xref: /llvm-project/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp (revision 89e767f1277b43862ca417810f23f70596536b81)
1 //===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass implements an idiom recognizer that transforms simple loops into a
10 // non-loop form.  In cases that this kicks in, it can be a significant
11 // performance win.
12 //
13 // If compiling for code size we avoid idiom recognition if the resulting
14 // code could be larger than the code for the original loop. One way this could
15 // happen is if the loop is not removable after idiom recognition due to the
16 // presence of non-idiom instructions. The initial implementation of the
17 // heuristics applies to idioms in multi-block loops.
18 //
19 //===----------------------------------------------------------------------===//
20 //
21 // TODO List:
22 //
23 // Future loop memory idioms to recognize:
24 //   memcmp, strlen, etc.
25 //
26 // This could recognize common matrix multiplies and dot product idioms and
27 // replace them with calls to BLAS (if linked in??).
28 //
29 //===----------------------------------------------------------------------===//
30 
31 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
32 #include "llvm/ADT/APInt.h"
33 #include "llvm/ADT/ArrayRef.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/MapVector.h"
36 #include "llvm/ADT/SetVector.h"
37 #include "llvm/ADT/SmallPtrSet.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/StringRef.h"
41 #include "llvm/Analysis/AliasAnalysis.h"
42 #include "llvm/Analysis/CmpInstAnalysis.h"
43 #include "llvm/Analysis/LoopInfo.h"
44 #include "llvm/Analysis/LoopPass.h"
45 #include "llvm/Analysis/MemoryLocation.h"
46 #include "llvm/Analysis/MemorySSA.h"
47 #include "llvm/Analysis/MemorySSAUpdater.h"
48 #include "llvm/Analysis/MustExecute.h"
49 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
50 #include "llvm/Analysis/ScalarEvolution.h"
51 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
52 #include "llvm/Analysis/TargetLibraryInfo.h"
53 #include "llvm/Analysis/TargetTransformInfo.h"
54 #include "llvm/Analysis/ValueTracking.h"
55 #include "llvm/IR/BasicBlock.h"
56 #include "llvm/IR/Constant.h"
57 #include "llvm/IR/Constants.h"
58 #include "llvm/IR/DataLayout.h"
59 #include "llvm/IR/DebugLoc.h"
60 #include "llvm/IR/DerivedTypes.h"
61 #include "llvm/IR/Dominators.h"
62 #include "llvm/IR/GlobalValue.h"
63 #include "llvm/IR/GlobalVariable.h"
64 #include "llvm/IR/IRBuilder.h"
65 #include "llvm/IR/InstrTypes.h"
66 #include "llvm/IR/Instruction.h"
67 #include "llvm/IR/Instructions.h"
68 #include "llvm/IR/IntrinsicInst.h"
69 #include "llvm/IR/Intrinsics.h"
70 #include "llvm/IR/LLVMContext.h"
71 #include "llvm/IR/Module.h"
72 #include "llvm/IR/PassManager.h"
73 #include "llvm/IR/PatternMatch.h"
74 #include "llvm/IR/Type.h"
75 #include "llvm/IR/User.h"
76 #include "llvm/IR/Value.h"
77 #include "llvm/IR/ValueHandle.h"
78 #include "llvm/Support/Casting.h"
79 #include "llvm/Support/CommandLine.h"
80 #include "llvm/Support/Debug.h"
81 #include "llvm/Support/InstructionCost.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Transforms/Utils/BuildLibCalls.h"
84 #include "llvm/Transforms/Utils/Local.h"
85 #include "llvm/Transforms/Utils/LoopUtils.h"
86 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
87 #include <algorithm>
88 #include <cassert>
89 #include <cstdint>
90 #include <utility>
91 #include <vector>
92 
93 using namespace llvm;
94 
95 #define DEBUG_TYPE "loop-idiom"
96 
97 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
98 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
99 STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores");
100 STATISTIC(
101     NumShiftUntilBitTest,
102     "Number of uncountable loops recognized as 'shift until bitttest' idiom");
103 STATISTIC(NumShiftUntilZero,
104           "Number of uncountable loops recognized as 'shift until zero' idiom");
105 
106 bool DisableLIRP::All;
107 static cl::opt<bool, true>
108     DisableLIRPAll("disable-" DEBUG_TYPE "-all",
109                    cl::desc("Options to disable Loop Idiom Recognize Pass."),
110                    cl::location(DisableLIRP::All), cl::init(false),
111                    cl::ReallyHidden);
112 
113 bool DisableLIRP::Memset;
114 static cl::opt<bool, true>
115     DisableLIRPMemset("disable-" DEBUG_TYPE "-memset",
116                       cl::desc("Proceed with loop idiom recognize pass, but do "
117                                "not convert loop(s) to memset."),
118                       cl::location(DisableLIRP::Memset), cl::init(false),
119                       cl::ReallyHidden);
120 
121 bool DisableLIRP::Memcpy;
122 static cl::opt<bool, true>
123     DisableLIRPMemcpy("disable-" DEBUG_TYPE "-memcpy",
124                       cl::desc("Proceed with loop idiom recognize pass, but do "
125                                "not convert loop(s) to memcpy."),
126                       cl::location(DisableLIRP::Memcpy), cl::init(false),
127                       cl::ReallyHidden);
128 
129 static cl::opt<bool> UseLIRCodeSizeHeurs(
130     "use-lir-code-size-heurs",
131     cl::desc("Use loop idiom recognition code size heuristics when compiling "
132              "with -Os/-Oz"),
133     cl::init(true), cl::Hidden);
134 
135 namespace {
136 
137 class LoopIdiomRecognize {
138   Loop *CurLoop = nullptr;
139   AliasAnalysis *AA;
140   DominatorTree *DT;
141   LoopInfo *LI;
142   ScalarEvolution *SE;
143   TargetLibraryInfo *TLI;
144   const TargetTransformInfo *TTI;
145   const DataLayout *DL;
146   OptimizationRemarkEmitter &ORE;
147   bool ApplyCodeSizeHeuristics;
148   std::unique_ptr<MemorySSAUpdater> MSSAU;
149 
150 public:
151   explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
152                               LoopInfo *LI, ScalarEvolution *SE,
153                               TargetLibraryInfo *TLI,
154                               const TargetTransformInfo *TTI, MemorySSA *MSSA,
155                               const DataLayout *DL,
156                               OptimizationRemarkEmitter &ORE)
157       : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) {
158     if (MSSA)
159       MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
160   }
161 
162   bool runOnLoop(Loop *L);
163 
164 private:
165   using StoreList = SmallVector<StoreInst *, 8>;
166   using StoreListMap = MapVector<Value *, StoreList>;
167 
168   StoreListMap StoreRefsForMemset;
169   StoreListMap StoreRefsForMemsetPattern;
170   StoreList StoreRefsForMemcpy;
171   bool HasMemset;
172   bool HasMemsetPattern;
173   bool HasMemcpy;
174 
175   /// Return code for isLegalStore()
176   enum LegalStoreKind {
177     None = 0,
178     Memset,
179     MemsetPattern,
180     Memcpy,
181     UnorderedAtomicMemcpy,
182     DontUse // Dummy retval never to be used. Allows catching errors in retval
183             // handling.
184   };
185 
186   /// \name Countable Loop Idiom Handling
187   /// @{
188 
189   bool runOnCountableLoop();
190   bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
191                       SmallVectorImpl<BasicBlock *> &ExitBlocks);
192 
193   void collectStores(BasicBlock *BB);
194   LegalStoreKind isLegalStore(StoreInst *SI);
195   enum class ForMemset { No, Yes };
196   bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
197                          ForMemset For);
198 
199   template <typename MemInst>
200   bool processLoopMemIntrinsic(
201       BasicBlock *BB,
202       bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
203       const SCEV *BECount);
204   bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount);
205   bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
206 
207   bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV,
208                                MaybeAlign StoreAlignment, Value *StoredVal,
209                                Instruction *TheStore,
210                                SmallPtrSetImpl<Instruction *> &Stores,
211                                const SCEVAddRecExpr *Ev, const SCEV *BECount,
212                                bool IsNegStride, bool IsLoopMemset = false);
213   bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
214   bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr,
215                                   const SCEV *StoreSize, MaybeAlign StoreAlign,
216                                   MaybeAlign LoadAlign, Instruction *TheStore,
217                                   Instruction *TheLoad,
218                                   const SCEVAddRecExpr *StoreEv,
219                                   const SCEVAddRecExpr *LoadEv,
220                                   const SCEV *BECount);
221   bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
222                                  bool IsLoopMemset = false);
223 
224   /// @}
225   /// \name Noncountable Loop Idiom Handling
226   /// @{
227 
228   bool runOnNoncountableLoop();
229 
230   bool recognizePopcount();
231   void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
232                                PHINode *CntPhi, Value *Var);
233   bool isProfitableToInsertFFS(Intrinsic::ID IntrinID, Value *InitX,
234                                bool ZeroCheck, size_t CanonicalSize);
235   bool insertFFSIfProfitable(Intrinsic::ID IntrinID, Value *InitX,
236                              Instruction *DefX, PHINode *CntPhi,
237                              Instruction *CntInst);
238   bool recognizeAndInsertFFS();  /// Find First Set: ctlz or cttz
239   bool recognizeShiftUntilLessThan();
240   void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB,
241                                 Instruction *CntInst, PHINode *CntPhi,
242                                 Value *Var, Instruction *DefX,
243                                 const DebugLoc &DL, bool ZeroCheck,
244                                 bool IsCntPhiUsedOutsideLoop,
245                                 bool InsertSub = false);
246 
247   bool recognizeShiftUntilBitTest();
248   bool recognizeShiftUntilZero();
249 
250   /// @}
251 };
252 } // end anonymous namespace
253 
254 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
255                                               LoopStandardAnalysisResults &AR,
256                                               LPMUpdater &) {
257   if (DisableLIRP::All)
258     return PreservedAnalyses::all();
259 
260   const auto *DL = &L.getHeader()->getDataLayout();
261 
262   // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
263   // pass.  Function analyses need to be preserved across loop transformations
264   // but ORE cannot be preserved (see comment before the pass definition).
265   OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
266 
267   LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI,
268                          AR.MSSA, DL, ORE);
269   if (!LIR.runOnLoop(&L))
270     return PreservedAnalyses::all();
271 
272   auto PA = getLoopPassPreservedAnalyses();
273   if (AR.MSSA)
274     PA.preserve<MemorySSAAnalysis>();
275   return PA;
276 }
277 
278 static void deleteDeadInstruction(Instruction *I) {
279   I->replaceAllUsesWith(PoisonValue::get(I->getType()));
280   I->eraseFromParent();
281 }
282 
283 //===----------------------------------------------------------------------===//
284 //
285 //          Implementation of LoopIdiomRecognize
286 //
287 //===----------------------------------------------------------------------===//
288 
289 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
290   CurLoop = L;
291   // If the loop could not be converted to canonical form, it must have an
292   // indirectbr in it, just give up.
293   if (!L->getLoopPreheader())
294     return false;
295 
296   // Disable loop idiom recognition if the function's name is a common idiom.
297   StringRef Name = L->getHeader()->getParent()->getName();
298   if (Name == "memset" || Name == "memcpy")
299     return false;
300 
301   // Determine if code size heuristics need to be applied.
302   ApplyCodeSizeHeuristics =
303       L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs;
304 
305   HasMemset = TLI->has(LibFunc_memset);
306   HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
307   HasMemcpy = TLI->has(LibFunc_memcpy);
308 
309   if (HasMemset || HasMemsetPattern || HasMemcpy)
310     if (SE->hasLoopInvariantBackedgeTakenCount(L))
311       return runOnCountableLoop();
312 
313   return runOnNoncountableLoop();
314 }
315 
316 bool LoopIdiomRecognize::runOnCountableLoop() {
317   const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
318   assert(!isa<SCEVCouldNotCompute>(BECount) &&
319          "runOnCountableLoop() called on a loop without a predictable"
320          "backedge-taken count");
321 
322   // If this loop executes exactly one time, then it should be peeled, not
323   // optimized by this pass.
324   if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
325     if (BECst->getAPInt() == 0)
326       return false;
327 
328   SmallVector<BasicBlock *, 8> ExitBlocks;
329   CurLoop->getUniqueExitBlocks(ExitBlocks);
330 
331   LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
332                     << CurLoop->getHeader()->getParent()->getName()
333                     << "] Countable Loop %" << CurLoop->getHeader()->getName()
334                     << "\n");
335 
336   // The following transforms hoist stores/memsets into the loop pre-header.
337   // Give up if the loop has instructions that may throw.
338   SimpleLoopSafetyInfo SafetyInfo;
339   SafetyInfo.computeLoopSafetyInfo(CurLoop);
340   if (SafetyInfo.anyBlockMayThrow())
341     return false;
342 
343   bool MadeChange = false;
344 
345   // Scan all the blocks in the loop that are not in subloops.
346   for (auto *BB : CurLoop->getBlocks()) {
347     // Ignore blocks in subloops.
348     if (LI->getLoopFor(BB) != CurLoop)
349       continue;
350 
351     MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
352   }
353   return MadeChange;
354 }
355 
356 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
357   const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
358   return ConstStride->getAPInt();
359 }
360 
361 /// getMemSetPatternValue - If a strided store of the specified value is safe to
362 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
363 /// be passed in.  Otherwise, return null.
364 ///
365 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
366 /// just replicate their input array and then pass on to memset_pattern16.
367 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
368   // FIXME: This could check for UndefValue because it can be merged into any
369   // other valid pattern.
370 
371   // If the value isn't a constant, we can't promote it to being in a constant
372   // array.  We could theoretically do a store to an alloca or something, but
373   // that doesn't seem worthwhile.
374   Constant *C = dyn_cast<Constant>(V);
375   if (!C || isa<ConstantExpr>(C))
376     return nullptr;
377 
378   // Only handle simple values that are a power of two bytes in size.
379   uint64_t Size = DL->getTypeSizeInBits(V->getType());
380   if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
381     return nullptr;
382 
383   // Don't care enough about darwin/ppc to implement this.
384   if (DL->isBigEndian())
385     return nullptr;
386 
387   // Convert to size in bytes.
388   Size /= 8;
389 
390   // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
391   // if the top and bottom are the same (e.g. for vectors and large integers).
392   if (Size > 16)
393     return nullptr;
394 
395   // If the constant is exactly 16 bytes, just use it.
396   if (Size == 16)
397     return C;
398 
399   // Otherwise, we'll use an array of the constants.
400   unsigned ArraySize = 16 / Size;
401   ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
402   return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
403 }
404 
405 LoopIdiomRecognize::LegalStoreKind
406 LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
407   // Don't touch volatile stores.
408   if (SI->isVolatile())
409     return LegalStoreKind::None;
410   // We only want simple or unordered-atomic stores.
411   if (!SI->isUnordered())
412     return LegalStoreKind::None;
413 
414   // Avoid merging nontemporal stores.
415   if (SI->getMetadata(LLVMContext::MD_nontemporal))
416     return LegalStoreKind::None;
417 
418   Value *StoredVal = SI->getValueOperand();
419   Value *StorePtr = SI->getPointerOperand();
420 
421   // Don't convert stores of non-integral pointer types to memsets (which stores
422   // integers).
423   if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
424     return LegalStoreKind::None;
425 
426   // Reject stores that are so large that they overflow an unsigned.
427   // When storing out scalable vectors we bail out for now, since the code
428   // below currently only works for constant strides.
429   TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
430   if (SizeInBits.isScalable() || (SizeInBits.getFixedValue() & 7) ||
431       (SizeInBits.getFixedValue() >> 32) != 0)
432     return LegalStoreKind::None;
433 
434   // See if the pointer expression is an AddRec like {base,+,1} on the current
435   // loop, which indicates a strided store.  If we have something else, it's a
436   // random store we can't handle.
437   const SCEVAddRecExpr *StoreEv =
438       dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
439   if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
440     return LegalStoreKind::None;
441 
442   // Check to see if we have a constant stride.
443   if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
444     return LegalStoreKind::None;
445 
446   // See if the store can be turned into a memset.
447 
448   // If the stored value is a byte-wise value (like i32 -1), then it may be
449   // turned into a memset of i8 -1, assuming that all the consecutive bytes
450   // are stored.  A store of i32 0x01020304 can never be turned into a memset,
451   // but it can be turned into memset_pattern if the target supports it.
452   Value *SplatValue = isBytewiseValue(StoredVal, *DL);
453 
454   // Note: memset and memset_pattern on unordered-atomic is yet not supported
455   bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple();
456 
457   // If we're allowed to form a memset, and the stored value would be
458   // acceptable for memset, use it.
459   if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset &&
460       // Verify that the stored value is loop invariant.  If not, we can't
461       // promote the memset.
462       CurLoop->isLoopInvariant(SplatValue)) {
463     // It looks like we can use SplatValue.
464     return LegalStoreKind::Memset;
465   }
466   if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset &&
467       // Don't create memset_pattern16s with address spaces.
468       StorePtr->getType()->getPointerAddressSpace() == 0 &&
469       getMemSetPatternValue(StoredVal, DL)) {
470     // It looks like we can use PatternValue!
471     return LegalStoreKind::MemsetPattern;
472   }
473 
474   // Otherwise, see if the store can be turned into a memcpy.
475   if (HasMemcpy && !DisableLIRP::Memcpy) {
476     // Check to see if the stride matches the size of the store.  If so, then we
477     // know that every byte is touched in the loop.
478     APInt Stride = getStoreStride(StoreEv);
479     unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
480     if (StoreSize != Stride && StoreSize != -Stride)
481       return LegalStoreKind::None;
482 
483     // The store must be feeding a non-volatile load.
484     LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
485 
486     // Only allow non-volatile loads
487     if (!LI || LI->isVolatile())
488       return LegalStoreKind::None;
489     // Only allow simple or unordered-atomic loads
490     if (!LI->isUnordered())
491       return LegalStoreKind::None;
492 
493     // See if the pointer expression is an AddRec like {base,+,1} on the current
494     // loop, which indicates a strided load.  If we have something else, it's a
495     // random load we can't handle.
496     const SCEVAddRecExpr *LoadEv =
497         dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
498     if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
499       return LegalStoreKind::None;
500 
501     // The store and load must share the same stride.
502     if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
503       return LegalStoreKind::None;
504 
505     // Success.  This store can be converted into a memcpy.
506     UnorderedAtomic = UnorderedAtomic || LI->isAtomic();
507     return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy
508                            : LegalStoreKind::Memcpy;
509   }
510   // This store can't be transformed into a memset/memcpy.
511   return LegalStoreKind::None;
512 }
513 
514 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
515   StoreRefsForMemset.clear();
516   StoreRefsForMemsetPattern.clear();
517   StoreRefsForMemcpy.clear();
518   for (Instruction &I : *BB) {
519     StoreInst *SI = dyn_cast<StoreInst>(&I);
520     if (!SI)
521       continue;
522 
523     // Make sure this is a strided store with a constant stride.
524     switch (isLegalStore(SI)) {
525     case LegalStoreKind::None:
526       // Nothing to do
527       break;
528     case LegalStoreKind::Memset: {
529       // Find the base pointer.
530       Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
531       StoreRefsForMemset[Ptr].push_back(SI);
532     } break;
533     case LegalStoreKind::MemsetPattern: {
534       // Find the base pointer.
535       Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
536       StoreRefsForMemsetPattern[Ptr].push_back(SI);
537     } break;
538     case LegalStoreKind::Memcpy:
539     case LegalStoreKind::UnorderedAtomicMemcpy:
540       StoreRefsForMemcpy.push_back(SI);
541       break;
542     default:
543       assert(false && "unhandled return value");
544       break;
545     }
546   }
547 }
548 
549 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
550 /// with the specified backedge count.  This block is known to be in the current
551 /// loop and not in any subloops.
552 bool LoopIdiomRecognize::runOnLoopBlock(
553     BasicBlock *BB, const SCEV *BECount,
554     SmallVectorImpl<BasicBlock *> &ExitBlocks) {
555   // We can only promote stores in this block if they are unconditionally
556   // executed in the loop.  For a block to be unconditionally executed, it has
557   // to dominate all the exit blocks of the loop.  Verify this now.
558   for (BasicBlock *ExitBlock : ExitBlocks)
559     if (!DT->dominates(BB, ExitBlock))
560       return false;
561 
562   bool MadeChange = false;
563   // Look for store instructions, which may be optimized to memset/memcpy.
564   collectStores(BB);
565 
566   // Look for a single store or sets of stores with a common base, which can be
567   // optimized into a memset (memset_pattern).  The latter most commonly happens
568   // with structs and handunrolled loops.
569   for (auto &SL : StoreRefsForMemset)
570     MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes);
571 
572   for (auto &SL : StoreRefsForMemsetPattern)
573     MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No);
574 
575   // Optimize the store into a memcpy, if it feeds an similarly strided load.
576   for (auto &SI : StoreRefsForMemcpy)
577     MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
578 
579   MadeChange |= processLoopMemIntrinsic<MemCpyInst>(
580       BB, &LoopIdiomRecognize::processLoopMemCpy, BECount);
581   MadeChange |= processLoopMemIntrinsic<MemSetInst>(
582       BB, &LoopIdiomRecognize::processLoopMemSet, BECount);
583 
584   return MadeChange;
585 }
586 
587 /// See if this store(s) can be promoted to a memset.
588 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
589                                            const SCEV *BECount, ForMemset For) {
590   // Try to find consecutive stores that can be transformed into memsets.
591   SetVector<StoreInst *> Heads, Tails;
592   SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
593 
594   // Do a quadratic search on all of the given stores and find
595   // all of the pairs of stores that follow each other.
596   SmallVector<unsigned, 16> IndexQueue;
597   for (unsigned i = 0, e = SL.size(); i < e; ++i) {
598     assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
599 
600     Value *FirstStoredVal = SL[i]->getValueOperand();
601     Value *FirstStorePtr = SL[i]->getPointerOperand();
602     const SCEVAddRecExpr *FirstStoreEv =
603         cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
604     APInt FirstStride = getStoreStride(FirstStoreEv);
605     unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType());
606 
607     // See if we can optimize just this store in isolation.
608     if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
609       Heads.insert(SL[i]);
610       continue;
611     }
612 
613     Value *FirstSplatValue = nullptr;
614     Constant *FirstPatternValue = nullptr;
615 
616     if (For == ForMemset::Yes)
617       FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL);
618     else
619       FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
620 
621     assert((FirstSplatValue || FirstPatternValue) &&
622            "Expected either splat value or pattern value.");
623 
624     IndexQueue.clear();
625     // If a store has multiple consecutive store candidates, search Stores
626     // array according to the sequence: from i+1 to e, then from i-1 to 0.
627     // This is because usually pairing with immediate succeeding or preceding
628     // candidate create the best chance to find memset opportunity.
629     unsigned j = 0;
630     for (j = i + 1; j < e; ++j)
631       IndexQueue.push_back(j);
632     for (j = i; j > 0; --j)
633       IndexQueue.push_back(j - 1);
634 
635     for (auto &k : IndexQueue) {
636       assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
637       Value *SecondStorePtr = SL[k]->getPointerOperand();
638       const SCEVAddRecExpr *SecondStoreEv =
639           cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
640       APInt SecondStride = getStoreStride(SecondStoreEv);
641 
642       if (FirstStride != SecondStride)
643         continue;
644 
645       Value *SecondStoredVal = SL[k]->getValueOperand();
646       Value *SecondSplatValue = nullptr;
647       Constant *SecondPatternValue = nullptr;
648 
649       if (For == ForMemset::Yes)
650         SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL);
651       else
652         SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
653 
654       assert((SecondSplatValue || SecondPatternValue) &&
655              "Expected either splat value or pattern value.");
656 
657       if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
658         if (For == ForMemset::Yes) {
659           if (isa<UndefValue>(FirstSplatValue))
660             FirstSplatValue = SecondSplatValue;
661           if (FirstSplatValue != SecondSplatValue)
662             continue;
663         } else {
664           if (isa<UndefValue>(FirstPatternValue))
665             FirstPatternValue = SecondPatternValue;
666           if (FirstPatternValue != SecondPatternValue)
667             continue;
668         }
669         Tails.insert(SL[k]);
670         Heads.insert(SL[i]);
671         ConsecutiveChain[SL[i]] = SL[k];
672         break;
673       }
674     }
675   }
676 
677   // We may run into multiple chains that merge into a single chain. We mark the
678   // stores that we transformed so that we don't visit the same store twice.
679   SmallPtrSet<Value *, 16> TransformedStores;
680   bool Changed = false;
681 
682   // For stores that start but don't end a link in the chain:
683   for (StoreInst *I : Heads) {
684     if (Tails.count(I))
685       continue;
686 
687     // We found a store instr that starts a chain. Now follow the chain and try
688     // to transform it.
689     SmallPtrSet<Instruction *, 8> AdjacentStores;
690     StoreInst *HeadStore = I;
691     unsigned StoreSize = 0;
692 
693     // Collect the chain into a list.
694     while (Tails.count(I) || Heads.count(I)) {
695       if (TransformedStores.count(I))
696         break;
697       AdjacentStores.insert(I);
698 
699       StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType());
700       // Move to the next value in the chain.
701       I = ConsecutiveChain[I];
702     }
703 
704     Value *StoredVal = HeadStore->getValueOperand();
705     Value *StorePtr = HeadStore->getPointerOperand();
706     const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
707     APInt Stride = getStoreStride(StoreEv);
708 
709     // Check to see if the stride matches the size of the stores.  If so, then
710     // we know that every byte is touched in the loop.
711     if (StoreSize != Stride && StoreSize != -Stride)
712       continue;
713 
714     bool IsNegStride = StoreSize == -Stride;
715 
716     Type *IntIdxTy = DL->getIndexType(StorePtr->getType());
717     const SCEV *StoreSizeSCEV = SE->getConstant(IntIdxTy, StoreSize);
718     if (processLoopStridedStore(StorePtr, StoreSizeSCEV,
719                                 MaybeAlign(HeadStore->getAlign()), StoredVal,
720                                 HeadStore, AdjacentStores, StoreEv, BECount,
721                                 IsNegStride)) {
722       TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
723       Changed = true;
724     }
725   }
726 
727   return Changed;
728 }
729 
730 /// processLoopMemIntrinsic - Template function for calling different processor
731 /// functions based on mem intrinsic type.
732 template <typename MemInst>
733 bool LoopIdiomRecognize::processLoopMemIntrinsic(
734     BasicBlock *BB,
735     bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
736     const SCEV *BECount) {
737   bool MadeChange = false;
738   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
739     Instruction *Inst = &*I++;
740     // Look for memory instructions, which may be optimized to a larger one.
741     if (MemInst *MI = dyn_cast<MemInst>(Inst)) {
742       WeakTrackingVH InstPtr(&*I);
743       if (!(this->*Processor)(MI, BECount))
744         continue;
745       MadeChange = true;
746 
747       // If processing the instruction invalidated our iterator, start over from
748       // the top of the block.
749       if (!InstPtr)
750         I = BB->begin();
751     }
752   }
753   return MadeChange;
754 }
755 
756 /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy
757 bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
758                                            const SCEV *BECount) {
759   // We can only handle non-volatile memcpys with a constant size.
760   if (MCI->isVolatile() || !isa<ConstantInt>(MCI->getLength()))
761     return false;
762 
763   // If we're not allowed to hack on memcpy, we fail.
764   if ((!HasMemcpy && !isa<MemCpyInlineInst>(MCI)) || DisableLIRP::Memcpy)
765     return false;
766 
767   Value *Dest = MCI->getDest();
768   Value *Source = MCI->getSource();
769   if (!Dest || !Source)
770     return false;
771 
772   // See if the load and store pointer expressions are AddRec like {base,+,1} on
773   // the current loop, which indicates a strided load and store.  If we have
774   // something else, it's a random load or store we can't handle.
775   const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest));
776   if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
777     return false;
778   const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source));
779   if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
780     return false;
781 
782   // Reject memcpys that are so large that they overflow an unsigned.
783   uint64_t SizeInBytes = cast<ConstantInt>(MCI->getLength())->getZExtValue();
784   if ((SizeInBytes >> 32) != 0)
785     return false;
786 
787   // Check if the stride matches the size of the memcpy. If so, then we know
788   // that every byte is touched in the loop.
789   const SCEVConstant *ConstStoreStride =
790       dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
791   const SCEVConstant *ConstLoadStride =
792       dyn_cast<SCEVConstant>(LoadEv->getOperand(1));
793   if (!ConstStoreStride || !ConstLoadStride)
794     return false;
795 
796   APInt StoreStrideValue = ConstStoreStride->getAPInt();
797   APInt LoadStrideValue = ConstLoadStride->getAPInt();
798   // Huge stride value - give up
799   if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64)
800     return false;
801 
802   if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) {
803     ORE.emit([&]() {
804       return OptimizationRemarkMissed(DEBUG_TYPE, "SizeStrideUnequal", MCI)
805              << ore::NV("Inst", "memcpy") << " in "
806              << ore::NV("Function", MCI->getFunction())
807              << " function will not be hoisted: "
808              << ore::NV("Reason", "memcpy size is not equal to stride");
809     });
810     return false;
811   }
812 
813   int64_t StoreStrideInt = StoreStrideValue.getSExtValue();
814   int64_t LoadStrideInt = LoadStrideValue.getSExtValue();
815   // Check if the load stride matches the store stride.
816   if (StoreStrideInt != LoadStrideInt)
817     return false;
818 
819   return processLoopStoreOfLoopLoad(
820       Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes),
821       MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv,
822       BECount);
823 }
824 
825 /// processLoopMemSet - See if this memset can be promoted to a large memset.
826 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
827                                            const SCEV *BECount) {
828   // We can only handle non-volatile memsets.
829   if (MSI->isVolatile())
830     return false;
831 
832   // If we're not allowed to hack on memset, we fail.
833   if (!HasMemset || DisableLIRP::Memset)
834     return false;
835 
836   Value *Pointer = MSI->getDest();
837 
838   // See if the pointer expression is an AddRec like {base,+,1} on the current
839   // loop, which indicates a strided store.  If we have something else, it's a
840   // random store we can't handle.
841   const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
842   if (!Ev || Ev->getLoop() != CurLoop)
843     return false;
844   if (!Ev->isAffine()) {
845     LLVM_DEBUG(dbgs() << "  Pointer is not affine, abort\n");
846     return false;
847   }
848 
849   const SCEV *PointerStrideSCEV = Ev->getOperand(1);
850   const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength());
851   if (!PointerStrideSCEV || !MemsetSizeSCEV)
852     return false;
853 
854   bool IsNegStride = false;
855   const bool IsConstantSize = isa<ConstantInt>(MSI->getLength());
856 
857   if (IsConstantSize) {
858     // Memset size is constant.
859     // Check if the pointer stride matches the memset size. If so, then
860     // we know that every byte is touched in the loop.
861     LLVM_DEBUG(dbgs() << "  memset size is constant\n");
862     uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
863     const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
864     if (!ConstStride)
865       return false;
866 
867     APInt Stride = ConstStride->getAPInt();
868     if (SizeInBytes != Stride && SizeInBytes != -Stride)
869       return false;
870 
871     IsNegStride = SizeInBytes == -Stride;
872   } else {
873     // Memset size is non-constant.
874     // Check if the pointer stride matches the memset size.
875     // To be conservative, the pass would not promote pointers that aren't in
876     // address space zero. Also, the pass only handles memset length and stride
877     // that are invariant for the top level loop.
878     LLVM_DEBUG(dbgs() << "  memset size is non-constant\n");
879     if (Pointer->getType()->getPointerAddressSpace() != 0) {
880       LLVM_DEBUG(dbgs() << "  pointer is not in address space zero, "
881                         << "abort\n");
882       return false;
883     }
884     if (!SE->isLoopInvariant(MemsetSizeSCEV, CurLoop)) {
885       LLVM_DEBUG(dbgs() << "  memset size is not a loop-invariant, "
886                         << "abort\n");
887       return false;
888     }
889 
890     // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV
891     IsNegStride = PointerStrideSCEV->isNonConstantNegative();
892     const SCEV *PositiveStrideSCEV =
893         IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV)
894                     : PointerStrideSCEV;
895     LLVM_DEBUG(dbgs() << "  MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"
896                       << "  PositiveStrideSCEV: " << *PositiveStrideSCEV
897                       << "\n");
898 
899     if (PositiveStrideSCEV != MemsetSizeSCEV) {
900       // If an expression is covered by the loop guard, compare again and
901       // proceed with optimization if equal.
902       const SCEV *FoldedPositiveStride =
903           SE->applyLoopGuards(PositiveStrideSCEV, CurLoop);
904       const SCEV *FoldedMemsetSize =
905           SE->applyLoopGuards(MemsetSizeSCEV, CurLoop);
906 
907       LLVM_DEBUG(dbgs() << "  Try to fold SCEV based on loop guard\n"
908                         << "    FoldedMemsetSize: " << *FoldedMemsetSize << "\n"
909                         << "    FoldedPositiveStride: " << *FoldedPositiveStride
910                         << "\n");
911 
912       if (FoldedPositiveStride != FoldedMemsetSize) {
913         LLVM_DEBUG(dbgs() << "  SCEV don't match, abort\n");
914         return false;
915       }
916     }
917   }
918 
919   // Verify that the memset value is loop invariant.  If not, we can't promote
920   // the memset.
921   Value *SplatValue = MSI->getValue();
922   if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
923     return false;
924 
925   SmallPtrSet<Instruction *, 1> MSIs;
926   MSIs.insert(MSI);
927   return processLoopStridedStore(Pointer, SE->getSCEV(MSI->getLength()),
928                                  MSI->getDestAlign(), SplatValue, MSI, MSIs, Ev,
929                                  BECount, IsNegStride, /*IsLoopMemset=*/true);
930 }
931 
932 /// mayLoopAccessLocation - Return true if the specified loop might access the
933 /// specified pointer location, which is a loop-strided access.  The 'Access'
934 /// argument specifies what the verboten forms of access are (read or write).
935 static bool
936 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
937                       const SCEV *BECount, const SCEV *StoreSizeSCEV,
938                       AliasAnalysis &AA,
939                       SmallPtrSetImpl<Instruction *> &IgnoredInsts) {
940   // Get the location that may be stored across the loop.  Since the access is
941   // strided positively through memory, we say that the modified location starts
942   // at the pointer and has infinite size.
943   LocationSize AccessSize = LocationSize::afterPointer();
944 
945   // If the loop iterates a fixed number of times, we can refine the access size
946   // to be exactly the size of the memset, which is (BECount+1)*StoreSize
947   const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount);
948   const SCEVConstant *ConstSize = dyn_cast<SCEVConstant>(StoreSizeSCEV);
949   if (BECst && ConstSize) {
950     std::optional<uint64_t> BEInt = BECst->getAPInt().tryZExtValue();
951     std::optional<uint64_t> SizeInt = ConstSize->getAPInt().tryZExtValue();
952     // FIXME: Should this check for overflow?
953     if (BEInt && SizeInt)
954       AccessSize = LocationSize::precise((*BEInt + 1) * *SizeInt);
955   }
956 
957   // TODO: For this to be really effective, we have to dive into the pointer
958   // operand in the store.  Store to &A[i] of 100 will always return may alias
959   // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
960   // which will then no-alias a store to &A[100].
961   MemoryLocation StoreLoc(Ptr, AccessSize);
962 
963   for (BasicBlock *B : L->blocks())
964     for (Instruction &I : *B)
965       if (!IgnoredInsts.contains(&I) &&
966           isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access))
967         return true;
968   return false;
969 }
970 
971 // If we have a negative stride, Start refers to the end of the memory location
972 // we're trying to memset.  Therefore, we need to recompute the base pointer,
973 // which is just Start - BECount*Size.
974 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
975                                         Type *IntPtr, const SCEV *StoreSizeSCEV,
976                                         ScalarEvolution *SE) {
977   const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
978   if (!StoreSizeSCEV->isOne()) {
979     // index = back edge count * store size
980     Index = SE->getMulExpr(Index,
981                            SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr),
982                            SCEV::FlagNUW);
983   }
984   // base pointer = start - index * store size
985   return SE->getMinusSCEV(Start, Index);
986 }
987 
988 /// Compute the number of bytes as a SCEV from the backedge taken count.
989 ///
990 /// This also maps the SCEV into the provided type and tries to handle the
991 /// computation in a way that will fold cleanly.
992 static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr,
993                                const SCEV *StoreSizeSCEV, Loop *CurLoop,
994                                const DataLayout *DL, ScalarEvolution *SE) {
995   const SCEV *TripCountSCEV =
996       SE->getTripCountFromExitCount(BECount, IntPtr, CurLoop);
997   return SE->getMulExpr(TripCountSCEV,
998                         SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr),
999                         SCEV::FlagNUW);
1000 }
1001 
1002 /// processLoopStridedStore - We see a strided store of some value.  If we can
1003 /// transform this into a memset or memset_pattern in the loop preheader, do so.
1004 bool LoopIdiomRecognize::processLoopStridedStore(
1005     Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment,
1006     Value *StoredVal, Instruction *TheStore,
1007     SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
1008     const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) {
1009   Module *M = TheStore->getModule();
1010   Value *SplatValue = isBytewiseValue(StoredVal, *DL);
1011   Constant *PatternValue = nullptr;
1012 
1013   if (!SplatValue)
1014     PatternValue = getMemSetPatternValue(StoredVal, DL);
1015 
1016   assert((SplatValue || PatternValue) &&
1017          "Expected either splat value or pattern value.");
1018 
1019   // The trip count of the loop and the base pointer of the addrec SCEV is
1020   // guaranteed to be loop invariant, which means that it should dominate the
1021   // header.  This allows us to insert code for it in the preheader.
1022   unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
1023   BasicBlock *Preheader = CurLoop->getLoopPreheader();
1024   IRBuilder<> Builder(Preheader->getTerminator());
1025   SCEVExpander Expander(*SE, *DL, "loop-idiom");
1026   SCEVExpanderCleaner ExpCleaner(Expander);
1027 
1028   Type *DestInt8PtrTy = Builder.getPtrTy(DestAS);
1029   Type *IntIdxTy = DL->getIndexType(DestPtr->getType());
1030 
1031   bool Changed = false;
1032   const SCEV *Start = Ev->getStart();
1033   // Handle negative strided loops.
1034   if (IsNegStride)
1035     Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSizeSCEV, SE);
1036 
1037   // TODO: ideally we should still be able to generate memset if SCEV expander
1038   // is taught to generate the dependencies at the latest point.
1039   if (!Expander.isSafeToExpand(Start))
1040     return Changed;
1041 
1042   // Okay, we have a strided store "p[i]" of a splattable value.  We can turn
1043   // this into a memset in the loop preheader now if we want.  However, this
1044   // would be unsafe to do if there is anything else in the loop that may read
1045   // or write to the aliased location.  Check for any overlap by generating the
1046   // base pointer and checking the region.
1047   Value *BasePtr =
1048       Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
1049 
1050   // From here on out, conservatively report to the pass manager that we've
1051   // changed the IR, even if we later clean up these added instructions. There
1052   // may be structural differences e.g. in the order of use lists not accounted
1053   // for in just a textual dump of the IR. This is written as a variable, even
1054   // though statically all the places this dominates could be replaced with
1055   // 'true', with the hope that anyone trying to be clever / "more precise" with
1056   // the return value will read this comment, and leave them alone.
1057   Changed = true;
1058 
1059   if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1060                             StoreSizeSCEV, *AA, Stores))
1061     return Changed;
1062 
1063   if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
1064     return Changed;
1065 
1066   // Okay, everything looks good, insert the memset.
1067 
1068   const SCEV *NumBytesS =
1069       getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1070 
1071   // TODO: ideally we should still be able to generate memset if SCEV expander
1072   // is taught to generate the dependencies at the latest point.
1073   if (!Expander.isSafeToExpand(NumBytesS))
1074     return Changed;
1075 
1076   Value *NumBytes =
1077       Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1078 
1079   if (!SplatValue && !isLibFuncEmittable(M, TLI, LibFunc_memset_pattern16))
1080     return Changed;
1081 
1082   AAMDNodes AATags = TheStore->getAAMetadata();
1083   for (Instruction *Store : Stores)
1084     AATags = AATags.merge(Store->getAAMetadata());
1085   if (auto CI = dyn_cast<ConstantInt>(NumBytes))
1086     AATags = AATags.extendTo(CI->getZExtValue());
1087   else
1088     AATags = AATags.extendTo(-1);
1089 
1090   CallInst *NewCall;
1091   if (SplatValue) {
1092     NewCall = Builder.CreateMemSet(
1093         BasePtr, SplatValue, NumBytes, MaybeAlign(StoreAlignment),
1094         /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias);
1095   } else {
1096     assert (isLibFuncEmittable(M, TLI, LibFunc_memset_pattern16));
1097     // Everything is emitted in default address space
1098     Type *Int8PtrTy = DestInt8PtrTy;
1099 
1100     StringRef FuncName = "memset_pattern16";
1101     FunctionCallee MSP = getOrInsertLibFunc(M, *TLI, LibFunc_memset_pattern16,
1102                             Builder.getVoidTy(), Int8PtrTy, Int8PtrTy, IntIdxTy);
1103     inferNonMandatoryLibFuncAttrs(M, FuncName, *TLI);
1104 
1105     // Otherwise we should form a memset_pattern16.  PatternValue is known to be
1106     // an constant array of 16-bytes.  Plop the value into a mergable global.
1107     GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1108                                             GlobalValue::PrivateLinkage,
1109                                             PatternValue, ".memset_pattern");
1110     GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
1111     GV->setAlignment(Align(16));
1112     Value *PatternPtr = GV;
1113     NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
1114 
1115     // Set the TBAA info if present.
1116     if (AATags.TBAA)
1117       NewCall->setMetadata(LLVMContext::MD_tbaa, AATags.TBAA);
1118 
1119     if (AATags.Scope)
1120       NewCall->setMetadata(LLVMContext::MD_alias_scope, AATags.Scope);
1121 
1122     if (AATags.NoAlias)
1123       NewCall->setMetadata(LLVMContext::MD_noalias, AATags.NoAlias);
1124   }
1125 
1126   NewCall->setDebugLoc(TheStore->getDebugLoc());
1127 
1128   if (MSSAU) {
1129     MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1130         NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1131     MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1132   }
1133 
1134   LLVM_DEBUG(dbgs() << "  Formed memset: " << *NewCall << "\n"
1135                     << "    from store to: " << *Ev << " at: " << *TheStore
1136                     << "\n");
1137 
1138   ORE.emit([&]() {
1139     OptimizationRemark R(DEBUG_TYPE, "ProcessLoopStridedStore",
1140                          NewCall->getDebugLoc(), Preheader);
1141     R << "Transformed loop-strided store in "
1142       << ore::NV("Function", TheStore->getFunction())
1143       << " function into a call to "
1144       << ore::NV("NewFunction", NewCall->getCalledFunction())
1145       << "() intrinsic";
1146     if (!Stores.empty())
1147       R << ore::setExtraArgs();
1148     for (auto *I : Stores) {
1149       R << ore::NV("FromBlock", I->getParent()->getName())
1150         << ore::NV("ToBlock", Preheader->getName());
1151     }
1152     return R;
1153   });
1154 
1155   // Okay, the memset has been formed.  Zap the original store and anything that
1156   // feeds into it.
1157   for (auto *I : Stores) {
1158     if (MSSAU)
1159       MSSAU->removeMemoryAccess(I, true);
1160     deleteDeadInstruction(I);
1161   }
1162   if (MSSAU && VerifyMemorySSA)
1163     MSSAU->getMemorySSA()->verifyMemorySSA();
1164   ++NumMemSet;
1165   ExpCleaner.markResultUsed();
1166   return true;
1167 }
1168 
1169 /// If the stored value is a strided load in the same loop with the same stride
1170 /// this may be transformable into a memcpy.  This kicks in for stuff like
1171 /// for (i) A[i] = B[i];
1172 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
1173                                                     const SCEV *BECount) {
1174   assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.");
1175 
1176   Value *StorePtr = SI->getPointerOperand();
1177   const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
1178   unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
1179 
1180   // The store must be feeding a non-volatile load.
1181   LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1182   assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.");
1183 
1184   // See if the pointer expression is an AddRec like {base,+,1} on the current
1185   // loop, which indicates a strided load.  If we have something else, it's a
1186   // random load we can't handle.
1187   Value *LoadPtr = LI->getPointerOperand();
1188   const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr));
1189 
1190   const SCEV *StoreSizeSCEV = SE->getConstant(StorePtr->getType(), StoreSize);
1191   return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSizeSCEV,
1192                                     SI->getAlign(), LI->getAlign(), SI, LI,
1193                                     StoreEv, LoadEv, BECount);
1194 }
1195 
1196 namespace {
1197 class MemmoveVerifier {
1198 public:
1199   explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr,
1200                            const DataLayout &DL)
1201       : DL(DL), BP1(llvm::GetPointerBaseWithConstantOffset(
1202                     LoadBasePtr.stripPointerCasts(), LoadOff, DL)),
1203         BP2(llvm::GetPointerBaseWithConstantOffset(
1204             StoreBasePtr.stripPointerCasts(), StoreOff, DL)),
1205         IsSameObject(BP1 == BP2) {}
1206 
1207   bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride,
1208                                   const Instruction &TheLoad,
1209                                   bool IsMemCpy) const {
1210     if (IsMemCpy) {
1211       // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1212       // for negative stride.
1213       if ((!IsNegStride && LoadOff <= StoreOff) ||
1214           (IsNegStride && LoadOff >= StoreOff))
1215         return false;
1216     } else {
1217       // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr
1218       // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr.
1219       int64_t LoadSize =
1220           DL.getTypeSizeInBits(TheLoad.getType()).getFixedValue() / 8;
1221       if (BP1 != BP2 || LoadSize != int64_t(StoreSize))
1222         return false;
1223       if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) ||
1224           (IsNegStride && LoadOff + LoadSize > StoreOff))
1225         return false;
1226     }
1227     return true;
1228   }
1229 
1230 private:
1231   const DataLayout &DL;
1232   int64_t LoadOff = 0;
1233   int64_t StoreOff = 0;
1234   const Value *BP1;
1235   const Value *BP2;
1236 
1237 public:
1238   const bool IsSameObject;
1239 };
1240 } // namespace
1241 
1242 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
1243     Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV,
1244     MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore,
1245     Instruction *TheLoad, const SCEVAddRecExpr *StoreEv,
1246     const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
1247 
1248   // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to
1249   // conservatively bail here, since otherwise we may have to transform
1250   // llvm.memcpy.inline into llvm.memcpy which is illegal.
1251   if (isa<MemCpyInlineInst>(TheStore))
1252     return false;
1253 
1254   // The trip count of the loop and the base pointer of the addrec SCEV is
1255   // guaranteed to be loop invariant, which means that it should dominate the
1256   // header.  This allows us to insert code for it in the preheader.
1257   BasicBlock *Preheader = CurLoop->getLoopPreheader();
1258   IRBuilder<> Builder(Preheader->getTerminator());
1259   SCEVExpander Expander(*SE, *DL, "loop-idiom");
1260 
1261   SCEVExpanderCleaner ExpCleaner(Expander);
1262 
1263   bool Changed = false;
1264   const SCEV *StrStart = StoreEv->getStart();
1265   unsigned StrAS = DestPtr->getType()->getPointerAddressSpace();
1266   Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS));
1267 
1268   APInt Stride = getStoreStride(StoreEv);
1269   const SCEVConstant *ConstStoreSize = dyn_cast<SCEVConstant>(StoreSizeSCEV);
1270 
1271   // TODO: Deal with non-constant size; Currently expect constant store size
1272   assert(ConstStoreSize && "store size is expected to be a constant");
1273 
1274   int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue();
1275   bool IsNegStride = StoreSize == -Stride;
1276 
1277   // Handle negative strided loops.
1278   if (IsNegStride)
1279     StrStart =
1280         getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSizeSCEV, SE);
1281 
1282   // Okay, we have a strided store "p[i]" of a loaded value.  We can turn
1283   // this into a memcpy in the loop preheader now if we want.  However, this
1284   // would be unsafe to do if there is anything else in the loop that may read
1285   // or write the memory region we're storing to.  This includes the load that
1286   // feeds the stores.  Check for an alias by generating the base address and
1287   // checking everything.
1288   Value *StoreBasePtr = Expander.expandCodeFor(
1289       StrStart, Builder.getPtrTy(StrAS), Preheader->getTerminator());
1290 
1291   // From here on out, conservatively report to the pass manager that we've
1292   // changed the IR, even if we later clean up these added instructions. There
1293   // may be structural differences e.g. in the order of use lists not accounted
1294   // for in just a textual dump of the IR. This is written as a variable, even
1295   // though statically all the places this dominates could be replaced with
1296   // 'true', with the hope that anyone trying to be clever / "more precise" with
1297   // the return value will read this comment, and leave them alone.
1298   Changed = true;
1299 
1300   SmallPtrSet<Instruction *, 2> IgnoredInsts;
1301   IgnoredInsts.insert(TheStore);
1302 
1303   bool IsMemCpy = isa<MemCpyInst>(TheStore);
1304   const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store";
1305 
1306   bool LoopAccessStore =
1307       mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1308                             StoreSizeSCEV, *AA, IgnoredInsts);
1309   if (LoopAccessStore) {
1310     // For memmove case it's not enough to guarantee that loop doesn't access
1311     // TheStore and TheLoad. Additionally we need to make sure that TheStore is
1312     // the only user of TheLoad.
1313     if (!TheLoad->hasOneUse())
1314       return Changed;
1315     IgnoredInsts.insert(TheLoad);
1316     if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
1317                               BECount, StoreSizeSCEV, *AA, IgnoredInsts)) {
1318       ORE.emit([&]() {
1319         return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessStore",
1320                                         TheStore)
1321                << ore::NV("Inst", InstRemark) << " in "
1322                << ore::NV("Function", TheStore->getFunction())
1323                << " function will not be hoisted: "
1324                << ore::NV("Reason", "The loop may access store location");
1325       });
1326       return Changed;
1327     }
1328     IgnoredInsts.erase(TheLoad);
1329   }
1330 
1331   const SCEV *LdStart = LoadEv->getStart();
1332   unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace();
1333 
1334   // Handle negative strided loops.
1335   if (IsNegStride)
1336     LdStart =
1337         getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSizeSCEV, SE);
1338 
1339   // For a memcpy, we have to make sure that the input array is not being
1340   // mutated by the loop.
1341   Value *LoadBasePtr = Expander.expandCodeFor(LdStart, Builder.getPtrTy(LdAS),
1342                                               Preheader->getTerminator());
1343 
1344   // If the store is a memcpy instruction, we must check if it will write to
1345   // the load memory locations. So remove it from the ignored stores.
1346   MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL);
1347   if (IsMemCpy && !Verifier.IsSameObject)
1348     IgnoredInsts.erase(TheStore);
1349   if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
1350                             StoreSizeSCEV, *AA, IgnoredInsts)) {
1351     ORE.emit([&]() {
1352       return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessLoad", TheLoad)
1353              << ore::NV("Inst", InstRemark) << " in "
1354              << ore::NV("Function", TheStore->getFunction())
1355              << " function will not be hoisted: "
1356              << ore::NV("Reason", "The loop may access load location");
1357     });
1358     return Changed;
1359   }
1360 
1361   bool IsAtomic = TheStore->isAtomic() || TheLoad->isAtomic();
1362   bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore;
1363 
1364   if (IsAtomic) {
1365     // For now don't support unordered atomic memmove.
1366     if (UseMemMove)
1367       return Changed;
1368 
1369     // We cannot allow unaligned ops for unordered load/store, so reject
1370     // anything where the alignment isn't at least the element size.
1371     assert((StoreAlign && LoadAlign) &&
1372            "Expect unordered load/store to have align.");
1373     if (*StoreAlign < StoreSize || *LoadAlign < StoreSize)
1374       return Changed;
1375 
1376     // If the element.atomic memcpy is not lowered into explicit
1377     // loads/stores later, then it will be lowered into an element-size
1378     // specific lib call. If the lib call doesn't exist for our store size, then
1379     // we shouldn't generate the memcpy.
1380     if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize())
1381       return Changed;
1382   }
1383 
1384   if (UseMemMove)
1385     if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, *TheLoad,
1386                                              IsMemCpy))
1387       return Changed;
1388 
1389   if (avoidLIRForMultiBlockLoop())
1390     return Changed;
1391 
1392   // Okay, everything is safe, we can transform this!
1393 
1394   const SCEV *NumBytesS =
1395       getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE);
1396 
1397   Value *NumBytes =
1398       Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1399 
1400   AAMDNodes AATags = TheLoad->getAAMetadata();
1401   AAMDNodes StoreAATags = TheStore->getAAMetadata();
1402   AATags = AATags.merge(StoreAATags);
1403   if (auto CI = dyn_cast<ConstantInt>(NumBytes))
1404     AATags = AATags.extendTo(CI->getZExtValue());
1405   else
1406     AATags = AATags.extendTo(-1);
1407 
1408   CallInst *NewCall = nullptr;
1409   // Check whether to generate an unordered atomic memcpy:
1410   //  If the load or store are atomic, then they must necessarily be unordered
1411   //  by previous checks.
1412   if (!IsAtomic) {
1413     if (UseMemMove)
1414       NewCall = Builder.CreateMemMove(
1415           StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes,
1416           /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias);
1417     else
1418       NewCall =
1419           Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign,
1420                                NumBytes, /*isVolatile=*/false, AATags.TBAA,
1421                                AATags.TBAAStruct, AATags.Scope, AATags.NoAlias);
1422   } else {
1423     // Create the call.
1424     // Note that unordered atomic loads/stores are *required* by the spec to
1425     // have an alignment but non-atomic loads/stores may not.
1426     NewCall = Builder.CreateElementUnorderedAtomicMemCpy(
1427         StoreBasePtr, *StoreAlign, LoadBasePtr, *LoadAlign, NumBytes, StoreSize,
1428         AATags.TBAA, AATags.TBAAStruct, AATags.Scope, AATags.NoAlias);
1429   }
1430   NewCall->setDebugLoc(TheStore->getDebugLoc());
1431 
1432   if (MSSAU) {
1433     MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1434         NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1435     MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1436   }
1437 
1438   LLVM_DEBUG(dbgs() << "  Formed new call: " << *NewCall << "\n"
1439                     << "    from load ptr=" << *LoadEv << " at: " << *TheLoad
1440                     << "\n"
1441                     << "    from store ptr=" << *StoreEv << " at: " << *TheStore
1442                     << "\n");
1443 
1444   ORE.emit([&]() {
1445     return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStoreOfLoopLoad",
1446                               NewCall->getDebugLoc(), Preheader)
1447            << "Formed a call to "
1448            << ore::NV("NewFunction", NewCall->getCalledFunction())
1449            << "() intrinsic from " << ore::NV("Inst", InstRemark)
1450            << " instruction in " << ore::NV("Function", TheStore->getFunction())
1451            << " function"
1452            << ore::setExtraArgs()
1453            << ore::NV("FromBlock", TheStore->getParent()->getName())
1454            << ore::NV("ToBlock", Preheader->getName());
1455   });
1456 
1457   // Okay, a new call to memcpy/memmove has been formed.  Zap the original store
1458   // and anything that feeds into it.
1459   if (MSSAU)
1460     MSSAU->removeMemoryAccess(TheStore, true);
1461   deleteDeadInstruction(TheStore);
1462   if (MSSAU && VerifyMemorySSA)
1463     MSSAU->getMemorySSA()->verifyMemorySSA();
1464   if (UseMemMove)
1465     ++NumMemMove;
1466   else
1467     ++NumMemCpy;
1468   ExpCleaner.markResultUsed();
1469   return true;
1470 }
1471 
1472 // When compiling for codesize we avoid idiom recognition for a multi-block loop
1473 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
1474 //
1475 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
1476                                                    bool IsLoopMemset) {
1477   if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
1478     if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) {
1479       LLVM_DEBUG(dbgs() << "  " << CurLoop->getHeader()->getParent()->getName()
1480                         << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1481                         << " avoided: multi-block top-level loop\n");
1482       return true;
1483     }
1484   }
1485 
1486   return false;
1487 }
1488 
1489 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1490   LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
1491                     << CurLoop->getHeader()->getParent()->getName()
1492                     << "] Noncountable Loop %"
1493                     << CurLoop->getHeader()->getName() << "\n");
1494 
1495   return recognizePopcount() || recognizeAndInsertFFS() ||
1496          recognizeShiftUntilBitTest() || recognizeShiftUntilZero() ||
1497          recognizeShiftUntilLessThan();
1498 }
1499 
1500 /// Check if the given conditional branch is based on the comparison between
1501 /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is
1502 /// true), the control yields to the loop entry. If the branch matches the
1503 /// behavior, the variable involved in the comparison is returned. This function
1504 /// will be called to see if the precondition and postcondition of the loop are
1505 /// in desirable form.
1506 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
1507                              bool JmpOnZero = false) {
1508   if (!BI || !BI->isConditional())
1509     return nullptr;
1510 
1511   ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1512   if (!Cond)
1513     return nullptr;
1514 
1515   ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1516   if (!CmpZero || !CmpZero->isZero())
1517     return nullptr;
1518 
1519   BasicBlock *TrueSucc = BI->getSuccessor(0);
1520   BasicBlock *FalseSucc = BI->getSuccessor(1);
1521   if (JmpOnZero)
1522     std::swap(TrueSucc, FalseSucc);
1523 
1524   ICmpInst::Predicate Pred = Cond->getPredicate();
1525   if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) ||
1526       (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry))
1527     return Cond->getOperand(0);
1528 
1529   return nullptr;
1530 }
1531 
1532 /// Check if the given conditional branch is based on an unsigned less-than
1533 /// comparison between a variable and a constant, and if the comparison is false
1534 /// the control yields to the loop entry. If the branch matches the behaviour,
1535 /// the variable involved in the comparison is returned.
1536 static Value *matchShiftULTCondition(BranchInst *BI, BasicBlock *LoopEntry,
1537                                      APInt &Threshold) {
1538   if (!BI || !BI->isConditional())
1539     return nullptr;
1540 
1541   ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1542   if (!Cond)
1543     return nullptr;
1544 
1545   ConstantInt *CmpConst = dyn_cast<ConstantInt>(Cond->getOperand(1));
1546   if (!CmpConst)
1547     return nullptr;
1548 
1549   BasicBlock *FalseSucc = BI->getSuccessor(1);
1550   ICmpInst::Predicate Pred = Cond->getPredicate();
1551 
1552   if (Pred == ICmpInst::ICMP_ULT && FalseSucc == LoopEntry) {
1553     Threshold = CmpConst->getValue();
1554     return Cond->getOperand(0);
1555   }
1556 
1557   return nullptr;
1558 }
1559 
1560 // Check if the recurrence variable `VarX` is in the right form to create
1561 // the idiom. Returns the value coerced to a PHINode if so.
1562 static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
1563                                  BasicBlock *LoopEntry) {
1564   auto *PhiX = dyn_cast<PHINode>(VarX);
1565   if (PhiX && PhiX->getParent() == LoopEntry &&
1566       (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX))
1567     return PhiX;
1568   return nullptr;
1569 }
1570 
1571 /// Return true if the idiom is detected in the loop.
1572 ///
1573 /// Additionally:
1574 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1575 ///       or nullptr if there is no such.
1576 /// 2) \p CntPhi is set to the corresponding phi node
1577 ///       or nullptr if there is no such.
1578 /// 3) \p InitX is set to the value whose CTLZ could be used.
1579 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1580 /// 5) \p Threshold is set to the constant involved in the unsigned less-than
1581 ///       comparison.
1582 ///
1583 /// The core idiom we are trying to detect is:
1584 /// \code
1585 ///    if (x0 < 2)
1586 ///      goto loop-exit // the precondition of the loop
1587 ///    cnt0 = init-val
1588 ///    do {
1589 ///      x = phi (x0, x.next);   //PhiX
1590 ///      cnt = phi (cnt0, cnt.next)
1591 ///
1592 ///      cnt.next = cnt + 1;
1593 ///       ...
1594 ///      x.next = x >> 1;   // DefX
1595 ///    } while (x >= 4)
1596 /// loop-exit:
1597 /// \endcode
1598 static bool detectShiftUntilLessThanIdiom(Loop *CurLoop, const DataLayout &DL,
1599                                           Intrinsic::ID &IntrinID,
1600                                           Value *&InitX, Instruction *&CntInst,
1601                                           PHINode *&CntPhi, Instruction *&DefX,
1602                                           APInt &Threshold) {
1603   BasicBlock *LoopEntry;
1604 
1605   DefX = nullptr;
1606   CntInst = nullptr;
1607   CntPhi = nullptr;
1608   LoopEntry = *(CurLoop->block_begin());
1609 
1610   // step 1: Check if the loop-back branch is in desirable form.
1611   if (Value *T = matchShiftULTCondition(
1612           dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry,
1613           Threshold))
1614     DefX = dyn_cast<Instruction>(T);
1615   else
1616     return false;
1617 
1618   // step 2: Check the recurrence of variable X
1619   if (!DefX || !isa<PHINode>(DefX))
1620     return false;
1621 
1622   PHINode *VarPhi = cast<PHINode>(DefX);
1623   int Idx = VarPhi->getBasicBlockIndex(LoopEntry);
1624   if (Idx == -1)
1625     return false;
1626 
1627   DefX = dyn_cast<Instruction>(VarPhi->getIncomingValue(Idx));
1628   if (!DefX || DefX->getNumOperands() == 0 || DefX->getOperand(0) != VarPhi)
1629     return false;
1630 
1631   // step 3: detect instructions corresponding to "x.next = x >> 1"
1632   if (DefX->getOpcode() != Instruction::LShr)
1633     return false;
1634 
1635   IntrinID = Intrinsic::ctlz;
1636   ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1));
1637   if (!Shft || !Shft->isOne())
1638     return false;
1639 
1640   InitX = VarPhi->getIncomingValueForBlock(CurLoop->getLoopPreheader());
1641 
1642   // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1643   //         or cnt.next = cnt + -1.
1644   // TODO: We can skip the step. If loop trip count is known (CTLZ),
1645   //       then all uses of "cnt.next" could be optimized to the trip count
1646   //       plus "cnt0". Currently it is not optimized.
1647   //       This step could be used to detect POPCNT instruction:
1648   //       cnt.next = cnt + (x.next & 1)
1649   for (Instruction &Inst :
1650        llvm::make_range(LoopEntry->getFirstNonPHIIt(), LoopEntry->end())) {
1651     if (Inst.getOpcode() != Instruction::Add)
1652       continue;
1653 
1654     ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1));
1655     if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
1656       continue;
1657 
1658     PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry);
1659     if (!Phi)
1660       continue;
1661 
1662     CntInst = &Inst;
1663     CntPhi = Phi;
1664     break;
1665   }
1666   if (!CntInst)
1667     return false;
1668 
1669   return true;
1670 }
1671 
1672 /// Return true iff the idiom is detected in the loop.
1673 ///
1674 /// Additionally:
1675 /// 1) \p CntInst is set to the instruction counting the population bit.
1676 /// 2) \p CntPhi is set to the corresponding phi node.
1677 /// 3) \p Var is set to the value whose population bits are being counted.
1678 ///
1679 /// The core idiom we are trying to detect is:
1680 /// \code
1681 ///    if (x0 != 0)
1682 ///      goto loop-exit // the precondition of the loop
1683 ///    cnt0 = init-val;
1684 ///    do {
1685 ///       x1 = phi (x0, x2);
1686 ///       cnt1 = phi(cnt0, cnt2);
1687 ///
1688 ///       cnt2 = cnt1 + 1;
1689 ///        ...
1690 ///       x2 = x1 & (x1 - 1);
1691 ///        ...
1692 ///    } while(x != 0);
1693 ///
1694 /// loop-exit:
1695 /// \endcode
1696 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1697                                 Instruction *&CntInst, PHINode *&CntPhi,
1698                                 Value *&Var) {
1699   // step 1: Check to see if the look-back branch match this pattern:
1700   //    "if (a!=0) goto loop-entry".
1701   BasicBlock *LoopEntry;
1702   Instruction *DefX2, *CountInst;
1703   Value *VarX1, *VarX0;
1704   PHINode *PhiX, *CountPhi;
1705 
1706   DefX2 = CountInst = nullptr;
1707   VarX1 = VarX0 = nullptr;
1708   PhiX = CountPhi = nullptr;
1709   LoopEntry = *(CurLoop->block_begin());
1710 
1711   // step 1: Check if the loop-back branch is in desirable form.
1712   {
1713     if (Value *T = matchCondition(
1714             dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1715       DefX2 = dyn_cast<Instruction>(T);
1716     else
1717       return false;
1718   }
1719 
1720   // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1721   {
1722     if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1723       return false;
1724 
1725     BinaryOperator *SubOneOp;
1726 
1727     if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1728       VarX1 = DefX2->getOperand(1);
1729     else {
1730       VarX1 = DefX2->getOperand(0);
1731       SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1732     }
1733     if (!SubOneOp || SubOneOp->getOperand(0) != VarX1)
1734       return false;
1735 
1736     ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1));
1737     if (!Dec ||
1738         !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1739           (SubOneOp->getOpcode() == Instruction::Add &&
1740            Dec->isMinusOne()))) {
1741       return false;
1742     }
1743   }
1744 
1745   // step 3: Check the recurrence of variable X
1746   PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry);
1747   if (!PhiX)
1748     return false;
1749 
1750   // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1751   {
1752     CountInst = nullptr;
1753     for (Instruction &Inst :
1754          llvm::make_range(LoopEntry->getFirstNonPHIIt(), LoopEntry->end())) {
1755       if (Inst.getOpcode() != Instruction::Add)
1756         continue;
1757 
1758       ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1));
1759       if (!Inc || !Inc->isOne())
1760         continue;
1761 
1762       PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry);
1763       if (!Phi)
1764         continue;
1765 
1766       // Check if the result of the instruction is live of the loop.
1767       bool LiveOutLoop = false;
1768       for (User *U : Inst.users()) {
1769         if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1770           LiveOutLoop = true;
1771           break;
1772         }
1773       }
1774 
1775       if (LiveOutLoop) {
1776         CountInst = &Inst;
1777         CountPhi = Phi;
1778         break;
1779       }
1780     }
1781 
1782     if (!CountInst)
1783       return false;
1784   }
1785 
1786   // step 5: check if the precondition is in this form:
1787   //   "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1788   {
1789     auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1790     Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1791     if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1792       return false;
1793 
1794     CntInst = CountInst;
1795     CntPhi = CountPhi;
1796     Var = T;
1797   }
1798 
1799   return true;
1800 }
1801 
1802 /// Return true if the idiom is detected in the loop.
1803 ///
1804 /// Additionally:
1805 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1806 ///       or nullptr if there is no such.
1807 /// 2) \p CntPhi is set to the corresponding phi node
1808 ///       or nullptr if there is no such.
1809 /// 3) \p Var is set to the value whose CTLZ could be used.
1810 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1811 ///
1812 /// The core idiom we are trying to detect is:
1813 /// \code
1814 ///    if (x0 == 0)
1815 ///      goto loop-exit // the precondition of the loop
1816 ///    cnt0 = init-val;
1817 ///    do {
1818 ///       x = phi (x0, x.next);   //PhiX
1819 ///       cnt = phi(cnt0, cnt.next);
1820 ///
1821 ///       cnt.next = cnt + 1;
1822 ///        ...
1823 ///       x.next = x >> 1;   // DefX
1824 ///        ...
1825 ///    } while(x.next != 0);
1826 ///
1827 /// loop-exit:
1828 /// \endcode
1829 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL,
1830                                       Intrinsic::ID &IntrinID, Value *&InitX,
1831                                       Instruction *&CntInst, PHINode *&CntPhi,
1832                                       Instruction *&DefX) {
1833   BasicBlock *LoopEntry;
1834   Value *VarX = nullptr;
1835 
1836   DefX = nullptr;
1837   CntInst = nullptr;
1838   CntPhi = nullptr;
1839   LoopEntry = *(CurLoop->block_begin());
1840 
1841   // step 1: Check if the loop-back branch is in desirable form.
1842   if (Value *T = matchCondition(
1843           dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1844     DefX = dyn_cast<Instruction>(T);
1845   else
1846     return false;
1847 
1848   // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1"
1849   if (!DefX || !DefX->isShift())
1850     return false;
1851   IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz :
1852                                                      Intrinsic::ctlz;
1853   ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1));
1854   if (!Shft || !Shft->isOne())
1855     return false;
1856   VarX = DefX->getOperand(0);
1857 
1858   // step 3: Check the recurrence of variable X
1859   PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
1860   if (!PhiX)
1861     return false;
1862 
1863   InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader());
1864 
1865   // Make sure the initial value can't be negative otherwise the ashr in the
1866   // loop might never reach zero which would make the loop infinite.
1867   if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL))
1868     return false;
1869 
1870   // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1871   //         or cnt.next = cnt + -1.
1872   // TODO: We can skip the step. If loop trip count is known (CTLZ),
1873   //       then all uses of "cnt.next" could be optimized to the trip count
1874   //       plus "cnt0". Currently it is not optimized.
1875   //       This step could be used to detect POPCNT instruction:
1876   //       cnt.next = cnt + (x.next & 1)
1877   for (Instruction &Inst :
1878        llvm::make_range(LoopEntry->getFirstNonPHIIt(), LoopEntry->end())) {
1879     if (Inst.getOpcode() != Instruction::Add)
1880       continue;
1881 
1882     ConstantInt *Inc = dyn_cast<ConstantInt>(Inst.getOperand(1));
1883     if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
1884       continue;
1885 
1886     PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry);
1887     if (!Phi)
1888       continue;
1889 
1890     CntInst = &Inst;
1891     CntPhi = Phi;
1892     break;
1893   }
1894   if (!CntInst)
1895     return false;
1896 
1897   return true;
1898 }
1899 
1900 // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always
1901 // profitable if we delete the loop.
1902 bool LoopIdiomRecognize::isProfitableToInsertFFS(Intrinsic::ID IntrinID,
1903                                                  Value *InitX, bool ZeroCheck,
1904                                                  size_t CanonicalSize) {
1905   const Value *Args[] = {InitX,
1906                          ConstantInt::getBool(InitX->getContext(), ZeroCheck)};
1907 
1908   // @llvm.dbg doesn't count as they have no semantic effect.
1909   auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug();
1910   uint32_t HeaderSize =
1911       std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end());
1912 
1913   IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args);
1914   InstructionCost Cost = TTI->getIntrinsicInstrCost(
1915       Attrs, TargetTransformInfo::TCK_SizeAndLatency);
1916   if (HeaderSize != CanonicalSize && Cost > TargetTransformInfo::TCC_Basic)
1917     return false;
1918 
1919   return true;
1920 }
1921 
1922 /// Convert CTLZ / CTTZ idiom loop into countable loop.
1923 /// If CTLZ / CTTZ inserted as a new trip count returns true; otherwise,
1924 /// returns false.
1925 bool LoopIdiomRecognize::insertFFSIfProfitable(Intrinsic::ID IntrinID,
1926                                                Value *InitX, Instruction *DefX,
1927                                                PHINode *CntPhi,
1928                                                Instruction *CntInst) {
1929   bool IsCntPhiUsedOutsideLoop = false;
1930   for (User *U : CntPhi->users())
1931     if (!CurLoop->contains(cast<Instruction>(U))) {
1932       IsCntPhiUsedOutsideLoop = true;
1933       break;
1934     }
1935   bool IsCntInstUsedOutsideLoop = false;
1936   for (User *U : CntInst->users())
1937     if (!CurLoop->contains(cast<Instruction>(U))) {
1938       IsCntInstUsedOutsideLoop = true;
1939       break;
1940     }
1941   // If both CntInst and CntPhi are used outside the loop the profitability
1942   // is questionable.
1943   if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1944     return false;
1945 
1946   // For some CPUs result of CTLZ(X) intrinsic is undefined
1947   // when X is 0. If we can not guarantee X != 0, we need to check this
1948   // when expand.
1949   bool ZeroCheck = false;
1950   // It is safe to assume Preheader exist as it was checked in
1951   // parent function RunOnLoop.
1952   BasicBlock *PH = CurLoop->getLoopPreheader();
1953 
1954   // If we are using the count instruction outside the loop, make sure we
1955   // have a zero check as a precondition. Without the check the loop would run
1956   // one iteration for before any check of the input value. This means 0 and 1
1957   // would have identical behavior in the original loop and thus
1958   if (!IsCntPhiUsedOutsideLoop) {
1959     auto *PreCondBB = PH->getSinglePredecessor();
1960     if (!PreCondBB)
1961       return false;
1962     auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1963     if (!PreCondBI)
1964       return false;
1965     if (matchCondition(PreCondBI, PH) != InitX)
1966       return false;
1967     ZeroCheck = true;
1968   }
1969 
1970   // FFS idiom loop has only 6 instructions:
1971   //  %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1972   //  %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1973   //  %shr = ashr %n.addr.0, 1
1974   //  %tobool = icmp eq %shr, 0
1975   //  %inc = add nsw %i.0, 1
1976   //  br i1 %tobool
1977   size_t IdiomCanonicalSize = 6;
1978   if (!isProfitableToInsertFFS(IntrinID, InitX, ZeroCheck, IdiomCanonicalSize))
1979     return false;
1980 
1981   transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX,
1982                            DefX->getDebugLoc(), ZeroCheck,
1983                            IsCntPhiUsedOutsideLoop);
1984   return true;
1985 }
1986 
1987 /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop
1988 /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new
1989 /// trip count returns true; otherwise, returns false.
1990 bool LoopIdiomRecognize::recognizeAndInsertFFS() {
1991   // Give up if the loop has multiple blocks or multiple backedges.
1992   if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1993     return false;
1994 
1995   Intrinsic::ID IntrinID;
1996   Value *InitX;
1997   Instruction *DefX = nullptr;
1998   PHINode *CntPhi = nullptr;
1999   Instruction *CntInst = nullptr;
2000 
2001   if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, CntInst, CntPhi,
2002                                  DefX))
2003     return false;
2004 
2005   return insertFFSIfProfitable(IntrinID, InitX, DefX, CntPhi, CntInst);
2006 }
2007 
2008 bool LoopIdiomRecognize::recognizeShiftUntilLessThan() {
2009   // Give up if the loop has multiple blocks or multiple backedges.
2010   if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
2011     return false;
2012 
2013   Intrinsic::ID IntrinID;
2014   Value *InitX;
2015   Instruction *DefX = nullptr;
2016   PHINode *CntPhi = nullptr;
2017   Instruction *CntInst = nullptr;
2018 
2019   APInt LoopThreshold;
2020   if (!detectShiftUntilLessThanIdiom(CurLoop, *DL, IntrinID, InitX, CntInst,
2021                                      CntPhi, DefX, LoopThreshold))
2022     return false;
2023 
2024   if (LoopThreshold == 2) {
2025     // Treat as regular FFS.
2026     return insertFFSIfProfitable(IntrinID, InitX, DefX, CntPhi, CntInst);
2027   }
2028 
2029   // Look for Floor Log2 Idiom.
2030   if (LoopThreshold != 4)
2031     return false;
2032 
2033   // Abort if CntPhi is used outside of the loop.
2034   for (User *U : CntPhi->users())
2035     if (!CurLoop->contains(cast<Instruction>(U)))
2036       return false;
2037 
2038   // It is safe to assume Preheader exist as it was checked in
2039   // parent function RunOnLoop.
2040   BasicBlock *PH = CurLoop->getLoopPreheader();
2041   auto *PreCondBB = PH->getSinglePredecessor();
2042   if (!PreCondBB)
2043     return false;
2044   auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
2045   if (!PreCondBI)
2046     return false;
2047 
2048   APInt PreLoopThreshold;
2049   if (matchShiftULTCondition(PreCondBI, PH, PreLoopThreshold) != InitX ||
2050       PreLoopThreshold != 2)
2051     return false;
2052 
2053   bool ZeroCheck = true;
2054 
2055   // the loop has only 6 instructions:
2056   //  %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
2057   //  %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
2058   //  %shr = ashr %n.addr.0, 1
2059   //  %tobool = icmp ult %n.addr.0, C
2060   //  %inc = add nsw %i.0, 1
2061   //  br i1 %tobool
2062   size_t IdiomCanonicalSize = 6;
2063   if (!isProfitableToInsertFFS(IntrinID, InitX, ZeroCheck, IdiomCanonicalSize))
2064     return false;
2065 
2066   // log2(x) = w − 1 − clz(x)
2067   transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX,
2068                            DefX->getDebugLoc(), ZeroCheck,
2069                            /*IsCntPhiUsedOutsideLoop=*/false,
2070                            /*InsertSub=*/true);
2071   return true;
2072 }
2073 
2074 /// Recognizes a population count idiom in a non-countable loop.
2075 ///
2076 /// If detected, transforms the relevant code to issue the popcount intrinsic
2077 /// function call, and returns true; otherwise, returns false.
2078 bool LoopIdiomRecognize::recognizePopcount() {
2079   if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
2080     return false;
2081 
2082   // Counting population are usually conducted by few arithmetic instructions.
2083   // Such instructions can be easily "absorbed" by vacant slots in a
2084   // non-compact loop. Therefore, recognizing popcount idiom only makes sense
2085   // in a compact loop.
2086 
2087   // Give up if the loop has multiple blocks or multiple backedges.
2088   if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
2089     return false;
2090 
2091   BasicBlock *LoopBody = *(CurLoop->block_begin());
2092   if (LoopBody->size() >= 20) {
2093     // The loop is too big, bail out.
2094     return false;
2095   }
2096 
2097   // It should have a preheader containing nothing but an unconditional branch.
2098   BasicBlock *PH = CurLoop->getLoopPreheader();
2099   if (!PH || &PH->front() != PH->getTerminator())
2100     return false;
2101   auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
2102   if (!EntryBI || EntryBI->isConditional())
2103     return false;
2104 
2105   // It should have a precondition block where the generated popcount intrinsic
2106   // function can be inserted.
2107   auto *PreCondBB = PH->getSinglePredecessor();
2108   if (!PreCondBB)
2109     return false;
2110   auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
2111   if (!PreCondBI || PreCondBI->isUnconditional())
2112     return false;
2113 
2114   Instruction *CntInst;
2115   PHINode *CntPhi;
2116   Value *Val;
2117   if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
2118     return false;
2119 
2120   transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
2121   return true;
2122 }
2123 
2124 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
2125                                        const DebugLoc &DL) {
2126   Value *Ops[] = {Val};
2127   Type *Tys[] = {Val->getType()};
2128 
2129   CallInst *CI = IRBuilder.CreateIntrinsic(Intrinsic::ctpop, Tys, Ops);
2130   CI->setDebugLoc(DL);
2131 
2132   return CI;
2133 }
2134 
2135 static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
2136                                     const DebugLoc &DL, bool ZeroCheck,
2137                                     Intrinsic::ID IID) {
2138   Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)};
2139   Type *Tys[] = {Val->getType()};
2140 
2141   CallInst *CI = IRBuilder.CreateIntrinsic(IID, Tys, Ops);
2142   CI->setDebugLoc(DL);
2143 
2144   return CI;
2145 }
2146 
2147 /// Transform the following loop (Using CTLZ, CTTZ is similar):
2148 /// loop:
2149 ///   CntPhi = PHI [Cnt0, CntInst]
2150 ///   PhiX = PHI [InitX, DefX]
2151 ///   CntInst = CntPhi + 1
2152 ///   DefX = PhiX >> 1
2153 ///   LOOP_BODY
2154 ///   Br: loop if (DefX != 0)
2155 /// Use(CntPhi) or Use(CntInst)
2156 ///
2157 /// Into:
2158 /// If CntPhi used outside the loop:
2159 ///   CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
2160 ///   Count = CountPrev + 1
2161 /// else
2162 ///   Count = BitWidth(InitX) - CTLZ(InitX)
2163 /// loop:
2164 ///   CntPhi = PHI [Cnt0, CntInst]
2165 ///   PhiX = PHI [InitX, DefX]
2166 ///   PhiCount = PHI [Count, Dec]
2167 ///   CntInst = CntPhi + 1
2168 ///   DefX = PhiX >> 1
2169 ///   Dec = PhiCount - 1
2170 ///   LOOP_BODY
2171 ///   Br: loop if (Dec != 0)
2172 /// Use(CountPrev + Cnt0) // Use(CntPhi)
2173 /// or
2174 /// Use(Count + Cnt0) // Use(CntInst)
2175 ///
2176 /// If LOOP_BODY is empty the loop will be deleted.
2177 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
2178 void LoopIdiomRecognize::transformLoopToCountable(
2179     Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst,
2180     PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL,
2181     bool ZeroCheck, bool IsCntPhiUsedOutsideLoop, bool InsertSub) {
2182   BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator());
2183 
2184   // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block
2185   IRBuilder<> Builder(PreheaderBr);
2186   Builder.SetCurrentDebugLocation(DL);
2187 
2188   // If there are no uses of CntPhi crate:
2189   //   Count = BitWidth - CTLZ(InitX);
2190   //   NewCount = Count;
2191   // If there are uses of CntPhi create:
2192   //   NewCount = BitWidth - CTLZ(InitX >> 1);
2193   //   Count = NewCount + 1;
2194   Value *InitXNext;
2195   if (IsCntPhiUsedOutsideLoop) {
2196     if (DefX->getOpcode() == Instruction::AShr)
2197       InitXNext = Builder.CreateAShr(InitX, 1);
2198     else if (DefX->getOpcode() == Instruction::LShr)
2199       InitXNext = Builder.CreateLShr(InitX, 1);
2200     else if (DefX->getOpcode() == Instruction::Shl) // cttz
2201       InitXNext = Builder.CreateShl(InitX, 1);
2202     else
2203       llvm_unreachable("Unexpected opcode!");
2204   } else
2205     InitXNext = InitX;
2206   Value *Count =
2207       createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID);
2208   Type *CountTy = Count->getType();
2209   Count = Builder.CreateSub(
2210       ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count);
2211   if (InsertSub)
2212     Count = Builder.CreateSub(Count, ConstantInt::get(CountTy, 1));
2213   Value *NewCount = Count;
2214   if (IsCntPhiUsedOutsideLoop)
2215     Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1));
2216 
2217   NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType());
2218 
2219   Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
2220   if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) {
2221     // If the counter was being incremented in the loop, add NewCount to the
2222     // counter's initial value, but only if the initial value is not zero.
2223     ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
2224     if (!InitConst || !InitConst->isZero())
2225       NewCount = Builder.CreateAdd(NewCount, CntInitVal);
2226   } else {
2227     // If the count was being decremented in the loop, subtract NewCount from
2228     // the counter's initial value.
2229     NewCount = Builder.CreateSub(CntInitVal, NewCount);
2230   }
2231 
2232   // Step 2: Insert new IV and loop condition:
2233   // loop:
2234   //   ...
2235   //   PhiCount = PHI [Count, Dec]
2236   //   ...
2237   //   Dec = PhiCount - 1
2238   //   ...
2239   //   Br: loop if (Dec != 0)
2240   BasicBlock *Body = *(CurLoop->block_begin());
2241   auto *LbBr = cast<BranchInst>(Body->getTerminator());
2242   ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
2243 
2244   PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi");
2245   TcPhi->insertBefore(Body->begin());
2246 
2247   Builder.SetInsertPoint(LbCond);
2248   Instruction *TcDec = cast<Instruction>(Builder.CreateSub(
2249       TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true));
2250 
2251   TcPhi->addIncoming(Count, Preheader);
2252   TcPhi->addIncoming(TcDec, Body);
2253 
2254   CmpInst::Predicate Pred =
2255       (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
2256   LbCond->setPredicate(Pred);
2257   LbCond->setOperand(0, TcDec);
2258   LbCond->setOperand(1, ConstantInt::get(CountTy, 0));
2259 
2260   // Step 3: All the references to the original counter outside
2261   //  the loop are replaced with the NewCount
2262   if (IsCntPhiUsedOutsideLoop)
2263     CntPhi->replaceUsesOutsideBlock(NewCount, Body);
2264   else
2265     CntInst->replaceUsesOutsideBlock(NewCount, Body);
2266 
2267   // step 4: Forget the "non-computable" trip-count SCEV associated with the
2268   //   loop. The loop would otherwise not be deleted even if it becomes empty.
2269   SE->forgetLoop(CurLoop);
2270 }
2271 
2272 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
2273                                                  Instruction *CntInst,
2274                                                  PHINode *CntPhi, Value *Var) {
2275   BasicBlock *PreHead = CurLoop->getLoopPreheader();
2276   auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator());
2277   const DebugLoc &DL = CntInst->getDebugLoc();
2278 
2279   // Assuming before transformation, the loop is following:
2280   //  if (x) // the precondition
2281   //     do { cnt++; x &= x - 1; } while(x);
2282 
2283   // Step 1: Insert the ctpop instruction at the end of the precondition block
2284   IRBuilder<> Builder(PreCondBr);
2285   Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
2286   {
2287     PopCnt = createPopcntIntrinsic(Builder, Var, DL);
2288     NewCount = PopCntZext =
2289         Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
2290 
2291     if (NewCount != PopCnt)
2292       (cast<Instruction>(NewCount))->setDebugLoc(DL);
2293 
2294     // TripCnt is exactly the number of iterations the loop has
2295     TripCnt = NewCount;
2296 
2297     // If the population counter's initial value is not zero, insert Add Inst.
2298     Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
2299     ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
2300     if (!InitConst || !InitConst->isZero()) {
2301       NewCount = Builder.CreateAdd(NewCount, CntInitVal);
2302       (cast<Instruction>(NewCount))->setDebugLoc(DL);
2303     }
2304   }
2305 
2306   // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
2307   //   "if (NewCount == 0) loop-exit". Without this change, the intrinsic
2308   //   function would be partial dead code, and downstream passes will drag
2309   //   it back from the precondition block to the preheader.
2310   {
2311     ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
2312 
2313     Value *Opnd0 = PopCntZext;
2314     Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
2315     if (PreCond->getOperand(0) != Var)
2316       std::swap(Opnd0, Opnd1);
2317 
2318     ICmpInst *NewPreCond = cast<ICmpInst>(
2319         Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
2320     PreCondBr->setCondition(NewPreCond);
2321 
2322     RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
2323   }
2324 
2325   // Step 3: Note that the population count is exactly the trip count of the
2326   // loop in question, which enable us to convert the loop from noncountable
2327   // loop into a countable one. The benefit is twofold:
2328   //
2329   //  - If the loop only counts population, the entire loop becomes dead after
2330   //    the transformation. It is a lot easier to prove a countable loop dead
2331   //    than to prove a noncountable one. (In some C dialects, an infinite loop
2332   //    isn't dead even if it computes nothing useful. In general, DCE needs
2333   //    to prove a noncountable loop finite before safely delete it.)
2334   //
2335   //  - If the loop also performs something else, it remains alive.
2336   //    Since it is transformed to countable form, it can be aggressively
2337   //    optimized by some optimizations which are in general not applicable
2338   //    to a noncountable loop.
2339   //
2340   // After this step, this loop (conceptually) would look like following:
2341   //   newcnt = __builtin_ctpop(x);
2342   //   t = newcnt;
2343   //   if (x)
2344   //     do { cnt++; x &= x-1; t--) } while (t > 0);
2345   BasicBlock *Body = *(CurLoop->block_begin());
2346   {
2347     auto *LbBr = cast<BranchInst>(Body->getTerminator());
2348     ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
2349     Type *Ty = TripCnt->getType();
2350 
2351     PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi");
2352     TcPhi->insertBefore(Body->begin());
2353 
2354     Builder.SetInsertPoint(LbCond);
2355     Instruction *TcDec = cast<Instruction>(
2356         Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
2357                           "tcdec", false, true));
2358 
2359     TcPhi->addIncoming(TripCnt, PreHead);
2360     TcPhi->addIncoming(TcDec, Body);
2361 
2362     CmpInst::Predicate Pred =
2363         (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
2364     LbCond->setPredicate(Pred);
2365     LbCond->setOperand(0, TcDec);
2366     LbCond->setOperand(1, ConstantInt::get(Ty, 0));
2367   }
2368 
2369   // Step 4: All the references to the original population counter outside
2370   //  the loop are replaced with the NewCount -- the value returned from
2371   //  __builtin_ctpop().
2372   CntInst->replaceUsesOutsideBlock(NewCount, Body);
2373 
2374   // step 5: Forget the "non-computable" trip-count SCEV associated with the
2375   //   loop. The loop would otherwise not be deleted even if it becomes empty.
2376   SE->forgetLoop(CurLoop);
2377 }
2378 
2379 /// Match loop-invariant value.
2380 template <typename SubPattern_t> struct match_LoopInvariant {
2381   SubPattern_t SubPattern;
2382   const Loop *L;
2383 
2384   match_LoopInvariant(const SubPattern_t &SP, const Loop *L)
2385       : SubPattern(SP), L(L) {}
2386 
2387   template <typename ITy> bool match(ITy *V) {
2388     return L->isLoopInvariant(V) && SubPattern.match(V);
2389   }
2390 };
2391 
2392 /// Matches if the value is loop-invariant.
2393 template <typename Ty>
2394 inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) {
2395   return match_LoopInvariant<Ty>(M, L);
2396 }
2397 
2398 /// Return true if the idiom is detected in the loop.
2399 ///
2400 /// The core idiom we are trying to detect is:
2401 /// \code
2402 ///   entry:
2403 ///     <...>
2404 ///     %bitmask = shl i32 1, %bitpos
2405 ///     br label %loop
2406 ///
2407 ///   loop:
2408 ///     %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2409 ///     %x.curr.bitmasked = and i32 %x.curr, %bitmask
2410 ///     %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2411 ///     %x.next = shl i32 %x.curr, 1
2412 ///     <...>
2413 ///     br i1 %x.curr.isbitunset, label %loop, label %end
2414 ///
2415 ///   end:
2416 ///     %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2417 ///     %x.next.res = phi i32 [ %x.next, %loop ] <...>
2418 ///     <...>
2419 /// \endcode
2420 static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX,
2421                                          Value *&BitMask, Value *&BitPos,
2422                                          Value *&CurrX, Instruction *&NextX) {
2423   LLVM_DEBUG(dbgs() << DEBUG_TYPE
2424              " Performing shift-until-bittest idiom detection.\n");
2425 
2426   // Give up if the loop has multiple blocks or multiple backedges.
2427   if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2428     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2429     return false;
2430   }
2431 
2432   BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2433   BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2434   assert(LoopPreheaderBB && "There is always a loop preheader.");
2435 
2436   using namespace PatternMatch;
2437 
2438   // Step 1: Check if the loop backedge is in desirable form.
2439 
2440   CmpPredicate Pred;
2441   Value *CmpLHS, *CmpRHS;
2442   BasicBlock *TrueBB, *FalseBB;
2443   if (!match(LoopHeaderBB->getTerminator(),
2444              m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)),
2445                   m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) {
2446     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2447     return false;
2448   }
2449 
2450   // Step 2: Check if the backedge's condition is in desirable form.
2451 
2452   auto MatchVariableBitMask = [&]() {
2453     return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2454            match(CmpLHS,
2455                  m_c_And(m_Value(CurrX),
2456                          m_CombineAnd(
2457                              m_Value(BitMask),
2458                              m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)),
2459                                              CurLoop))));
2460   };
2461   auto MatchConstantBitMask = [&]() {
2462     return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2463            match(CmpLHS, m_And(m_Value(CurrX),
2464                                m_CombineAnd(m_Value(BitMask), m_Power2()))) &&
2465            (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask)));
2466   };
2467   auto MatchDecomposableConstantBitMask = [&]() {
2468     auto Res = llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred);
2469     if (Res && Res->Mask.isPowerOf2()) {
2470       assert(ICmpInst::isEquality(Res->Pred));
2471       Pred = Res->Pred;
2472       CurrX = Res->X;
2473       BitMask = ConstantInt::get(CurrX->getType(), Res->Mask);
2474       BitPos = ConstantInt::get(CurrX->getType(), Res->Mask.logBase2());
2475       return true;
2476     }
2477     return false;
2478   };
2479 
2480   if (!MatchVariableBitMask() && !MatchConstantBitMask() &&
2481       !MatchDecomposableConstantBitMask()) {
2482     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n");
2483     return false;
2484   }
2485 
2486   // Step 3: Check if the recurrence is in desirable form.
2487   auto *CurrXPN = dyn_cast<PHINode>(CurrX);
2488   if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) {
2489     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2490     return false;
2491   }
2492 
2493   BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB);
2494   NextX =
2495       dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB));
2496 
2497   assert(CurLoop->isLoopInvariant(BaseX) &&
2498          "Expected BaseX to be available in the preheader!");
2499 
2500   if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) {
2501     // FIXME: support right-shift?
2502     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2503     return false;
2504   }
2505 
2506   // Step 4: Check if the backedge's destinations are in desirable form.
2507 
2508   assert(ICmpInst::isEquality(Pred) &&
2509          "Should only get equality predicates here.");
2510 
2511   // cmp-br is commutative, so canonicalize to a single variant.
2512   if (Pred != ICmpInst::Predicate::ICMP_EQ) {
2513     Pred = ICmpInst::getInversePredicate(Pred);
2514     std::swap(TrueBB, FalseBB);
2515   }
2516 
2517   // We expect to exit loop when comparison yields false,
2518   // so when it yields true we should branch back to loop header.
2519   if (TrueBB != LoopHeaderBB) {
2520     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2521     return false;
2522   }
2523 
2524   // Okay, idiom checks out.
2525   return true;
2526 }
2527 
2528 /// Look for the following loop:
2529 /// \code
2530 ///   entry:
2531 ///     <...>
2532 ///     %bitmask = shl i32 1, %bitpos
2533 ///     br label %loop
2534 ///
2535 ///   loop:
2536 ///     %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2537 ///     %x.curr.bitmasked = and i32 %x.curr, %bitmask
2538 ///     %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2539 ///     %x.next = shl i32 %x.curr, 1
2540 ///     <...>
2541 ///     br i1 %x.curr.isbitunset, label %loop, label %end
2542 ///
2543 ///   end:
2544 ///     %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2545 ///     %x.next.res = phi i32 [ %x.next, %loop ] <...>
2546 ///     <...>
2547 /// \endcode
2548 ///
2549 /// And transform it into:
2550 /// \code
2551 ///   entry:
2552 ///     %bitmask = shl i32 1, %bitpos
2553 ///     %lowbitmask = add i32 %bitmask, -1
2554 ///     %mask = or i32 %lowbitmask, %bitmask
2555 ///     %x.masked = and i32 %x, %mask
2556 ///     %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked,
2557 ///                                                         i1 true)
2558 ///     %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros
2559 ///     %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1
2560 ///     %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos
2561 ///     %tripcount = add i32 %backedgetakencount, 1
2562 ///     %x.curr = shl i32 %x, %backedgetakencount
2563 ///     %x.next = shl i32 %x, %tripcount
2564 ///     br label %loop
2565 ///
2566 ///   loop:
2567 ///     %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ]
2568 ///     %loop.iv.next = add nuw i32 %loop.iv, 1
2569 ///     %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount
2570 ///     <...>
2571 ///     br i1 %loop.ivcheck, label %end, label %loop
2572 ///
2573 ///   end:
2574 ///     %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2575 ///     %x.next.res = phi i32 [ %x.next, %loop ] <...>
2576 ///     <...>
2577 /// \endcode
2578 bool LoopIdiomRecognize::recognizeShiftUntilBitTest() {
2579   bool MadeChange = false;
2580 
2581   Value *X, *BitMask, *BitPos, *XCurr;
2582   Instruction *XNext;
2583   if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr,
2584                                     XNext)) {
2585     LLVM_DEBUG(dbgs() << DEBUG_TYPE
2586                " shift-until-bittest idiom detection failed.\n");
2587     return MadeChange;
2588   }
2589   LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n");
2590 
2591   // Ok, it is the idiom we were looking for, we *could* transform this loop,
2592   // but is it profitable to transform?
2593 
2594   BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2595   BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2596   assert(LoopPreheaderBB && "There is always a loop preheader.");
2597 
2598   BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2599   assert(SuccessorBB && "There is only a single successor.");
2600 
2601   IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2602   Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc());
2603 
2604   Intrinsic::ID IntrID = Intrinsic::ctlz;
2605   Type *Ty = X->getType();
2606   unsigned Bitwidth = Ty->getScalarSizeInBits();
2607 
2608   TargetTransformInfo::TargetCostKind CostKind =
2609       TargetTransformInfo::TCK_SizeAndLatency;
2610 
2611   // The rewrite is considered to be unprofitable iff and only iff the
2612   // intrinsic/shift we'll use are not cheap. Note that we are okay with *just*
2613   // making the loop countable, even if nothing else changes.
2614   IntrinsicCostAttributes Attrs(
2615       IntrID, Ty, {PoisonValue::get(Ty), /*is_zero_poison=*/Builder.getTrue()});
2616   InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2617   if (Cost > TargetTransformInfo::TCC_Basic) {
2618     LLVM_DEBUG(dbgs() << DEBUG_TYPE
2619                " Intrinsic is too costly, not beneficial\n");
2620     return MadeChange;
2621   }
2622   if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) >
2623       TargetTransformInfo::TCC_Basic) {
2624     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n");
2625     return MadeChange;
2626   }
2627 
2628   // Ok, transform appears worthwhile.
2629   MadeChange = true;
2630 
2631   if (!isGuaranteedNotToBeUndefOrPoison(BitPos)) {
2632     // BitMask may be computed from BitPos, Freeze BitPos so we can increase
2633     // it's use count.
2634     std::optional<BasicBlock::iterator> InsertPt = std::nullopt;
2635     if (auto *BitPosI = dyn_cast<Instruction>(BitPos))
2636       InsertPt = BitPosI->getInsertionPointAfterDef();
2637     else
2638       InsertPt = DT->getRoot()->getFirstNonPHIOrDbgOrAlloca();
2639     if (!InsertPt)
2640       return false;
2641     FreezeInst *BitPosFrozen =
2642         new FreezeInst(BitPos, BitPos->getName() + ".fr", *InsertPt);
2643     BitPos->replaceUsesWithIf(BitPosFrozen, [BitPosFrozen](Use &U) {
2644       return U.getUser() != BitPosFrozen;
2645     });
2646     BitPos = BitPosFrozen;
2647   }
2648 
2649   // Step 1: Compute the loop trip count.
2650 
2651   Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty),
2652                                         BitPos->getName() + ".lowbitmask");
2653   Value *Mask =
2654       Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask");
2655   Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked");
2656   CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic(
2657       IntrID, Ty, {XMasked, /*is_zero_poison=*/Builder.getTrue()},
2658       /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros");
2659   Value *XMaskedNumActiveBits = Builder.CreateSub(
2660       ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros,
2661       XMasked->getName() + ".numactivebits", /*HasNUW=*/true,
2662       /*HasNSW=*/Bitwidth != 2);
2663   Value *XMaskedLeadingOnePos =
2664       Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty),
2665                         XMasked->getName() + ".leadingonepos", /*HasNUW=*/false,
2666                         /*HasNSW=*/Bitwidth > 2);
2667 
2668   Value *LoopBackedgeTakenCount = Builder.CreateSub(
2669       BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount",
2670       /*HasNUW=*/true, /*HasNSW=*/true);
2671   // We know loop's backedge-taken count, but what's loop's trip count?
2672   // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2673   Value *LoopTripCount =
2674       Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
2675                         CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
2676                         /*HasNSW=*/Bitwidth != 2);
2677 
2678   // Step 2: Compute the recurrence's final value without a loop.
2679 
2680   // NewX is always safe to compute, because `LoopBackedgeTakenCount`
2681   // will always be smaller than `bitwidth(X)`, i.e. we never get poison.
2682   Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount);
2683   NewX->takeName(XCurr);
2684   if (auto *I = dyn_cast<Instruction>(NewX))
2685     I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2686 
2687   Value *NewXNext;
2688   // Rewriting XNext is more complicated, however, because `X << LoopTripCount`
2689   // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen
2690   // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know
2691   // that isn't the case, we'll need to emit an alternative, safe IR.
2692   if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() ||
2693       PatternMatch::match(
2694           BitPos, PatternMatch::m_SpecificInt_ICMP(
2695                       ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(),
2696                                                Ty->getScalarSizeInBits() - 1))))
2697     NewXNext = Builder.CreateShl(X, LoopTripCount);
2698   else {
2699     // Otherwise, just additionally shift by one. It's the smallest solution,
2700     // alternatively, we could check that NewX is INT_MIN (or BitPos is )
2701     // and select 0 instead.
2702     NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1));
2703   }
2704 
2705   NewXNext->takeName(XNext);
2706   if (auto *I = dyn_cast<Instruction>(NewXNext))
2707     I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2708 
2709   // Step 3: Adjust the successor basic block to recieve the computed
2710   //         recurrence's final value instead of the recurrence itself.
2711 
2712   XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB);
2713   XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB);
2714 
2715   // Step 4: Rewrite the loop into a countable form, with canonical IV.
2716 
2717   // The new canonical induction variable.
2718   Builder.SetInsertPoint(LoopHeaderBB, LoopHeaderBB->begin());
2719   auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
2720 
2721   // The induction itself.
2722   // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2723   Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
2724   auto *IVNext =
2725       Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next",
2726                         /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
2727 
2728   // The loop trip count check.
2729   auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount,
2730                                        CurLoop->getName() + ".ivcheck");
2731   Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB);
2732   LoopHeaderBB->getTerminator()->eraseFromParent();
2733 
2734   // Populate the IV PHI.
2735   IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
2736   IV->addIncoming(IVNext, LoopHeaderBB);
2737 
2738   // Step 5: Forget the "non-computable" trip-count SCEV associated with the
2739   //   loop. The loop would otherwise not be deleted even if it becomes empty.
2740 
2741   SE->forgetLoop(CurLoop);
2742 
2743   // Other passes will take care of actually deleting the loop if possible.
2744 
2745   LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n");
2746 
2747   ++NumShiftUntilBitTest;
2748   return MadeChange;
2749 }
2750 
2751 /// Return true if the idiom is detected in the loop.
2752 ///
2753 /// The core idiom we are trying to detect is:
2754 /// \code
2755 ///   entry:
2756 ///     <...>
2757 ///     %start = <...>
2758 ///     %extraoffset = <...>
2759 ///     <...>
2760 ///     br label %for.cond
2761 ///
2762 ///   loop:
2763 ///     %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2764 ///     %nbits = add nsw i8 %iv, %extraoffset
2765 ///     %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2766 ///     %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2767 ///     %iv.next = add i8 %iv, 1
2768 ///     <...>
2769 ///     br i1 %val.shifted.iszero, label %end, label %loop
2770 ///
2771 ///   end:
2772 ///     %iv.res = phi i8 [ %iv, %loop ] <...>
2773 ///     %nbits.res = phi i8 [ %nbits, %loop ] <...>
2774 ///     %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2775 ///     %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2776 ///     %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2777 ///     <...>
2778 /// \endcode
2779 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE,
2780                                       Instruction *&ValShiftedIsZero,
2781                                       Intrinsic::ID &IntrinID, Instruction *&IV,
2782                                       Value *&Start, Value *&Val,
2783                                       const SCEV *&ExtraOffsetExpr,
2784                                       bool &InvertedCond) {
2785   LLVM_DEBUG(dbgs() << DEBUG_TYPE
2786              " Performing shift-until-zero idiom detection.\n");
2787 
2788   // Give up if the loop has multiple blocks or multiple backedges.
2789   if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2790     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2791     return false;
2792   }
2793 
2794   Instruction *ValShifted, *NBits, *IVNext;
2795   Value *ExtraOffset;
2796 
2797   BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2798   BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2799   assert(LoopPreheaderBB && "There is always a loop preheader.");
2800 
2801   using namespace PatternMatch;
2802 
2803   // Step 1: Check if the loop backedge, condition is in desirable form.
2804 
2805   CmpPredicate Pred;
2806   BasicBlock *TrueBB, *FalseBB;
2807   if (!match(LoopHeaderBB->getTerminator(),
2808              m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB),
2809                   m_BasicBlock(FalseBB))) ||
2810       !match(ValShiftedIsZero,
2811              m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) ||
2812       !ICmpInst::isEquality(Pred)) {
2813     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2814     return false;
2815   }
2816 
2817   // Step 2: Check if the comparison's operand is in desirable form.
2818   // FIXME: Val could be a one-input PHI node, which we should look past.
2819   if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop),
2820                                  m_Instruction(NBits)))) {
2821     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n");
2822     return false;
2823   }
2824   IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz
2825                                                          : Intrinsic::ctlz;
2826 
2827   // Step 3: Check if the shift amount is in desirable form.
2828 
2829   if (match(NBits, m_c_Add(m_Instruction(IV),
2830                            m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2831       (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap()))
2832     ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset));
2833   else if (match(NBits,
2834                  m_Sub(m_Instruction(IV),
2835                        m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2836            NBits->hasNoSignedWrap())
2837     ExtraOffsetExpr = SE->getSCEV(ExtraOffset);
2838   else {
2839     IV = NBits;
2840     ExtraOffsetExpr = SE->getZero(NBits->getType());
2841   }
2842 
2843   // Step 4: Check if the recurrence is in desirable form.
2844   auto *IVPN = dyn_cast<PHINode>(IV);
2845   if (!IVPN || IVPN->getParent() != LoopHeaderBB) {
2846     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2847     return false;
2848   }
2849 
2850   Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB);
2851   IVNext = dyn_cast<Instruction>(IVPN->getIncomingValueForBlock(LoopHeaderBB));
2852 
2853   if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) {
2854     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2855     return false;
2856   }
2857 
2858   // Step 4: Check if the backedge's destinations are in desirable form.
2859 
2860   assert(ICmpInst::isEquality(Pred) &&
2861          "Should only get equality predicates here.");
2862 
2863   // cmp-br is commutative, so canonicalize to a single variant.
2864   InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ;
2865   if (InvertedCond) {
2866     Pred = ICmpInst::getInversePredicate(Pred);
2867     std::swap(TrueBB, FalseBB);
2868   }
2869 
2870   // We expect to exit loop when comparison yields true,
2871   // so when it yields false we should branch back to loop header.
2872   if (FalseBB != LoopHeaderBB) {
2873     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2874     return false;
2875   }
2876 
2877   // The new, countable, loop will certainly only run a known number of
2878   // iterations, It won't be infinite. But the old loop might be infinite
2879   // under certain conditions. For logical shifts, the value will become zero
2880   // after at most bitwidth(%Val) loop iterations. However, for arithmetic
2881   // right-shift, iff the sign bit was set, the value will never become zero,
2882   // and the loop may never finish.
2883   if (ValShifted->getOpcode() == Instruction::AShr &&
2884       !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) {
2885     LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n");
2886     return false;
2887   }
2888 
2889   // Okay, idiom checks out.
2890   return true;
2891 }
2892 
2893 /// Look for the following loop:
2894 /// \code
2895 ///   entry:
2896 ///     <...>
2897 ///     %start = <...>
2898 ///     %extraoffset = <...>
2899 ///     <...>
2900 ///     br label %for.cond
2901 ///
2902 ///   loop:
2903 ///     %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2904 ///     %nbits = add nsw i8 %iv, %extraoffset
2905 ///     %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2906 ///     %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2907 ///     %iv.next = add i8 %iv, 1
2908 ///     <...>
2909 ///     br i1 %val.shifted.iszero, label %end, label %loop
2910 ///
2911 ///   end:
2912 ///     %iv.res = phi i8 [ %iv, %loop ] <...>
2913 ///     %nbits.res = phi i8 [ %nbits, %loop ] <...>
2914 ///     %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2915 ///     %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2916 ///     %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2917 ///     <...>
2918 /// \endcode
2919 ///
2920 /// And transform it into:
2921 /// \code
2922 ///   entry:
2923 ///     <...>
2924 ///     %start = <...>
2925 ///     %extraoffset = <...>
2926 ///     <...>
2927 ///     %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0)
2928 ///     %val.numactivebits = sub i8 8, %val.numleadingzeros
2929 ///     %extraoffset.neg = sub i8 0, %extraoffset
2930 ///     %tmp = add i8 %val.numactivebits, %extraoffset.neg
2931 ///     %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start)
2932 ///     %loop.tripcount = sub i8 %iv.final, %start
2933 ///     br label %loop
2934 ///
2935 ///   loop:
2936 ///     %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ]
2937 ///     %loop.iv.next = add i8 %loop.iv, 1
2938 ///     %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount
2939 ///     %iv = add i8 %loop.iv, %start
2940 ///     <...>
2941 ///     br i1 %loop.ivcheck, label %end, label %loop
2942 ///
2943 ///   end:
2944 ///     %iv.res = phi i8 [ %iv.final, %loop ] <...>
2945 ///     <...>
2946 /// \endcode
2947 bool LoopIdiomRecognize::recognizeShiftUntilZero() {
2948   bool MadeChange = false;
2949 
2950   Instruction *ValShiftedIsZero;
2951   Intrinsic::ID IntrID;
2952   Instruction *IV;
2953   Value *Start, *Val;
2954   const SCEV *ExtraOffsetExpr;
2955   bool InvertedCond;
2956   if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV,
2957                                  Start, Val, ExtraOffsetExpr, InvertedCond)) {
2958     LLVM_DEBUG(dbgs() << DEBUG_TYPE
2959                " shift-until-zero idiom detection failed.\n");
2960     return MadeChange;
2961   }
2962   LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n");
2963 
2964   // Ok, it is the idiom we were looking for, we *could* transform this loop,
2965   // but is it profitable to transform?
2966 
2967   BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2968   BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2969   assert(LoopPreheaderBB && "There is always a loop preheader.");
2970 
2971   BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2972   assert(SuccessorBB && "There is only a single successor.");
2973 
2974   IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2975   Builder.SetCurrentDebugLocation(IV->getDebugLoc());
2976 
2977   Type *Ty = Val->getType();
2978   unsigned Bitwidth = Ty->getScalarSizeInBits();
2979 
2980   TargetTransformInfo::TargetCostKind CostKind =
2981       TargetTransformInfo::TCK_SizeAndLatency;
2982 
2983   // The rewrite is considered to be unprofitable iff and only iff the
2984   // intrinsic we'll use are not cheap. Note that we are okay with *just*
2985   // making the loop countable, even if nothing else changes.
2986   IntrinsicCostAttributes Attrs(
2987       IntrID, Ty, {PoisonValue::get(Ty), /*is_zero_poison=*/Builder.getFalse()});
2988   InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2989   if (Cost > TargetTransformInfo::TCC_Basic) {
2990     LLVM_DEBUG(dbgs() << DEBUG_TYPE
2991                " Intrinsic is too costly, not beneficial\n");
2992     return MadeChange;
2993   }
2994 
2995   // Ok, transform appears worthwhile.
2996   MadeChange = true;
2997 
2998   bool OffsetIsZero = false;
2999   if (auto *ExtraOffsetExprC = dyn_cast<SCEVConstant>(ExtraOffsetExpr))
3000     OffsetIsZero = ExtraOffsetExprC->isZero();
3001 
3002   // Step 1: Compute the loop's final IV value / trip count.
3003 
3004   CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic(
3005       IntrID, Ty, {Val, /*is_zero_poison=*/Builder.getFalse()},
3006       /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros");
3007   Value *ValNumActiveBits = Builder.CreateSub(
3008       ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros,
3009       Val->getName() + ".numactivebits", /*HasNUW=*/true,
3010       /*HasNSW=*/Bitwidth != 2);
3011 
3012   SCEVExpander Expander(*SE, *DL, "loop-idiom");
3013   Expander.setInsertPoint(&*Builder.GetInsertPoint());
3014   Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr);
3015 
3016   Value *ValNumActiveBitsOffset = Builder.CreateAdd(
3017       ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset",
3018       /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true);
3019   Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty},
3020                                            {ValNumActiveBitsOffset, Start},
3021                                            /*FMFSource=*/nullptr, "iv.final");
3022 
3023   auto *LoopBackedgeTakenCount = cast<Instruction>(Builder.CreateSub(
3024       IVFinal, Start, CurLoop->getName() + ".backedgetakencount",
3025       /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true));
3026   // FIXME: or when the offset was `add nuw`
3027 
3028   // We know loop's backedge-taken count, but what's loop's trip count?
3029   Value *LoopTripCount =
3030       Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
3031                         CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
3032                         /*HasNSW=*/Bitwidth != 2);
3033 
3034   // Step 2: Adjust the successor basic block to recieve the original
3035   //         induction variable's final value instead of the orig. IV itself.
3036 
3037   IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB);
3038 
3039   // Step 3: Rewrite the loop into a countable form, with canonical IV.
3040 
3041   // The new canonical induction variable.
3042   Builder.SetInsertPoint(LoopHeaderBB, LoopHeaderBB->begin());
3043   auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
3044 
3045   // The induction itself.
3046   Builder.SetInsertPoint(LoopHeaderBB, LoopHeaderBB->getFirstNonPHIIt());
3047   auto *CIVNext =
3048       Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next",
3049                         /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
3050 
3051   // The loop trip count check.
3052   auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount,
3053                                         CurLoop->getName() + ".ivcheck");
3054   auto *NewIVCheck = CIVCheck;
3055   if (InvertedCond) {
3056     NewIVCheck = Builder.CreateNot(CIVCheck);
3057     NewIVCheck->takeName(ValShiftedIsZero);
3058   }
3059 
3060   // The original IV, but rebased to be an offset to the CIV.
3061   auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false,
3062                                      /*HasNSW=*/true); // FIXME: what about NUW?
3063   IVDePHId->takeName(IV);
3064 
3065   // The loop terminator.
3066   Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
3067   Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB);
3068   LoopHeaderBB->getTerminator()->eraseFromParent();
3069 
3070   // Populate the IV PHI.
3071   CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
3072   CIV->addIncoming(CIVNext, LoopHeaderBB);
3073 
3074   // Step 4: Forget the "non-computable" trip-count SCEV associated with the
3075   //   loop. The loop would otherwise not be deleted even if it becomes empty.
3076 
3077   SE->forgetLoop(CurLoop);
3078 
3079   // Step 5: Try to cleanup the loop's body somewhat.
3080   IV->replaceAllUsesWith(IVDePHId);
3081   IV->eraseFromParent();
3082 
3083   ValShiftedIsZero->replaceAllUsesWith(NewIVCheck);
3084   ValShiftedIsZero->eraseFromParent();
3085 
3086   // Other passes will take care of actually deleting the loop if possible.
3087 
3088   LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n");
3089 
3090   ++NumShiftUntilZero;
3091   return MadeChange;
3092 }
3093