetInsertPoint(TheBB: LoopHeaderBB, IP: LoopHeaderBB->begin());
3185 auto *IV = Builder.CreatePHI(Ty, NumReservedValues: 2, Name: CurLoop->getName() + ".iv");
3186
3187 // The induction itself.
3188 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
3189 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
3190 auto *IVNext =
3191 Builder.CreateAdd(LHS: IV, RHS: ConstantInt::get(Ty, V: 1), Name: IV->getName() + ".next",
3192 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
3193
3194 // The loop trip count check.
3195 auto *IVCheck = Builder.CreateICmpEQ(LHS: IVNext, RHS: LoopTripCount,
3196 Name: CurLoop->getName() + ".ivcheck");
3197 SmallVector<uint32_t> BranchWeights;
3198 const bool HasBranchWeights =
3199 !ProfcheckDisableMetadataFixes &&
3200 extractBranchWeights(I: *LoopHeaderBB->getTerminator(), Weights&: BranchWeights);
3201
3202 auto *BI = Builder.CreateCondBr(Cond: IVCheck, True: SuccessorBB, False: LoopHeaderBB);
3203 if (HasBranchWeights) {
3204 if (SuccessorBB == LoopHeaderBB->getTerminator()->getSuccessor(Idx: 1))
3205 std::swap(a&: BranchWeights[0], b&: BranchWeights[1]);
3206 // We're not changing the loop profile, so we can reuse the original loop's
3207 // profile.
3208 setBranchWeights(I&: *BI, Weights: BranchWeights,
3209 /*IsExpected=*/false);
3210 }
3211
3212 LoopHeaderBB->getTerminator()->eraseFromParent();
3213
3214 // Populate the IV PHI.
3215 IV->addIncoming(V: ConstantInt::get(Ty, V: 0), BB: LoopPreheaderBB);
3216 IV->addIncoming(V: IVNext, BB: LoopHeaderBB);
3217
3218 // Step 5: Forget the "non-computable" trip-count SCEV associated with the
3219 // loop. The loop would otherwise not be deleted even if it becomes empty.
3220
3221 SE->forgetLoop(L: CurLoop);
3222
3223 // Other passes will take care of actually deleting the loop if possible.
3224
3225 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n");
3226
3227 ++NumShiftUntilBitTest;
3228 return MadeChange;
3229}
3230
3231/// Return true if the idiom is detected in the loop.
3232///
3233/// The core idiom we are trying to detect is:
3234/// \code
3235/// entry:
3236/// <...>
3237/// %start = <...>
3238/// %extraoffset = <...>
3239/// <...>
3240/// br label %for.cond
3241///
3242/// loop:
3243/// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
3244/// %nbits = add nsw i8 %iv, %extraoffset
3245/// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
3246/// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
3247/// %iv.next = add i8 %iv, 1
3248/// <...>
3249/// br i1 %val.shifted.iszero, label %end, label %loop
3250///
3251/// end:
3252/// %iv.res = phi i8 [ %iv, %loop ] <...>
3253/// %nbits.res = phi i8 [ %nbits, %loop ] <...>
3254/// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
3255/// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
3256/// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
3257/// <...>
3258/// \endcode
3259static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE,
3260 Instruction *&ValShiftedIsZero,
3261 Intrinsic::ID &IntrinID, Instruction *&IV,
3262 Value *&Start, Value *&Val,
3263 const SCEV *&ExtraOffsetExpr,
3264 bool &InvertedCond) {
3265 LLVM_DEBUG(dbgs() << DEBUG_TYPE
3266 " Performing shift-until-zero idiom detection.\n");
3267
3268 // Give up if the loop has multiple blocks or multiple backedges.
3269 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
3270 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
3271 return false;
3272 }
3273
3274 Instruction *ValShifted, *NBits, *IVNext;
3275 Value *ExtraOffset;
3276
3277 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
3278 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
3279 assert(LoopPreheaderBB && "There is always a loop preheader.");
3280
3281 using namespace PatternMatch;
3282
3283 // Step 1: Check if the loop backedge, condition is in desirable form.
3284
3285 CmpPredicate Pred;
3286 BasicBlock *TrueBB, *FalseBB;
3287 if (!match(V: LoopHeaderBB->getTerminator(),
3288 P: m_Br(C: m_Instruction(I&: ValShiftedIsZero), T: m_BasicBlock(V&: TrueBB),
3289 F: m_BasicBlock(V&: FalseBB))) ||
3290 !match(V: ValShiftedIsZero,
3291 P: m_ICmp(Pred, L: m_Instruction(I&: ValShifted), R: m_Zero())) ||
3292 !ICmpInst::isEquality(P: Pred)) {
3293 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
3294 return false;
3295 }
3296
3297 // Step 2: Check if the comparison's operand is in desirable form.
3298 // FIXME: Val could be a one-input PHI node, which we should look past.
3299 if (!match(V: ValShifted, P: m_Shift(L: m_LoopInvariant(M: m_Value(V&: Val), L: CurLoop),
3300 R: m_Instruction(I&: NBits)))) {
3301 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n");
3302 return false;
3303 }
3304 IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz
3305 : Intrinsic::ctlz;
3306
3307 // Step 3: Check if the shift amount is in desirable form.
3308
3309 if (match(V: NBits, P: m_c_Add(L: m_Instruction(I&: IV),
3310 R: m_LoopInvariant(M: m_Value(V&: ExtraOffset), L: CurLoop))) &&
3311 (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap()))
3312 ExtraOffsetExpr = SE->getNegativeSCEV(V: SE->getSCEV(V: ExtraOffset));
3313 else if (match(V: NBits,
3314 P: m_Sub(L: m_Instruction(I&: IV),
3315 R: m_LoopInvariant(M: m_Value(V&: ExtraOffset), L: CurLoop))) &&
3316 NBits->hasNoSignedWrap())
3317 ExtraOffsetExpr = SE->getSCEV(V: ExtraOffset);
3318 else {
3319 IV = NBits;
3320 ExtraOffsetExpr = SE->getZero(Ty: NBits->getType());
3321 }
3322
3323 // Step 4: Check if the recurrence is in desirable form.
3324 auto *IVPN = dyn_cast<PHINode>(Val: IV);
3325 if (!IVPN || IVPN->getParent() != LoopHeaderBB) {
3326 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
3327 return false;
3328 }
3329
3330 Start = IVPN->getIncomingValueForBlock(BB: LoopPreheaderBB);
3331 IVNext = dyn_cast<Instruction>(Val: IVPN->getIncomingValueForBlock(BB: LoopHeaderBB));
3332
3333 if (!IVNext || !match(V: IVNext, P: m_Add(L: m_Specific(V: IVPN), R: m_One()))) {
3334 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
3335 return false;
3336 }
3337
3338 // Step 4: Check if the backedge's destinations are in desirable form.
3339
3340 assert(ICmpInst::isEquality(Pred) &&
3341 "Should only get equality predicates here.");
3342
3343 // cmp-br is commutative, so canonicalize to a single variant.
3344 InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ;
3345 if (InvertedCond) {
3346 Pred = ICmpInst::getInversePredicate(pred: Pred);
3347 std::swap(a&: TrueBB, b&: FalseBB);
3348 }
3349
3350 // We expect to exit loop when comparison yields true,
3351 // so when it yields false we should branch back to loop header.
3352 if (FalseBB != LoopHeaderBB) {
3353 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
3354 return false;
3355 }
3356
3357 // The new, countable, loop will certainly only run a known number of
3358 // iterations, It won't be infinite. But the old loop might be infinite
3359 // under certain conditions. For logical shifts, the value will become zero
3360 // after at most bitwidth(%Val) loop iterations. However, for arithmetic
3361 // right-shift, iff the sign bit was set, the value will never become zero,
3362 // and the loop may never finish.
3363 if (ValShifted->getOpcode() == Instruction::AShr &&
3364 !isMustProgress(L: CurLoop) && !SE->isKnownNonNegative(S: SE->getSCEV(V: Val))) {
3365 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n");
3366 return false;
3367 }
3368
3369 // Okay, idiom checks out.
3370 return true;
3371}
3372
3373/// Look for the following loop:
3374/// \code
3375/// entry:
3376/// <...>
3377/// %start = <...>
3378/// %extraoffset = <...>
3379/// <...>
3380/// br label %loop
3381///
3382/// loop:
3383/// %iv = phi i8 [ %start, %entry ], [ %iv.next, %loop ]
3384/// %nbits = add nsw i8 %iv, %extraoffset
3385/// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
3386/// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
3387/// %iv.next = add i8 %iv, 1
3388/// <...>
3389/// br i1 %val.shifted.iszero, label %end, label %loop
3390///
3391/// end:
3392/// %iv.res = phi i8 [ %iv, %loop ] <...>
3393/// %nbits.res = phi i8 [ %nbits, %loop ] <...>
3394/// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
3395/// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
3396/// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
3397/// <...>
3398/// \endcode
3399///
3400/// And transform it into:
3401/// \code
3402/// entry:
3403/// <...>
3404/// %start = <...>
3405/// %extraoffset = <...>
3406/// <...>
3407/// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0)
3408/// %val.numactivebits = sub i8 8, %val.numleadingzeros
3409/// %extraoffset.neg = sub i8 0, %extraoffset
3410/// %tmp = add i8 %val.numactivebits, %extraoffset.neg
3411/// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start)
3412/// %loop.tripcount = sub i8 %iv.final, %start
3413/// br label %loop
3414///
3415/// loop:
3416/// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ]
3417/// %loop.iv.next = add i8 %loop.iv, 1
3418/// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount
3419/// %iv = add i8 %loop.iv, %start
3420/// <...>
3421/// br i1 %loop.ivcheck, label %end, label %loop
3422///
3423/// end:
3424/// %iv.res = phi i8 [ %iv.final, %loop ] <...>
3425/// <...>
3426/// \endcode
3427bool LoopIdiomRecognize::recognizeShiftUntilZero() {
3428 bool MadeChange = false;
3429
3430 Instruction *ValShiftedIsZero;
3431 Intrinsic::ID IntrID;
3432 Instruction *IV;
3433 Value *Start, *Val;
3434 const SCEV *ExtraOffsetExpr;
3435 bool InvertedCond;
3436 if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrinID&: IntrID, IV,
3437 Start, Val, ExtraOffsetExpr, InvertedCond)) {
3438 LLVM_DEBUG(dbgs() << DEBUG_TYPE
3439 " shift-until-zero idiom detection failed.\n");
3440 return MadeChange;
3441 }
3442 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n");
3443
3444 // Ok, it is the idiom we were looking for, we *could* transform this loop,
3445 // but is it profitable to transform?
3446
3447 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
3448 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
3449 assert(LoopPreheaderBB && "There is always a loop preheader.");
3450
3451 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
3452 assert(SuccessorBB && "There is only a single successor.");
3453
3454 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
3455 Builder.SetCurrentDebugLocation(IV->getDebugLoc());
3456
3457 Type *Ty = Val->getType();
3458 unsigned Bitwidth = Ty->getScalarSizeInBits();
3459
3460 TargetTransformInfo::TargetCostKind CostKind =
3461 TargetTransformInfo::TCK_SizeAndLatency;
3462
3463 // The rewrite is considered to be unprofitable iff and only iff the
3464 // intrinsic we'll use are not cheap. Note that we are okay with *just*
3465 // making the loop countable, even if nothing else changes.
3466 IntrinsicCostAttributes Attrs(
3467 IntrID, Ty, {PoisonValue::get(T: Ty), /*is_zero_poison=*/Builder.getFalse()});
3468 InstructionCost Cost = TTI->getIntrinsicInstrCost(ICA: Attrs, CostKind);
3469 if (Cost > TargetTransformInfo::TCC_Basic) {
3470 LLVM_DEBUG(dbgs() << DEBUG_TYPE
3471 " Intrinsic is too costly, not beneficial\n");
3472 return MadeChange;
3473 }
3474
3475 // Ok, transform appears worthwhile.
3476 MadeChange = true;
3477
3478 bool OffsetIsZero = ExtraOffsetExpr->isZero();
3479
3480 // Step 1: Compute the loop's final IV value / trip count.
3481
3482 CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic(
3483 ID: IntrID, Types: Ty, Args: {Val, /*is_zero_poison=*/Builder.getFalse()},
3484 /*FMFSource=*/nullptr, Name: Val->getName() + ".numleadingzeros");
3485 Value *ValNumActiveBits = Builder.CreateSub(
3486 LHS: ConstantInt::get(Ty, V: Ty->getScalarSizeInBits()), RHS: ValNumLeadingZeros,
3487 Name: Val->getName() + ".numactivebits", /*HasNUW=*/true,
3488 /*HasNSW=*/Bitwidth != 2);
3489
3490 SCEVExpander Expander(*SE, "loop-idiom");
3491 Expander.setInsertPoint(&*Builder.GetInsertPoint());
3492 Value *ExtraOffset = Expander.expandCodeFor(SH: ExtraOffsetExpr);
3493
3494 Value *ValNumActiveBitsOffset = Builder.CreateAdd(
3495 LHS: ValNumActiveBits, RHS: ExtraOffset, Name: ValNumActiveBits->getName() + ".offset",
3496 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true);
3497 Value *IVFinal = Builder.CreateIntrinsic(ID: Intrinsic::smax, Types: {Ty},
3498 Args: {ValNumActiveBitsOffset, Start},
3499 /*FMFSource=*/nullptr, Name: "iv.final");
3500
3501 auto *LoopBackedgeTakenCount = cast<Instruction>(Val: Builder.CreateSub(
3502 LHS: IVFinal, RHS: Start, Name: CurLoop->getName() + ".backedgetakencount",
3503 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true));
3504 // FIXME: or when the offset was `add nuw`
3505
3506 // We know loop's backedge-taken count, but what's loop's trip count?
3507 Value *LoopTripCount =
3508 Builder.CreateAdd(LHS: LoopBackedgeTakenCount, RHS: ConstantInt::get(Ty, V: 1),
3509 Name: CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
3510 /*HasNSW=*/Bitwidth != 2);
3511
3512 // Step 2: Adjust the successor basic block to receive the original
3513 // induction variable's final value instead of the orig. IV itself.
3514
3515 IV->replaceUsesOutsideBlock(V: IVFinal, BB: LoopHeaderBB);
3516
3517 // Step 3: Rewrite the loop into a countable form, with canonical IV.
3518
3519 // The new canonical induction variable.
3520 Builder.SetInsertPoint(TheBB: LoopHeaderBB, IP: LoopHeaderBB->begin());
3521 auto *CIV = Builder.CreatePHI(Ty, NumReservedValues: 2, Name: CurLoop->getName() + ".iv");
3522
3523 // The induction itself.
3524 Builder.SetInsertPoint(TheBB: LoopHeaderBB, IP: LoopHeaderBB->getFirstNonPHIIt());
3525 auto *CIVNext =
3526 Builder.CreateAdd(LHS: CIV, RHS: ConstantInt::get(Ty, V: 1), Name: CIV->getName() + ".next",
3527 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
3528
3529 // The loop trip count check.
3530 auto *CIVCheck = Builder.CreateICmpEQ(LHS: CIVNext, RHS: LoopTripCount,
3531 Name: CurLoop->getName() + ".ivcheck");
3532 auto *NewIVCheck = CIVCheck;
3533 if (InvertedCond) {
3534 NewIVCheck = Builder.CreateNot(V: CIVCheck);
3535 NewIVCheck->takeName(V: ValShiftedIsZero);
3536 }
3537
3538 // The original IV, but rebased to be an offset to the CIV.
3539 auto *IVDePHId = Builder.CreateAdd(LHS: CIV, RHS: Start, Name: "", /*HasNUW=*/false,
3540 /*HasNSW=*/true); // FIXME: what about NUW?
3541 IVDePHId->takeName(V: IV);
3542
3543 // The loop terminator.
3544 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
3545 SmallVector<uint32_t> BranchWeights;
3546 const bool HasBranchWeights =
3547 !ProfcheckDisableMetadataFixes &&
3548 extractBranchWeights(I: *LoopHeaderBB->getTerminator(), Weights&: BranchWeights);
3549
3550 auto *BI = Builder.CreateCondBr(Cond: CIVCheck, True: SuccessorBB, False: LoopHeaderBB);
3551 if (HasBranchWeights) {
3552 if (InvertedCond)
3553 std::swap(a&: BranchWeights[0], b&: BranchWeights[1]);
3554 // We're not changing the loop profile, so we can reuse the original loop's
3555 // profile.
3556 setBranchWeights(I&: *BI, Weights: BranchWeights, /*IsExpected=*/false);
3557 }
3558 LoopHeaderBB->getTerminator()->eraseFromParent();
3559
3560 // Populate the IV PHI.
3561 CIV->addIncoming(V: ConstantInt::get(Ty, V: 0), BB: LoopPreheaderBB);
3562 CIV->addIncoming(V: CIVNext, BB: LoopHeaderBB);
3563
3564 // Step 4: Forget the "non-computable" trip-count SCEV associated with the
3565 // loop. The loop would otherwise not be deleted even if it becomes empty.
3566
3567 SE->forgetLoop(L: CurLoop);
3568
3569 // Step 5: Try to cleanup the loop's body somewhat.
3570 IV->replaceAllUsesWith(V: IVDePHId);
3571 IV->eraseFromParent();
3572
3573 ValShiftedIsZero->replaceAllUsesWith(V: NewIVCheck);
3574 ValShiftedIsZero->eraseFromParent();
3575
3576 // Other passes will take care of actually deleting the loop if possible.
3577
3578 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n");
3579
3580 ++NumShiftUntilZero;
3581 return MadeChange;
3582}
3583