1 //===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass implements an idiom recognizer that transforms simple loops into a
10 // non-loop form. In cases that this kicks in, it can be a significant
11 // performance win.
12 //
13 // If compiling for code size we avoid idiom recognition if the resulting
14 // code could be larger than the code for the original loop. One way this could
15 // happen is if the loop is not removable after idiom recognition due to the
16 // presence of non-idiom instructions. The initial implementation of the
17 // heuristics applies to idioms in multi-block loops.
18 //
19 //===----------------------------------------------------------------------===//
20 //
21 // TODO List:
22 //
23 // Future loop memory idioms to recognize:
24 // memcmp, strlen, etc.
25 // Future floating point idioms to recognize in -ffast-math mode:
26 // fpowi
27 // Future integer operation idioms to recognize:
28 // ctpop
29 //
30 // Beware that isel's default lowering for ctpop is highly inefficient for
31 // i64 and larger types when i64 is legal and the value has few bits set. It
32 // would be good to enhance isel to emit a loop for ctpop in this case.
33 //
34 // This could recognize common matrix multiplies and dot product idioms and
35 // replace them with calls to BLAS (if linked in??).
36 //
37 //===----------------------------------------------------------------------===//
38
39 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
40 #include "llvm/ADT/APInt.h"
41 #include "llvm/ADT/ArrayRef.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/MapVector.h"
44 #include "llvm/ADT/SetVector.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/SmallVector.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/StringRef.h"
49 #include "llvm/Analysis/AliasAnalysis.h"
50 #include "llvm/Analysis/CmpInstAnalysis.h"
51 #include "llvm/Analysis/LoopAccessAnalysis.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Analysis/MemoryLocation.h"
55 #include "llvm/Analysis/MemorySSA.h"
56 #include "llvm/Analysis/MemorySSAUpdater.h"
57 #include "llvm/Analysis/MustExecute.h"
58 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
59 #include "llvm/Analysis/ScalarEvolution.h"
60 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
61 #include "llvm/Analysis/TargetLibraryInfo.h"
62 #include "llvm/Analysis/TargetTransformInfo.h"
63 #include "llvm/Analysis/ValueTracking.h"
64 #include "llvm/IR/Attributes.h"
65 #include "llvm/IR/BasicBlock.h"
66 #include "llvm/IR/Constant.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugLoc.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/GlobalValue.h"
73 #include "llvm/IR/GlobalVariable.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/InstrTypes.h"
76 #include "llvm/IR/Instruction.h"
77 #include "llvm/IR/Instructions.h"
78 #include "llvm/IR/IntrinsicInst.h"
79 #include "llvm/IR/Intrinsics.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PassManager.h"
83 #include "llvm/IR/PatternMatch.h"
84 #include "llvm/IR/Type.h"
85 #include "llvm/IR/User.h"
86 #include "llvm/IR/Value.h"
87 #include "llvm/IR/ValueHandle.h"
88 #include "llvm/InitializePasses.h"
89 #include "llvm/Pass.h"
90 #include "llvm/Support/Casting.h"
91 #include "llvm/Support/CommandLine.h"
92 #include "llvm/Support/Debug.h"
93 #include "llvm/Support/InstructionCost.h"
94 #include "llvm/Support/raw_ostream.h"
95 #include "llvm/Transforms/Scalar.h"
96 #include "llvm/Transforms/Utils/BuildLibCalls.h"
97 #include "llvm/Transforms/Utils/Local.h"
98 #include "llvm/Transforms/Utils/LoopUtils.h"
99 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
100 #include <algorithm>
101 #include <cassert>
102 #include <cstdint>
103 #include <utility>
104 #include <vector>
105
106 using namespace llvm;
107
108 #define DEBUG_TYPE "loop-idiom"
109
110 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
111 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
112 STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores");
113 STATISTIC(
114 NumShiftUntilBitTest,
115 "Number of uncountable loops recognized as 'shift until bitttest' idiom");
116 STATISTIC(NumShiftUntilZero,
117 "Number of uncountable loops recognized as 'shift until zero' idiom");
118
119 bool DisableLIRP::All;
120 static cl::opt<bool, true>
121 DisableLIRPAll("disable-" DEBUG_TYPE "-all",
122 cl::desc("Options to disable Loop Idiom Recognize Pass."),
123 cl::location(DisableLIRP::All), cl::init(false),
124 cl::ReallyHidden);
125
126 bool DisableLIRP::Memset;
127 static cl::opt<bool, true>
128 DisableLIRPMemset("disable-" DEBUG_TYPE "-memset",
129 cl::desc("Proceed with loop idiom recognize pass, but do "
130 "not convert loop(s) to memset."),
131 cl::location(DisableLIRP::Memset), cl::init(false),
132 cl::ReallyHidden);
133
134 bool DisableLIRP::Memcpy;
135 static cl::opt<bool, true>
136 DisableLIRPMemcpy("disable-" DEBUG_TYPE "-memcpy",
137 cl::desc("Proceed with loop idiom recognize pass, but do "
138 "not convert loop(s) to memcpy."),
139 cl::location(DisableLIRP::Memcpy), cl::init(false),
140 cl::ReallyHidden);
141
142 static cl::opt<bool> UseLIRCodeSizeHeurs(
143 "use-lir-code-size-heurs",
144 cl::desc("Use loop idiom recognition code size heuristics when compiling"
145 "with -Os/-Oz"),
146 cl::init(true), cl::Hidden);
147
148 namespace {
149
150 class LoopIdiomRecognize {
151 Loop *CurLoop = nullptr;
152 AliasAnalysis *AA;
153 DominatorTree *DT;
154 LoopInfo *LI;
155 ScalarEvolution *SE;
156 TargetLibraryInfo *TLI;
157 const TargetTransformInfo *TTI;
158 const DataLayout *DL;
159 OptimizationRemarkEmitter &ORE;
160 bool ApplyCodeSizeHeuristics;
161 std::unique_ptr<MemorySSAUpdater> MSSAU;
162
163 public:
LoopIdiomRecognize(AliasAnalysis * AA,DominatorTree * DT,LoopInfo * LI,ScalarEvolution * SE,TargetLibraryInfo * TLI,const TargetTransformInfo * TTI,MemorySSA * MSSA,const DataLayout * DL,OptimizationRemarkEmitter & ORE)164 explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
165 LoopInfo *LI, ScalarEvolution *SE,
166 TargetLibraryInfo *TLI,
167 const TargetTransformInfo *TTI, MemorySSA *MSSA,
168 const DataLayout *DL,
169 OptimizationRemarkEmitter &ORE)
170 : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) {
171 if (MSSA)
172 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
173 }
174
175 bool runOnLoop(Loop *L);
176
177 private:
178 using StoreList = SmallVector<StoreInst *, 8>;
179 using StoreListMap = MapVector<Value *, StoreList>;
180
181 StoreListMap StoreRefsForMemset;
182 StoreListMap StoreRefsForMemsetPattern;
183 StoreList StoreRefsForMemcpy;
184 bool HasMemset;
185 bool HasMemsetPattern;
186 bool HasMemcpy;
187
188 /// Return code for isLegalStore()
189 enum LegalStoreKind {
190 None = 0,
191 Memset,
192 MemsetPattern,
193 Memcpy,
194 UnorderedAtomicMemcpy,
195 DontUse // Dummy retval never to be used. Allows catching errors in retval
196 // handling.
197 };
198
199 /// \name Countable Loop Idiom Handling
200 /// @{
201
202 bool runOnCountableLoop();
203 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
204 SmallVectorImpl<BasicBlock *> &ExitBlocks);
205
206 void collectStores(BasicBlock *BB);
207 LegalStoreKind isLegalStore(StoreInst *SI);
208 enum class ForMemset { No, Yes };
209 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
210 ForMemset For);
211
212 template <typename MemInst>
213 bool processLoopMemIntrinsic(
214 BasicBlock *BB,
215 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
216 const SCEV *BECount);
217 bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount);
218 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
219
220 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
221 MaybeAlign StoreAlignment, Value *StoredVal,
222 Instruction *TheStore,
223 SmallPtrSetImpl<Instruction *> &Stores,
224 const SCEVAddRecExpr *Ev, const SCEV *BECount,
225 bool NegStride, bool IsLoopMemset = false);
226 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
227 bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr,
228 unsigned StoreSize, MaybeAlign StoreAlign,
229 MaybeAlign LoadAlign, Instruction *TheStore,
230 Instruction *TheLoad,
231 const SCEVAddRecExpr *StoreEv,
232 const SCEVAddRecExpr *LoadEv,
233 const SCEV *BECount);
234 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
235 bool IsLoopMemset = false);
236
237 /// @}
238 /// \name Noncountable Loop Idiom Handling
239 /// @{
240
241 bool runOnNoncountableLoop();
242
243 bool recognizePopcount();
244 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
245 PHINode *CntPhi, Value *Var);
246 bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz
247 void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB,
248 Instruction *CntInst, PHINode *CntPhi,
249 Value *Var, Instruction *DefX,
250 const DebugLoc &DL, bool ZeroCheck,
251 bool IsCntPhiUsedOutsideLoop);
252
253 bool recognizeShiftUntilBitTest();
254 bool recognizeShiftUntilZero();
255
256 /// @}
257 };
258
259 class LoopIdiomRecognizeLegacyPass : public LoopPass {
260 public:
261 static char ID;
262
LoopIdiomRecognizeLegacyPass()263 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
264 initializeLoopIdiomRecognizeLegacyPassPass(
265 *PassRegistry::getPassRegistry());
266 }
267
runOnLoop(Loop * L,LPPassManager & LPM)268 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
269 if (DisableLIRP::All)
270 return false;
271
272 if (skipLoop(L))
273 return false;
274
275 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
276 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
277 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
278 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
279 TargetLibraryInfo *TLI =
280 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
281 *L->getHeader()->getParent());
282 const TargetTransformInfo *TTI =
283 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
284 *L->getHeader()->getParent());
285 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
286 auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>();
287 MemorySSA *MSSA = nullptr;
288 if (MSSAAnalysis)
289 MSSA = &MSSAAnalysis->getMSSA();
290
291 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
292 // pass. Function analyses need to be preserved across loop transformations
293 // but ORE cannot be preserved (see comment before the pass definition).
294 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
295
296 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, MSSA, DL, ORE);
297 return LIR.runOnLoop(L);
298 }
299
300 /// This transformation requires natural loop information & requires that
301 /// loop preheaders be inserted into the CFG.
getAnalysisUsage(AnalysisUsage & AU) const302 void getAnalysisUsage(AnalysisUsage &AU) const override {
303 AU.addRequired<TargetLibraryInfoWrapperPass>();
304 AU.addRequired<TargetTransformInfoWrapperPass>();
305 AU.addPreserved<MemorySSAWrapperPass>();
306 getLoopAnalysisUsage(AU);
307 }
308 };
309
310 } // end anonymous namespace
311
312 char LoopIdiomRecognizeLegacyPass::ID = 0;
313
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater &)314 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
315 LoopStandardAnalysisResults &AR,
316 LPMUpdater &) {
317 if (DisableLIRP::All)
318 return PreservedAnalyses::all();
319
320 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
321
322 // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
323 // pass. Function analyses need to be preserved across loop transformations
324 // but ORE cannot be preserved (see comment before the pass definition).
325 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
326
327 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI,
328 AR.MSSA, DL, ORE);
329 if (!LIR.runOnLoop(&L))
330 return PreservedAnalyses::all();
331
332 auto PA = getLoopPassPreservedAnalyses();
333 if (AR.MSSA)
334 PA.preserve<MemorySSAAnalysis>();
335 return PA;
336 }
337
338 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
339 "Recognize loop idioms", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)340 INITIALIZE_PASS_DEPENDENCY(LoopPass)
341 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
342 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
343 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
344 "Recognize loop idioms", false, false)
345
346 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
347
deleteDeadInstruction(Instruction * I)348 static void deleteDeadInstruction(Instruction *I) {
349 I->replaceAllUsesWith(UndefValue::get(I->getType()));
350 I->eraseFromParent();
351 }
352
353 //===----------------------------------------------------------------------===//
354 //
355 // Implementation of LoopIdiomRecognize
356 //
357 //===----------------------------------------------------------------------===//
358
runOnLoop(Loop * L)359 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
360 CurLoop = L;
361 // If the loop could not be converted to canonical form, it must have an
362 // indirectbr in it, just give up.
363 if (!L->getLoopPreheader())
364 return false;
365
366 // Disable loop idiom recognition if the function's name is a common idiom.
367 StringRef Name = L->getHeader()->getParent()->getName();
368 if (Name == "memset" || Name == "memcpy")
369 return false;
370
371 // Determine if code size heuristics need to be applied.
372 ApplyCodeSizeHeuristics =
373 L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs;
374
375 HasMemset = TLI->has(LibFunc_memset);
376 HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
377 HasMemcpy = TLI->has(LibFunc_memcpy);
378
379 if (HasMemset || HasMemsetPattern || HasMemcpy)
380 if (SE->hasLoopInvariantBackedgeTakenCount(L))
381 return runOnCountableLoop();
382
383 return runOnNoncountableLoop();
384 }
385
runOnCountableLoop()386 bool LoopIdiomRecognize::runOnCountableLoop() {
387 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
388 assert(!isa<SCEVCouldNotCompute>(BECount) &&
389 "runOnCountableLoop() called on a loop without a predictable"
390 "backedge-taken count");
391
392 // If this loop executes exactly one time, then it should be peeled, not
393 // optimized by this pass.
394 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
395 if (BECst->getAPInt() == 0)
396 return false;
397
398 SmallVector<BasicBlock *, 8> ExitBlocks;
399 CurLoop->getUniqueExitBlocks(ExitBlocks);
400
401 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
402 << CurLoop->getHeader()->getParent()->getName()
403 << "] Countable Loop %" << CurLoop->getHeader()->getName()
404 << "\n");
405
406 // The following transforms hoist stores/memsets into the loop pre-header.
407 // Give up if the loop has instructions that may throw.
408 SimpleLoopSafetyInfo SafetyInfo;
409 SafetyInfo.computeLoopSafetyInfo(CurLoop);
410 if (SafetyInfo.anyBlockMayThrow())
411 return false;
412
413 bool MadeChange = false;
414
415 // Scan all the blocks in the loop that are not in subloops.
416 for (auto *BB : CurLoop->getBlocks()) {
417 // Ignore blocks in subloops.
418 if (LI->getLoopFor(BB) != CurLoop)
419 continue;
420
421 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
422 }
423 return MadeChange;
424 }
425
getStoreStride(const SCEVAddRecExpr * StoreEv)426 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
427 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
428 return ConstStride->getAPInt();
429 }
430
431 /// getMemSetPatternValue - If a strided store of the specified value is safe to
432 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
433 /// be passed in. Otherwise, return null.
434 ///
435 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
436 /// just replicate their input array and then pass on to memset_pattern16.
getMemSetPatternValue(Value * V,const DataLayout * DL)437 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
438 // FIXME: This could check for UndefValue because it can be merged into any
439 // other valid pattern.
440
441 // If the value isn't a constant, we can't promote it to being in a constant
442 // array. We could theoretically do a store to an alloca or something, but
443 // that doesn't seem worthwhile.
444 Constant *C = dyn_cast<Constant>(V);
445 if (!C)
446 return nullptr;
447
448 // Only handle simple values that are a power of two bytes in size.
449 uint64_t Size = DL->getTypeSizeInBits(V->getType());
450 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
451 return nullptr;
452
453 // Don't care enough about darwin/ppc to implement this.
454 if (DL->isBigEndian())
455 return nullptr;
456
457 // Convert to size in bytes.
458 Size /= 8;
459
460 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
461 // if the top and bottom are the same (e.g. for vectors and large integers).
462 if (Size > 16)
463 return nullptr;
464
465 // If the constant is exactly 16 bytes, just use it.
466 if (Size == 16)
467 return C;
468
469 // Otherwise, we'll use an array of the constants.
470 unsigned ArraySize = 16 / Size;
471 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
472 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
473 }
474
475 LoopIdiomRecognize::LegalStoreKind
isLegalStore(StoreInst * SI)476 LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
477 // Don't touch volatile stores.
478 if (SI->isVolatile())
479 return LegalStoreKind::None;
480 // We only want simple or unordered-atomic stores.
481 if (!SI->isUnordered())
482 return LegalStoreKind::None;
483
484 // Avoid merging nontemporal stores.
485 if (SI->getMetadata(LLVMContext::MD_nontemporal))
486 return LegalStoreKind::None;
487
488 Value *StoredVal = SI->getValueOperand();
489 Value *StorePtr = SI->getPointerOperand();
490
491 // Don't convert stores of non-integral pointer types to memsets (which stores
492 // integers).
493 if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
494 return LegalStoreKind::None;
495
496 // Reject stores that are so large that they overflow an unsigned.
497 // When storing out scalable vectors we bail out for now, since the code
498 // below currently only works for constant strides.
499 TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
500 if (SizeInBits.isScalable() || (SizeInBits.getFixedSize() & 7) ||
501 (SizeInBits.getFixedSize() >> 32) != 0)
502 return LegalStoreKind::None;
503
504 // See if the pointer expression is an AddRec like {base,+,1} on the current
505 // loop, which indicates a strided store. If we have something else, it's a
506 // random store we can't handle.
507 const SCEVAddRecExpr *StoreEv =
508 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
509 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
510 return LegalStoreKind::None;
511
512 // Check to see if we have a constant stride.
513 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
514 return LegalStoreKind::None;
515
516 // See if the store can be turned into a memset.
517
518 // If the stored value is a byte-wise value (like i32 -1), then it may be
519 // turned into a memset of i8 -1, assuming that all the consecutive bytes
520 // are stored. A store of i32 0x01020304 can never be turned into a memset,
521 // but it can be turned into memset_pattern if the target supports it.
522 Value *SplatValue = isBytewiseValue(StoredVal, *DL);
523
524 // Note: memset and memset_pattern on unordered-atomic is yet not supported
525 bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple();
526
527 // If we're allowed to form a memset, and the stored value would be
528 // acceptable for memset, use it.
529 if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset &&
530 // Verify that the stored value is loop invariant. If not, we can't
531 // promote the memset.
532 CurLoop->isLoopInvariant(SplatValue)) {
533 // It looks like we can use SplatValue.
534 return LegalStoreKind::Memset;
535 }
536 if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset &&
537 // Don't create memset_pattern16s with address spaces.
538 StorePtr->getType()->getPointerAddressSpace() == 0 &&
539 getMemSetPatternValue(StoredVal, DL)) {
540 // It looks like we can use PatternValue!
541 return LegalStoreKind::MemsetPattern;
542 }
543
544 // Otherwise, see if the store can be turned into a memcpy.
545 if (HasMemcpy && !DisableLIRP::Memcpy) {
546 // Check to see if the stride matches the size of the store. If so, then we
547 // know that every byte is touched in the loop.
548 APInt Stride = getStoreStride(StoreEv);
549 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
550 if (StoreSize != Stride && StoreSize != -Stride)
551 return LegalStoreKind::None;
552
553 // The store must be feeding a non-volatile load.
554 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
555
556 // Only allow non-volatile loads
557 if (!LI || LI->isVolatile())
558 return LegalStoreKind::None;
559 // Only allow simple or unordered-atomic loads
560 if (!LI->isUnordered())
561 return LegalStoreKind::None;
562
563 // See if the pointer expression is an AddRec like {base,+,1} on the current
564 // loop, which indicates a strided load. If we have something else, it's a
565 // random load we can't handle.
566 const SCEVAddRecExpr *LoadEv =
567 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
568 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
569 return LegalStoreKind::None;
570
571 // The store and load must share the same stride.
572 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
573 return LegalStoreKind::None;
574
575 // Success. This store can be converted into a memcpy.
576 UnorderedAtomic = UnorderedAtomic || LI->isAtomic();
577 return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy
578 : LegalStoreKind::Memcpy;
579 }
580 // This store can't be transformed into a memset/memcpy.
581 return LegalStoreKind::None;
582 }
583
collectStores(BasicBlock * BB)584 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
585 StoreRefsForMemset.clear();
586 StoreRefsForMemsetPattern.clear();
587 StoreRefsForMemcpy.clear();
588 for (Instruction &I : *BB) {
589 StoreInst *SI = dyn_cast<StoreInst>(&I);
590 if (!SI)
591 continue;
592
593 // Make sure this is a strided store with a constant stride.
594 switch (isLegalStore(SI)) {
595 case LegalStoreKind::None:
596 // Nothing to do
597 break;
598 case LegalStoreKind::Memset: {
599 // Find the base pointer.
600 Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
601 StoreRefsForMemset[Ptr].push_back(SI);
602 } break;
603 case LegalStoreKind::MemsetPattern: {
604 // Find the base pointer.
605 Value *Ptr = getUnderlyingObject(SI->getPointerOperand());
606 StoreRefsForMemsetPattern[Ptr].push_back(SI);
607 } break;
608 case LegalStoreKind::Memcpy:
609 case LegalStoreKind::UnorderedAtomicMemcpy:
610 StoreRefsForMemcpy.push_back(SI);
611 break;
612 default:
613 assert(false && "unhandled return value");
614 break;
615 }
616 }
617 }
618
619 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
620 /// with the specified backedge count. This block is known to be in the current
621 /// loop and not in any subloops.
runOnLoopBlock(BasicBlock * BB,const SCEV * BECount,SmallVectorImpl<BasicBlock * > & ExitBlocks)622 bool LoopIdiomRecognize::runOnLoopBlock(
623 BasicBlock *BB, const SCEV *BECount,
624 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
625 // We can only promote stores in this block if they are unconditionally
626 // executed in the loop. For a block to be unconditionally executed, it has
627 // to dominate all the exit blocks of the loop. Verify this now.
628 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
629 if (!DT->dominates(BB, ExitBlocks[i]))
630 return false;
631
632 bool MadeChange = false;
633 // Look for store instructions, which may be optimized to memset/memcpy.
634 collectStores(BB);
635
636 // Look for a single store or sets of stores with a common base, which can be
637 // optimized into a memset (memset_pattern). The latter most commonly happens
638 // with structs and handunrolled loops.
639 for (auto &SL : StoreRefsForMemset)
640 MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes);
641
642 for (auto &SL : StoreRefsForMemsetPattern)
643 MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No);
644
645 // Optimize the store into a memcpy, if it feeds an similarly strided load.
646 for (auto &SI : StoreRefsForMemcpy)
647 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
648
649 MadeChange |= processLoopMemIntrinsic<MemCpyInst>(
650 BB, &LoopIdiomRecognize::processLoopMemCpy, BECount);
651 MadeChange |= processLoopMemIntrinsic<MemSetInst>(
652 BB, &LoopIdiomRecognize::processLoopMemSet, BECount);
653
654 return MadeChange;
655 }
656
657 /// See if this store(s) can be promoted to a memset.
processLoopStores(SmallVectorImpl<StoreInst * > & SL,const SCEV * BECount,ForMemset For)658 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
659 const SCEV *BECount, ForMemset For) {
660 // Try to find consecutive stores that can be transformed into memsets.
661 SetVector<StoreInst *> Heads, Tails;
662 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
663
664 // Do a quadratic search on all of the given stores and find
665 // all of the pairs of stores that follow each other.
666 SmallVector<unsigned, 16> IndexQueue;
667 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
668 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
669
670 Value *FirstStoredVal = SL[i]->getValueOperand();
671 Value *FirstStorePtr = SL[i]->getPointerOperand();
672 const SCEVAddRecExpr *FirstStoreEv =
673 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
674 APInt FirstStride = getStoreStride(FirstStoreEv);
675 unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType());
676
677 // See if we can optimize just this store in isolation.
678 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
679 Heads.insert(SL[i]);
680 continue;
681 }
682
683 Value *FirstSplatValue = nullptr;
684 Constant *FirstPatternValue = nullptr;
685
686 if (For == ForMemset::Yes)
687 FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL);
688 else
689 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
690
691 assert((FirstSplatValue || FirstPatternValue) &&
692 "Expected either splat value or pattern value.");
693
694 IndexQueue.clear();
695 // If a store has multiple consecutive store candidates, search Stores
696 // array according to the sequence: from i+1 to e, then from i-1 to 0.
697 // This is because usually pairing with immediate succeeding or preceding
698 // candidate create the best chance to find memset opportunity.
699 unsigned j = 0;
700 for (j = i + 1; j < e; ++j)
701 IndexQueue.push_back(j);
702 for (j = i; j > 0; --j)
703 IndexQueue.push_back(j - 1);
704
705 for (auto &k : IndexQueue) {
706 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
707 Value *SecondStorePtr = SL[k]->getPointerOperand();
708 const SCEVAddRecExpr *SecondStoreEv =
709 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
710 APInt SecondStride = getStoreStride(SecondStoreEv);
711
712 if (FirstStride != SecondStride)
713 continue;
714
715 Value *SecondStoredVal = SL[k]->getValueOperand();
716 Value *SecondSplatValue = nullptr;
717 Constant *SecondPatternValue = nullptr;
718
719 if (For == ForMemset::Yes)
720 SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL);
721 else
722 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
723
724 assert((SecondSplatValue || SecondPatternValue) &&
725 "Expected either splat value or pattern value.");
726
727 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
728 if (For == ForMemset::Yes) {
729 if (isa<UndefValue>(FirstSplatValue))
730 FirstSplatValue = SecondSplatValue;
731 if (FirstSplatValue != SecondSplatValue)
732 continue;
733 } else {
734 if (isa<UndefValue>(FirstPatternValue))
735 FirstPatternValue = SecondPatternValue;
736 if (FirstPatternValue != SecondPatternValue)
737 continue;
738 }
739 Tails.insert(SL[k]);
740 Heads.insert(SL[i]);
741 ConsecutiveChain[SL[i]] = SL[k];
742 break;
743 }
744 }
745 }
746
747 // We may run into multiple chains that merge into a single chain. We mark the
748 // stores that we transformed so that we don't visit the same store twice.
749 SmallPtrSet<Value *, 16> TransformedStores;
750 bool Changed = false;
751
752 // For stores that start but don't end a link in the chain:
753 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
754 it != e; ++it) {
755 if (Tails.count(*it))
756 continue;
757
758 // We found a store instr that starts a chain. Now follow the chain and try
759 // to transform it.
760 SmallPtrSet<Instruction *, 8> AdjacentStores;
761 StoreInst *I = *it;
762
763 StoreInst *HeadStore = I;
764 unsigned StoreSize = 0;
765
766 // Collect the chain into a list.
767 while (Tails.count(I) || Heads.count(I)) {
768 if (TransformedStores.count(I))
769 break;
770 AdjacentStores.insert(I);
771
772 StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType());
773 // Move to the next value in the chain.
774 I = ConsecutiveChain[I];
775 }
776
777 Value *StoredVal = HeadStore->getValueOperand();
778 Value *StorePtr = HeadStore->getPointerOperand();
779 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
780 APInt Stride = getStoreStride(StoreEv);
781
782 // Check to see if the stride matches the size of the stores. If so, then
783 // we know that every byte is touched in the loop.
784 if (StoreSize != Stride && StoreSize != -Stride)
785 continue;
786
787 bool NegStride = StoreSize == -Stride;
788
789 if (processLoopStridedStore(StorePtr, StoreSize,
790 MaybeAlign(HeadStore->getAlignment()),
791 StoredVal, HeadStore, AdjacentStores, StoreEv,
792 BECount, NegStride)) {
793 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
794 Changed = true;
795 }
796 }
797
798 return Changed;
799 }
800
801 /// processLoopMemIntrinsic - Template function for calling different processor
802 /// functions based on mem instrinsic type.
803 template <typename MemInst>
processLoopMemIntrinsic(BasicBlock * BB,bool (LoopIdiomRecognize::* Processor)(MemInst *,const SCEV *),const SCEV * BECount)804 bool LoopIdiomRecognize::processLoopMemIntrinsic(
805 BasicBlock *BB,
806 bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *),
807 const SCEV *BECount) {
808 bool MadeChange = false;
809 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
810 Instruction *Inst = &*I++;
811 // Look for memory instructions, which may be optimized to a larger one.
812 if (MemInst *MI = dyn_cast<MemInst>(Inst)) {
813 WeakTrackingVH InstPtr(&*I);
814 if (!(this->*Processor)(MI, BECount))
815 continue;
816 MadeChange = true;
817
818 // If processing the instruction invalidated our iterator, start over from
819 // the top of the block.
820 if (!InstPtr)
821 I = BB->begin();
822 }
823 }
824 return MadeChange;
825 }
826
827 /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy
processLoopMemCpy(MemCpyInst * MCI,const SCEV * BECount)828 bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
829 const SCEV *BECount) {
830 // We can only handle non-volatile memcpys with a constant size.
831 if (MCI->isVolatile() || !isa<ConstantInt>(MCI->getLength()))
832 return false;
833
834 // If we're not allowed to hack on memcpy, we fail.
835 if ((!HasMemcpy && !isa<MemCpyInlineInst>(MCI)) || DisableLIRP::Memcpy)
836 return false;
837
838 Value *Dest = MCI->getDest();
839 Value *Source = MCI->getSource();
840 if (!Dest || !Source)
841 return false;
842
843 // See if the load and store pointer expressions are AddRec like {base,+,1} on
844 // the current loop, which indicates a strided load and store. If we have
845 // something else, it's a random load or store we can't handle.
846 const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest));
847 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
848 return false;
849 const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source));
850 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
851 return false;
852
853 // Reject memcpys that are so large that they overflow an unsigned.
854 uint64_t SizeInBytes = cast<ConstantInt>(MCI->getLength())->getZExtValue();
855 if ((SizeInBytes >> 32) != 0)
856 return false;
857
858 // Check if the stride matches the size of the memcpy. If so, then we know
859 // that every byte is touched in the loop.
860 const SCEVConstant *StoreStride =
861 dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
862 const SCEVConstant *LoadStride =
863 dyn_cast<SCEVConstant>(LoadEv->getOperand(1));
864 if (!StoreStride || !LoadStride)
865 return false;
866
867 APInt StoreStrideValue = StoreStride->getAPInt();
868 APInt LoadStrideValue = LoadStride->getAPInt();
869 // Huge stride value - give up
870 if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64)
871 return false;
872
873 if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) {
874 ORE.emit([&]() {
875 return OptimizationRemarkMissed(DEBUG_TYPE, "SizeStrideUnequal", MCI)
876 << ore::NV("Inst", "memcpy") << " in "
877 << ore::NV("Function", MCI->getFunction())
878 << " function will not be hoised: "
879 << ore::NV("Reason", "memcpy size is not equal to stride");
880 });
881 return false;
882 }
883
884 int64_t StoreStrideInt = StoreStrideValue.getSExtValue();
885 int64_t LoadStrideInt = LoadStrideValue.getSExtValue();
886 // Check if the load stride matches the store stride.
887 if (StoreStrideInt != LoadStrideInt)
888 return false;
889
890 return processLoopStoreOfLoopLoad(Dest, Source, (unsigned)SizeInBytes,
891 MCI->getDestAlign(), MCI->getSourceAlign(),
892 MCI, MCI, StoreEv, LoadEv, BECount);
893 }
894
895 /// processLoopMemSet - See if this memset can be promoted to a large memset.
processLoopMemSet(MemSetInst * MSI,const SCEV * BECount)896 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
897 const SCEV *BECount) {
898 // We can only handle non-volatile memsets with a constant size.
899 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
900 return false;
901
902 // If we're not allowed to hack on memset, we fail.
903 if (!HasMemset || DisableLIRP::Memset)
904 return false;
905
906 Value *Pointer = MSI->getDest();
907
908 // See if the pointer expression is an AddRec like {base,+,1} on the current
909 // loop, which indicates a strided store. If we have something else, it's a
910 // random store we can't handle.
911 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
912 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
913 return false;
914
915 // Reject memsets that are so large that they overflow an unsigned.
916 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
917 if ((SizeInBytes >> 32) != 0)
918 return false;
919
920 // Check to see if the stride matches the size of the memset. If so, then we
921 // know that every byte is touched in the loop.
922 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
923 if (!ConstStride)
924 return false;
925
926 APInt Stride = ConstStride->getAPInt();
927 if (SizeInBytes != Stride && SizeInBytes != -Stride)
928 return false;
929
930 // Verify that the memset value is loop invariant. If not, we can't promote
931 // the memset.
932 Value *SplatValue = MSI->getValue();
933 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
934 return false;
935
936 SmallPtrSet<Instruction *, 1> MSIs;
937 MSIs.insert(MSI);
938 bool NegStride = SizeInBytes == -Stride;
939 return processLoopStridedStore(
940 Pointer, (unsigned)SizeInBytes, MaybeAlign(MSI->getDestAlignment()),
941 SplatValue, MSI, MSIs, Ev, BECount, NegStride, /*IsLoopMemset=*/true);
942 }
943
944 /// mayLoopAccessLocation - Return true if the specified loop might access the
945 /// specified pointer location, which is a loop-strided access. The 'Access'
946 /// argument specifies what the verboten forms of access are (read or write).
947 static bool
mayLoopAccessLocation(Value * Ptr,ModRefInfo Access,Loop * L,const SCEV * BECount,unsigned StoreSize,AliasAnalysis & AA,SmallPtrSetImpl<Instruction * > & IgnoredStores)948 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
949 const SCEV *BECount, unsigned StoreSize,
950 AliasAnalysis &AA,
951 SmallPtrSetImpl<Instruction *> &IgnoredStores) {
952 // Get the location that may be stored across the loop. Since the access is
953 // strided positively through memory, we say that the modified location starts
954 // at the pointer and has infinite size.
955 LocationSize AccessSize = LocationSize::afterPointer();
956
957 // If the loop iterates a fixed number of times, we can refine the access size
958 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
959 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
960 AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) *
961 StoreSize);
962
963 // TODO: For this to be really effective, we have to dive into the pointer
964 // operand in the store. Store to &A[i] of 100 will always return may alias
965 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
966 // which will then no-alias a store to &A[100].
967 MemoryLocation StoreLoc(Ptr, AccessSize);
968
969 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
970 ++BI)
971 for (Instruction &I : **BI)
972 if (IgnoredStores.count(&I) == 0 &&
973 isModOrRefSet(
974 intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access)))
975 return true;
976
977 return false;
978 }
979
980 // If we have a negative stride, Start refers to the end of the memory location
981 // we're trying to memset. Therefore, we need to recompute the base pointer,
982 // which is just Start - BECount*Size.
getStartForNegStride(const SCEV * Start,const SCEV * BECount,Type * IntPtr,unsigned StoreSize,ScalarEvolution * SE)983 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
984 Type *IntPtr, unsigned StoreSize,
985 ScalarEvolution *SE) {
986 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
987 if (StoreSize != 1)
988 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
989 SCEV::FlagNUW);
990 return SE->getMinusSCEV(Start, Index);
991 }
992
993 /// Compute the number of bytes as a SCEV from the backedge taken count.
994 ///
995 /// This also maps the SCEV into the provided type and tries to handle the
996 /// computation in a way that will fold cleanly.
getNumBytes(const SCEV * BECount,Type * IntPtr,unsigned StoreSize,Loop * CurLoop,const DataLayout * DL,ScalarEvolution * SE)997 static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr,
998 unsigned StoreSize, Loop *CurLoop,
999 const DataLayout *DL, ScalarEvolution *SE) {
1000 const SCEV *NumBytesS;
1001 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1002 // pointer size if it isn't already.
1003 //
1004 // If we're going to need to zero extend the BE count, check if we can add
1005 // one to it prior to zero extending without overflow. Provided this is safe,
1006 // it allows better simplification of the +1.
1007 if (DL->getTypeSizeInBits(BECount->getType()).getFixedSize() <
1008 DL->getTypeSizeInBits(IntPtr).getFixedSize() &&
1009 SE->isLoopEntryGuardedByCond(
1010 CurLoop, ICmpInst::ICMP_NE, BECount,
1011 SE->getNegativeSCEV(SE->getOne(BECount->getType())))) {
1012 NumBytesS = SE->getZeroExtendExpr(
1013 SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW),
1014 IntPtr);
1015 } else {
1016 NumBytesS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr),
1017 SE->getOne(IntPtr), SCEV::FlagNUW);
1018 }
1019
1020 // And scale it based on the store size.
1021 if (StoreSize != 1) {
1022 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
1023 SCEV::FlagNUW);
1024 }
1025 return NumBytesS;
1026 }
1027
1028 /// processLoopStridedStore - We see a strided store of some value. If we can
1029 /// transform this into a memset or memset_pattern in the loop preheader, do so.
processLoopStridedStore(Value * DestPtr,unsigned StoreSize,MaybeAlign StoreAlignment,Value * StoredVal,Instruction * TheStore,SmallPtrSetImpl<Instruction * > & Stores,const SCEVAddRecExpr * Ev,const SCEV * BECount,bool NegStride,bool IsLoopMemset)1030 bool LoopIdiomRecognize::processLoopStridedStore(
1031 Value *DestPtr, unsigned StoreSize, MaybeAlign StoreAlignment,
1032 Value *StoredVal, Instruction *TheStore,
1033 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
1034 const SCEV *BECount, bool NegStride, bool IsLoopMemset) {
1035 Value *SplatValue = isBytewiseValue(StoredVal, *DL);
1036 Constant *PatternValue = nullptr;
1037
1038 if (!SplatValue)
1039 PatternValue = getMemSetPatternValue(StoredVal, DL);
1040
1041 assert((SplatValue || PatternValue) &&
1042 "Expected either splat value or pattern value.");
1043
1044 // The trip count of the loop and the base pointer of the addrec SCEV is
1045 // guaranteed to be loop invariant, which means that it should dominate the
1046 // header. This allows us to insert code for it in the preheader.
1047 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
1048 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1049 IRBuilder<> Builder(Preheader->getTerminator());
1050 SCEVExpander Expander(*SE, *DL, "loop-idiom");
1051 SCEVExpanderCleaner ExpCleaner(Expander, *DT);
1052
1053 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
1054 Type *IntIdxTy = DL->getIndexType(DestPtr->getType());
1055
1056 bool Changed = false;
1057 const SCEV *Start = Ev->getStart();
1058 // Handle negative strided loops.
1059 if (NegStride)
1060 Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSize, SE);
1061
1062 // TODO: ideally we should still be able to generate memset if SCEV expander
1063 // is taught to generate the dependencies at the latest point.
1064 if (!isSafeToExpand(Start, *SE))
1065 return Changed;
1066
1067 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
1068 // this into a memset in the loop preheader now if we want. However, this
1069 // would be unsafe to do if there is anything else in the loop that may read
1070 // or write to the aliased location. Check for any overlap by generating the
1071 // base pointer and checking the region.
1072 Value *BasePtr =
1073 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
1074
1075 // From here on out, conservatively report to the pass manager that we've
1076 // changed the IR, even if we later clean up these added instructions. There
1077 // may be structural differences e.g. in the order of use lists not accounted
1078 // for in just a textual dump of the IR. This is written as a variable, even
1079 // though statically all the places this dominates could be replaced with
1080 // 'true', with the hope that anyone trying to be clever / "more precise" with
1081 // the return value will read this comment, and leave them alone.
1082 Changed = true;
1083
1084 if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1085 StoreSize, *AA, Stores))
1086 return Changed;
1087
1088 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
1089 return Changed;
1090
1091 // Okay, everything looks good, insert the memset.
1092
1093 const SCEV *NumBytesS =
1094 getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE);
1095
1096 // TODO: ideally we should still be able to generate memset if SCEV expander
1097 // is taught to generate the dependencies at the latest point.
1098 if (!isSafeToExpand(NumBytesS, *SE))
1099 return Changed;
1100
1101 Value *NumBytes =
1102 Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1103
1104 CallInst *NewCall;
1105 if (SplatValue) {
1106 NewCall = Builder.CreateMemSet(BasePtr, SplatValue, NumBytes,
1107 MaybeAlign(StoreAlignment));
1108 } else {
1109 // Everything is emitted in default address space
1110 Type *Int8PtrTy = DestInt8PtrTy;
1111
1112 Module *M = TheStore->getModule();
1113 StringRef FuncName = "memset_pattern16";
1114 FunctionCallee MSP = M->getOrInsertFunction(FuncName, Builder.getVoidTy(),
1115 Int8PtrTy, Int8PtrTy, IntIdxTy);
1116 inferLibFuncAttributes(M, FuncName, *TLI);
1117
1118 // Otherwise we should form a memset_pattern16. PatternValue is known to be
1119 // an constant array of 16-bytes. Plop the value into a mergable global.
1120 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1121 GlobalValue::PrivateLinkage,
1122 PatternValue, ".memset_pattern");
1123 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
1124 GV->setAlignment(Align(16));
1125 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
1126 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
1127 }
1128 NewCall->setDebugLoc(TheStore->getDebugLoc());
1129
1130 if (MSSAU) {
1131 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1132 NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1133 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1134 }
1135
1136 LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1137 << " from store to: " << *Ev << " at: " << *TheStore
1138 << "\n");
1139
1140 ORE.emit([&]() {
1141 return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStridedStore",
1142 NewCall->getDebugLoc(), Preheader)
1143 << "Transformed loop-strided store in "
1144 << ore::NV("Function", TheStore->getFunction())
1145 << " function into a call to "
1146 << ore::NV("NewFunction", NewCall->getCalledFunction())
1147 << "() intrinsic";
1148 });
1149
1150 // Okay, the memset has been formed. Zap the original store and anything that
1151 // feeds into it.
1152 for (auto *I : Stores) {
1153 if (MSSAU)
1154 MSSAU->removeMemoryAccess(I, true);
1155 deleteDeadInstruction(I);
1156 }
1157 if (MSSAU && VerifyMemorySSA)
1158 MSSAU->getMemorySSA()->verifyMemorySSA();
1159 ++NumMemSet;
1160 ExpCleaner.markResultUsed();
1161 return true;
1162 }
1163
1164 /// If the stored value is a strided load in the same loop with the same stride
1165 /// this may be transformable into a memcpy. This kicks in for stuff like
1166 /// for (i) A[i] = B[i];
processLoopStoreOfLoopLoad(StoreInst * SI,const SCEV * BECount)1167 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
1168 const SCEV *BECount) {
1169 assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.");
1170
1171 Value *StorePtr = SI->getPointerOperand();
1172 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
1173 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
1174
1175 // The store must be feeding a non-volatile load.
1176 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1177 assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.");
1178
1179 // See if the pointer expression is an AddRec like {base,+,1} on the current
1180 // loop, which indicates a strided load. If we have something else, it's a
1181 // random load we can't handle.
1182 Value *LoadPtr = LI->getPointerOperand();
1183 const SCEVAddRecExpr *LoadEv = cast<SCEVAddRecExpr>(SE->getSCEV(LoadPtr));
1184 return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSize,
1185 SI->getAlign(), LI->getAlign(), SI, LI,
1186 StoreEv, LoadEv, BECount);
1187 }
1188
processLoopStoreOfLoopLoad(Value * DestPtr,Value * SourcePtr,unsigned StoreSize,MaybeAlign StoreAlign,MaybeAlign LoadAlign,Instruction * TheStore,Instruction * TheLoad,const SCEVAddRecExpr * StoreEv,const SCEVAddRecExpr * LoadEv,const SCEV * BECount)1189 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
1190 Value *DestPtr, Value *SourcePtr, unsigned StoreSize, MaybeAlign StoreAlign,
1191 MaybeAlign LoadAlign, Instruction *TheStore, Instruction *TheLoad,
1192 const SCEVAddRecExpr *StoreEv, const SCEVAddRecExpr *LoadEv,
1193 const SCEV *BECount) {
1194
1195 // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to
1196 // conservatively bail here, since otherwise we may have to transform
1197 // llvm.memcpy.inline into llvm.memcpy which is illegal.
1198 if (isa<MemCpyInlineInst>(TheStore))
1199 return false;
1200
1201 // The trip count of the loop and the base pointer of the addrec SCEV is
1202 // guaranteed to be loop invariant, which means that it should dominate the
1203 // header. This allows us to insert code for it in the preheader.
1204 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1205 IRBuilder<> Builder(Preheader->getTerminator());
1206 SCEVExpander Expander(*SE, *DL, "loop-idiom");
1207
1208 SCEVExpanderCleaner ExpCleaner(Expander, *DT);
1209
1210 bool Changed = false;
1211 const SCEV *StrStart = StoreEv->getStart();
1212 unsigned StrAS = DestPtr->getType()->getPointerAddressSpace();
1213 Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS));
1214
1215 APInt Stride = getStoreStride(StoreEv);
1216 bool NegStride = StoreSize == -Stride;
1217
1218 // Handle negative strided loops.
1219 if (NegStride)
1220 StrStart = getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSize, SE);
1221
1222 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1223 // this into a memcpy in the loop preheader now if we want. However, this
1224 // would be unsafe to do if there is anything else in the loop that may read
1225 // or write the memory region we're storing to. This includes the load that
1226 // feeds the stores. Check for an alias by generating the base address and
1227 // checking everything.
1228 Value *StoreBasePtr = Expander.expandCodeFor(
1229 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
1230
1231 // From here on out, conservatively report to the pass manager that we've
1232 // changed the IR, even if we later clean up these added instructions. There
1233 // may be structural differences e.g. in the order of use lists not accounted
1234 // for in just a textual dump of the IR. This is written as a variable, even
1235 // though statically all the places this dominates could be replaced with
1236 // 'true', with the hope that anyone trying to be clever / "more precise" with
1237 // the return value will read this comment, and leave them alone.
1238 Changed = true;
1239
1240 SmallPtrSet<Instruction *, 2> Stores;
1241 Stores.insert(TheStore);
1242
1243 bool IsMemCpy = isa<MemCpyInst>(TheStore);
1244 const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store";
1245
1246 bool UseMemMove =
1247 mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1248 StoreSize, *AA, Stores);
1249 if (UseMemMove) {
1250 // For memmove case it's not enough to guarantee that loop doesn't access
1251 // TheStore and TheLoad. Additionally we need to make sure that TheStore is
1252 // the only user of TheLoad.
1253 if (!TheLoad->hasOneUse())
1254 return Changed;
1255 Stores.insert(TheLoad);
1256 if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop,
1257 BECount, StoreSize, *AA, Stores)) {
1258 ORE.emit([&]() {
1259 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessStore",
1260 TheStore)
1261 << ore::NV("Inst", InstRemark) << " in "
1262 << ore::NV("Function", TheStore->getFunction())
1263 << " function will not be hoisted: "
1264 << ore::NV("Reason", "The loop may access store location");
1265 });
1266 return Changed;
1267 }
1268 Stores.erase(TheLoad);
1269 }
1270
1271 const SCEV *LdStart = LoadEv->getStart();
1272 unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace();
1273
1274 // Handle negative strided loops.
1275 if (NegStride)
1276 LdStart = getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSize, SE);
1277
1278 // For a memcpy, we have to make sure that the input array is not being
1279 // mutated by the loop.
1280 Value *LoadBasePtr = Expander.expandCodeFor(
1281 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
1282
1283 // If the store is a memcpy instruction, we must check if it will write to
1284 // the load memory locations. So remove it from the ignored stores.
1285 if (IsMemCpy)
1286 Stores.erase(TheStore);
1287 if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
1288 StoreSize, *AA, Stores)) {
1289 ORE.emit([&]() {
1290 return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessLoad", TheLoad)
1291 << ore::NV("Inst", InstRemark) << " in "
1292 << ore::NV("Function", TheStore->getFunction())
1293 << " function will not be hoisted: "
1294 << ore::NV("Reason", "The loop may access load location");
1295 });
1296 return Changed;
1297 }
1298 if (UseMemMove) {
1299 // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr for
1300 // negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr.
1301 int64_t LoadOff = 0, StoreOff = 0;
1302 const Value *BP1 = llvm::GetPointerBaseWithConstantOffset(
1303 LoadBasePtr->stripPointerCasts(), LoadOff, *DL);
1304 const Value *BP2 = llvm::GetPointerBaseWithConstantOffset(
1305 StoreBasePtr->stripPointerCasts(), StoreOff, *DL);
1306 int64_t LoadSize =
1307 DL->getTypeSizeInBits(TheLoad->getType()).getFixedSize() / 8;
1308 if (BP1 != BP2 || LoadSize != int64_t(StoreSize))
1309 return Changed;
1310 if ((!NegStride && LoadOff < StoreOff + int64_t(StoreSize)) ||
1311 (NegStride && LoadOff + LoadSize > StoreOff))
1312 return Changed;
1313 }
1314
1315 if (avoidLIRForMultiBlockLoop())
1316 return Changed;
1317
1318 // Okay, everything is safe, we can transform this!
1319
1320 const SCEV *NumBytesS =
1321 getNumBytes(BECount, IntIdxTy, StoreSize, CurLoop, DL, SE);
1322
1323 Value *NumBytes =
1324 Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator());
1325
1326 CallInst *NewCall = nullptr;
1327 // Check whether to generate an unordered atomic memcpy:
1328 // If the load or store are atomic, then they must necessarily be unordered
1329 // by previous checks.
1330 if (!TheStore->isAtomic() && !TheLoad->isAtomic()) {
1331 if (UseMemMove)
1332 NewCall = Builder.CreateMemMove(StoreBasePtr, StoreAlign, LoadBasePtr,
1333 LoadAlign, NumBytes);
1334 else
1335 NewCall = Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr,
1336 LoadAlign, NumBytes);
1337 } else {
1338 // For now don't support unordered atomic memmove.
1339 if (UseMemMove)
1340 return Changed;
1341 // We cannot allow unaligned ops for unordered load/store, so reject
1342 // anything where the alignment isn't at least the element size.
1343 assert((StoreAlign.hasValue() && LoadAlign.hasValue()) &&
1344 "Expect unordered load/store to have align.");
1345 if (StoreAlign.getValue() < StoreSize || LoadAlign.getValue() < StoreSize)
1346 return Changed;
1347
1348 // If the element.atomic memcpy is not lowered into explicit
1349 // loads/stores later, then it will be lowered into an element-size
1350 // specific lib call. If the lib call doesn't exist for our store size, then
1351 // we shouldn't generate the memcpy.
1352 if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize())
1353 return Changed;
1354
1355 // Create the call.
1356 // Note that unordered atomic loads/stores are *required* by the spec to
1357 // have an alignment but non-atomic loads/stores may not.
1358 NewCall = Builder.CreateElementUnorderedAtomicMemCpy(
1359 StoreBasePtr, StoreAlign.getValue(), LoadBasePtr, LoadAlign.getValue(),
1360 NumBytes, StoreSize);
1361 }
1362 NewCall->setDebugLoc(TheStore->getDebugLoc());
1363
1364 if (MSSAU) {
1365 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1366 NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator);
1367 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1368 }
1369
1370 LLVM_DEBUG(dbgs() << " Formed new call: " << *NewCall << "\n"
1371 << " from load ptr=" << *LoadEv << " at: " << *TheLoad
1372 << "\n"
1373 << " from store ptr=" << *StoreEv << " at: " << *TheStore
1374 << "\n");
1375
1376 ORE.emit([&]() {
1377 return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStoreOfLoopLoad",
1378 NewCall->getDebugLoc(), Preheader)
1379 << "Formed a call to "
1380 << ore::NV("NewFunction", NewCall->getCalledFunction())
1381 << "() intrinsic from " << ore::NV("Inst", InstRemark)
1382 << " instruction in " << ore::NV("Function", TheStore->getFunction())
1383 << " function";
1384 });
1385
1386 // Okay, the memcpy has been formed. Zap the original store and anything that
1387 // feeds into it.
1388 if (MSSAU)
1389 MSSAU->removeMemoryAccess(TheStore, true);
1390 deleteDeadInstruction(TheStore);
1391 if (MSSAU && VerifyMemorySSA)
1392 MSSAU->getMemorySSA()->verifyMemorySSA();
1393 if (UseMemMove)
1394 ++NumMemMove;
1395 else
1396 ++NumMemCpy;
1397 ExpCleaner.markResultUsed();
1398 return true;
1399 }
1400
1401 // When compiling for codesize we avoid idiom recognition for a multi-block loop
1402 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
1403 //
avoidLIRForMultiBlockLoop(bool IsMemset,bool IsLoopMemset)1404 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
1405 bool IsLoopMemset) {
1406 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
1407 if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) {
1408 LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
1409 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1410 << " avoided: multi-block top-level loop\n");
1411 return true;
1412 }
1413 }
1414
1415 return false;
1416 }
1417
runOnNoncountableLoop()1418 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1419 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F["
1420 << CurLoop->getHeader()->getParent()->getName()
1421 << "] Noncountable Loop %"
1422 << CurLoop->getHeader()->getName() << "\n");
1423
1424 return recognizePopcount() || recognizeAndInsertFFS() ||
1425 recognizeShiftUntilBitTest() || recognizeShiftUntilZero();
1426 }
1427
1428 /// Check if the given conditional branch is based on the comparison between
1429 /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is
1430 /// true), the control yields to the loop entry. If the branch matches the
1431 /// behavior, the variable involved in the comparison is returned. This function
1432 /// will be called to see if the precondition and postcondition of the loop are
1433 /// in desirable form.
matchCondition(BranchInst * BI,BasicBlock * LoopEntry,bool JmpOnZero=false)1434 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
1435 bool JmpOnZero = false) {
1436 if (!BI || !BI->isConditional())
1437 return nullptr;
1438
1439 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1440 if (!Cond)
1441 return nullptr;
1442
1443 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1444 if (!CmpZero || !CmpZero->isZero())
1445 return nullptr;
1446
1447 BasicBlock *TrueSucc = BI->getSuccessor(0);
1448 BasicBlock *FalseSucc = BI->getSuccessor(1);
1449 if (JmpOnZero)
1450 std::swap(TrueSucc, FalseSucc);
1451
1452 ICmpInst::Predicate Pred = Cond->getPredicate();
1453 if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) ||
1454 (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry))
1455 return Cond->getOperand(0);
1456
1457 return nullptr;
1458 }
1459
1460 // Check if the recurrence variable `VarX` is in the right form to create
1461 // the idiom. Returns the value coerced to a PHINode if so.
getRecurrenceVar(Value * VarX,Instruction * DefX,BasicBlock * LoopEntry)1462 static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
1463 BasicBlock *LoopEntry) {
1464 auto *PhiX = dyn_cast<PHINode>(VarX);
1465 if (PhiX && PhiX->getParent() == LoopEntry &&
1466 (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX))
1467 return PhiX;
1468 return nullptr;
1469 }
1470
1471 /// Return true iff the idiom is detected in the loop.
1472 ///
1473 /// Additionally:
1474 /// 1) \p CntInst is set to the instruction counting the population bit.
1475 /// 2) \p CntPhi is set to the corresponding phi node.
1476 /// 3) \p Var is set to the value whose population bits are being counted.
1477 ///
1478 /// The core idiom we are trying to detect is:
1479 /// \code
1480 /// if (x0 != 0)
1481 /// goto loop-exit // the precondition of the loop
1482 /// cnt0 = init-val;
1483 /// do {
1484 /// x1 = phi (x0, x2);
1485 /// cnt1 = phi(cnt0, cnt2);
1486 ///
1487 /// cnt2 = cnt1 + 1;
1488 /// ...
1489 /// x2 = x1 & (x1 - 1);
1490 /// ...
1491 /// } while(x != 0);
1492 ///
1493 /// loop-exit:
1494 /// \endcode
detectPopcountIdiom(Loop * CurLoop,BasicBlock * PreCondBB,Instruction * & CntInst,PHINode * & CntPhi,Value * & Var)1495 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1496 Instruction *&CntInst, PHINode *&CntPhi,
1497 Value *&Var) {
1498 // step 1: Check to see if the look-back branch match this pattern:
1499 // "if (a!=0) goto loop-entry".
1500 BasicBlock *LoopEntry;
1501 Instruction *DefX2, *CountInst;
1502 Value *VarX1, *VarX0;
1503 PHINode *PhiX, *CountPhi;
1504
1505 DefX2 = CountInst = nullptr;
1506 VarX1 = VarX0 = nullptr;
1507 PhiX = CountPhi = nullptr;
1508 LoopEntry = *(CurLoop->block_begin());
1509
1510 // step 1: Check if the loop-back branch is in desirable form.
1511 {
1512 if (Value *T = matchCondition(
1513 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1514 DefX2 = dyn_cast<Instruction>(T);
1515 else
1516 return false;
1517 }
1518
1519 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1520 {
1521 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1522 return false;
1523
1524 BinaryOperator *SubOneOp;
1525
1526 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1527 VarX1 = DefX2->getOperand(1);
1528 else {
1529 VarX1 = DefX2->getOperand(0);
1530 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1531 }
1532 if (!SubOneOp || SubOneOp->getOperand(0) != VarX1)
1533 return false;
1534
1535 ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1));
1536 if (!Dec ||
1537 !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1538 (SubOneOp->getOpcode() == Instruction::Add &&
1539 Dec->isMinusOne()))) {
1540 return false;
1541 }
1542 }
1543
1544 // step 3: Check the recurrence of variable X
1545 PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry);
1546 if (!PhiX)
1547 return false;
1548
1549 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1550 {
1551 CountInst = nullptr;
1552 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1553 IterE = LoopEntry->end();
1554 Iter != IterE; Iter++) {
1555 Instruction *Inst = &*Iter;
1556 if (Inst->getOpcode() != Instruction::Add)
1557 continue;
1558
1559 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1560 if (!Inc || !Inc->isOne())
1561 continue;
1562
1563 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1564 if (!Phi)
1565 continue;
1566
1567 // Check if the result of the instruction is live of the loop.
1568 bool LiveOutLoop = false;
1569 for (User *U : Inst->users()) {
1570 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1571 LiveOutLoop = true;
1572 break;
1573 }
1574 }
1575
1576 if (LiveOutLoop) {
1577 CountInst = Inst;
1578 CountPhi = Phi;
1579 break;
1580 }
1581 }
1582
1583 if (!CountInst)
1584 return false;
1585 }
1586
1587 // step 5: check if the precondition is in this form:
1588 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1589 {
1590 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1591 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1592 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1593 return false;
1594
1595 CntInst = CountInst;
1596 CntPhi = CountPhi;
1597 Var = T;
1598 }
1599
1600 return true;
1601 }
1602
1603 /// Return true if the idiom is detected in the loop.
1604 ///
1605 /// Additionally:
1606 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1607 /// or nullptr if there is no such.
1608 /// 2) \p CntPhi is set to the corresponding phi node
1609 /// or nullptr if there is no such.
1610 /// 3) \p Var is set to the value whose CTLZ could be used.
1611 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1612 ///
1613 /// The core idiom we are trying to detect is:
1614 /// \code
1615 /// if (x0 == 0)
1616 /// goto loop-exit // the precondition of the loop
1617 /// cnt0 = init-val;
1618 /// do {
1619 /// x = phi (x0, x.next); //PhiX
1620 /// cnt = phi(cnt0, cnt.next);
1621 ///
1622 /// cnt.next = cnt + 1;
1623 /// ...
1624 /// x.next = x >> 1; // DefX
1625 /// ...
1626 /// } while(x.next != 0);
1627 ///
1628 /// loop-exit:
1629 /// \endcode
detectShiftUntilZeroIdiom(Loop * CurLoop,const DataLayout & DL,Intrinsic::ID & IntrinID,Value * & InitX,Instruction * & CntInst,PHINode * & CntPhi,Instruction * & DefX)1630 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL,
1631 Intrinsic::ID &IntrinID, Value *&InitX,
1632 Instruction *&CntInst, PHINode *&CntPhi,
1633 Instruction *&DefX) {
1634 BasicBlock *LoopEntry;
1635 Value *VarX = nullptr;
1636
1637 DefX = nullptr;
1638 CntInst = nullptr;
1639 CntPhi = nullptr;
1640 LoopEntry = *(CurLoop->block_begin());
1641
1642 // step 1: Check if the loop-back branch is in desirable form.
1643 if (Value *T = matchCondition(
1644 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1645 DefX = dyn_cast<Instruction>(T);
1646 else
1647 return false;
1648
1649 // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1"
1650 if (!DefX || !DefX->isShift())
1651 return false;
1652 IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz :
1653 Intrinsic::ctlz;
1654 ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1));
1655 if (!Shft || !Shft->isOne())
1656 return false;
1657 VarX = DefX->getOperand(0);
1658
1659 // step 3: Check the recurrence of variable X
1660 PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
1661 if (!PhiX)
1662 return false;
1663
1664 InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader());
1665
1666 // Make sure the initial value can't be negative otherwise the ashr in the
1667 // loop might never reach zero which would make the loop infinite.
1668 if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL))
1669 return false;
1670
1671 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1672 // or cnt.next = cnt + -1.
1673 // TODO: We can skip the step. If loop trip count is known (CTLZ),
1674 // then all uses of "cnt.next" could be optimized to the trip count
1675 // plus "cnt0". Currently it is not optimized.
1676 // This step could be used to detect POPCNT instruction:
1677 // cnt.next = cnt + (x.next & 1)
1678 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1679 IterE = LoopEntry->end();
1680 Iter != IterE; Iter++) {
1681 Instruction *Inst = &*Iter;
1682 if (Inst->getOpcode() != Instruction::Add)
1683 continue;
1684
1685 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1686 if (!Inc || (!Inc->isOne() && !Inc->isMinusOne()))
1687 continue;
1688
1689 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1690 if (!Phi)
1691 continue;
1692
1693 CntInst = Inst;
1694 CntPhi = Phi;
1695 break;
1696 }
1697 if (!CntInst)
1698 return false;
1699
1700 return true;
1701 }
1702
1703 /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop
1704 /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new
1705 /// trip count returns true; otherwise, returns false.
recognizeAndInsertFFS()1706 bool LoopIdiomRecognize::recognizeAndInsertFFS() {
1707 // Give up if the loop has multiple blocks or multiple backedges.
1708 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1709 return false;
1710
1711 Intrinsic::ID IntrinID;
1712 Value *InitX;
1713 Instruction *DefX = nullptr;
1714 PHINode *CntPhi = nullptr;
1715 Instruction *CntInst = nullptr;
1716 // Help decide if transformation is profitable. For ShiftUntilZero idiom,
1717 // this is always 6.
1718 size_t IdiomCanonicalSize = 6;
1719
1720 if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX,
1721 CntInst, CntPhi, DefX))
1722 return false;
1723
1724 bool IsCntPhiUsedOutsideLoop = false;
1725 for (User *U : CntPhi->users())
1726 if (!CurLoop->contains(cast<Instruction>(U))) {
1727 IsCntPhiUsedOutsideLoop = true;
1728 break;
1729 }
1730 bool IsCntInstUsedOutsideLoop = false;
1731 for (User *U : CntInst->users())
1732 if (!CurLoop->contains(cast<Instruction>(U))) {
1733 IsCntInstUsedOutsideLoop = true;
1734 break;
1735 }
1736 // If both CntInst and CntPhi are used outside the loop the profitability
1737 // is questionable.
1738 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1739 return false;
1740
1741 // For some CPUs result of CTLZ(X) intrinsic is undefined
1742 // when X is 0. If we can not guarantee X != 0, we need to check this
1743 // when expand.
1744 bool ZeroCheck = false;
1745 // It is safe to assume Preheader exist as it was checked in
1746 // parent function RunOnLoop.
1747 BasicBlock *PH = CurLoop->getLoopPreheader();
1748
1749 // If we are using the count instruction outside the loop, make sure we
1750 // have a zero check as a precondition. Without the check the loop would run
1751 // one iteration for before any check of the input value. This means 0 and 1
1752 // would have identical behavior in the original loop and thus
1753 if (!IsCntPhiUsedOutsideLoop) {
1754 auto *PreCondBB = PH->getSinglePredecessor();
1755 if (!PreCondBB)
1756 return false;
1757 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1758 if (!PreCondBI)
1759 return false;
1760 if (matchCondition(PreCondBI, PH) != InitX)
1761 return false;
1762 ZeroCheck = true;
1763 }
1764
1765 // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always
1766 // profitable if we delete the loop.
1767
1768 // the loop has only 6 instructions:
1769 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1770 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1771 // %shr = ashr %n.addr.0, 1
1772 // %tobool = icmp eq %shr, 0
1773 // %inc = add nsw %i.0, 1
1774 // br i1 %tobool
1775
1776 const Value *Args[] = {InitX,
1777 ConstantInt::getBool(InitX->getContext(), ZeroCheck)};
1778
1779 // @llvm.dbg doesn't count as they have no semantic effect.
1780 auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug();
1781 uint32_t HeaderSize =
1782 std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end());
1783
1784 IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args);
1785 InstructionCost Cost =
1786 TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency);
1787 if (HeaderSize != IdiomCanonicalSize &&
1788 Cost > TargetTransformInfo::TCC_Basic)
1789 return false;
1790
1791 transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX,
1792 DefX->getDebugLoc(), ZeroCheck,
1793 IsCntPhiUsedOutsideLoop);
1794 return true;
1795 }
1796
1797 /// Recognizes a population count idiom in a non-countable loop.
1798 ///
1799 /// If detected, transforms the relevant code to issue the popcount intrinsic
1800 /// function call, and returns true; otherwise, returns false.
recognizePopcount()1801 bool LoopIdiomRecognize::recognizePopcount() {
1802 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1803 return false;
1804
1805 // Counting population are usually conducted by few arithmetic instructions.
1806 // Such instructions can be easily "absorbed" by vacant slots in a
1807 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1808 // in a compact loop.
1809
1810 // Give up if the loop has multiple blocks or multiple backedges.
1811 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1812 return false;
1813
1814 BasicBlock *LoopBody = *(CurLoop->block_begin());
1815 if (LoopBody->size() >= 20) {
1816 // The loop is too big, bail out.
1817 return false;
1818 }
1819
1820 // It should have a preheader containing nothing but an unconditional branch.
1821 BasicBlock *PH = CurLoop->getLoopPreheader();
1822 if (!PH || &PH->front() != PH->getTerminator())
1823 return false;
1824 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1825 if (!EntryBI || EntryBI->isConditional())
1826 return false;
1827
1828 // It should have a precondition block where the generated popcount intrinsic
1829 // function can be inserted.
1830 auto *PreCondBB = PH->getSinglePredecessor();
1831 if (!PreCondBB)
1832 return false;
1833 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1834 if (!PreCondBI || PreCondBI->isUnconditional())
1835 return false;
1836
1837 Instruction *CntInst;
1838 PHINode *CntPhi;
1839 Value *Val;
1840 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1841 return false;
1842
1843 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1844 return true;
1845 }
1846
createPopcntIntrinsic(IRBuilder<> & IRBuilder,Value * Val,const DebugLoc & DL)1847 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1848 const DebugLoc &DL) {
1849 Value *Ops[] = {Val};
1850 Type *Tys[] = {Val->getType()};
1851
1852 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1853 Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1854 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1855 CI->setDebugLoc(DL);
1856
1857 return CI;
1858 }
1859
createFFSIntrinsic(IRBuilder<> & IRBuilder,Value * Val,const DebugLoc & DL,bool ZeroCheck,Intrinsic::ID IID)1860 static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1861 const DebugLoc &DL, bool ZeroCheck,
1862 Intrinsic::ID IID) {
1863 Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)};
1864 Type *Tys[] = {Val->getType()};
1865
1866 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1867 Function *Func = Intrinsic::getDeclaration(M, IID, Tys);
1868 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1869 CI->setDebugLoc(DL);
1870
1871 return CI;
1872 }
1873
1874 /// Transform the following loop (Using CTLZ, CTTZ is similar):
1875 /// loop:
1876 /// CntPhi = PHI [Cnt0, CntInst]
1877 /// PhiX = PHI [InitX, DefX]
1878 /// CntInst = CntPhi + 1
1879 /// DefX = PhiX >> 1
1880 /// LOOP_BODY
1881 /// Br: loop if (DefX != 0)
1882 /// Use(CntPhi) or Use(CntInst)
1883 ///
1884 /// Into:
1885 /// If CntPhi used outside the loop:
1886 /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
1887 /// Count = CountPrev + 1
1888 /// else
1889 /// Count = BitWidth(InitX) - CTLZ(InitX)
1890 /// loop:
1891 /// CntPhi = PHI [Cnt0, CntInst]
1892 /// PhiX = PHI [InitX, DefX]
1893 /// PhiCount = PHI [Count, Dec]
1894 /// CntInst = CntPhi + 1
1895 /// DefX = PhiX >> 1
1896 /// Dec = PhiCount - 1
1897 /// LOOP_BODY
1898 /// Br: loop if (Dec != 0)
1899 /// Use(CountPrev + Cnt0) // Use(CntPhi)
1900 /// or
1901 /// Use(Count + Cnt0) // Use(CntInst)
1902 ///
1903 /// If LOOP_BODY is empty the loop will be deleted.
1904 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
transformLoopToCountable(Intrinsic::ID IntrinID,BasicBlock * Preheader,Instruction * CntInst,PHINode * CntPhi,Value * InitX,Instruction * DefX,const DebugLoc & DL,bool ZeroCheck,bool IsCntPhiUsedOutsideLoop)1905 void LoopIdiomRecognize::transformLoopToCountable(
1906 Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst,
1907 PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL,
1908 bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) {
1909 BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator());
1910
1911 // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block
1912 IRBuilder<> Builder(PreheaderBr);
1913 Builder.SetCurrentDebugLocation(DL);
1914
1915 // If there are no uses of CntPhi crate:
1916 // Count = BitWidth - CTLZ(InitX);
1917 // NewCount = Count;
1918 // If there are uses of CntPhi create:
1919 // NewCount = BitWidth - CTLZ(InitX >> 1);
1920 // Count = NewCount + 1;
1921 Value *InitXNext;
1922 if (IsCntPhiUsedOutsideLoop) {
1923 if (DefX->getOpcode() == Instruction::AShr)
1924 InitXNext = Builder.CreateAShr(InitX, 1);
1925 else if (DefX->getOpcode() == Instruction::LShr)
1926 InitXNext = Builder.CreateLShr(InitX, 1);
1927 else if (DefX->getOpcode() == Instruction::Shl) // cttz
1928 InitXNext = Builder.CreateShl(InitX, 1);
1929 else
1930 llvm_unreachable("Unexpected opcode!");
1931 } else
1932 InitXNext = InitX;
1933 Value *Count =
1934 createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID);
1935 Type *CountTy = Count->getType();
1936 Count = Builder.CreateSub(
1937 ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count);
1938 Value *NewCount = Count;
1939 if (IsCntPhiUsedOutsideLoop)
1940 Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1));
1941
1942 NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType());
1943
1944 Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
1945 if (cast<ConstantInt>(CntInst->getOperand(1))->isOne()) {
1946 // If the counter was being incremented in the loop, add NewCount to the
1947 // counter's initial value, but only if the initial value is not zero.
1948 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1949 if (!InitConst || !InitConst->isZero())
1950 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1951 } else {
1952 // If the count was being decremented in the loop, subtract NewCount from
1953 // the counter's initial value.
1954 NewCount = Builder.CreateSub(CntInitVal, NewCount);
1955 }
1956
1957 // Step 2: Insert new IV and loop condition:
1958 // loop:
1959 // ...
1960 // PhiCount = PHI [Count, Dec]
1961 // ...
1962 // Dec = PhiCount - 1
1963 // ...
1964 // Br: loop if (Dec != 0)
1965 BasicBlock *Body = *(CurLoop->block_begin());
1966 auto *LbBr = cast<BranchInst>(Body->getTerminator());
1967 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1968
1969 PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front());
1970
1971 Builder.SetInsertPoint(LbCond);
1972 Instruction *TcDec = cast<Instruction>(Builder.CreateSub(
1973 TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true));
1974
1975 TcPhi->addIncoming(Count, Preheader);
1976 TcPhi->addIncoming(TcDec, Body);
1977
1978 CmpInst::Predicate Pred =
1979 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
1980 LbCond->setPredicate(Pred);
1981 LbCond->setOperand(0, TcDec);
1982 LbCond->setOperand(1, ConstantInt::get(CountTy, 0));
1983
1984 // Step 3: All the references to the original counter outside
1985 // the loop are replaced with the NewCount
1986 if (IsCntPhiUsedOutsideLoop)
1987 CntPhi->replaceUsesOutsideBlock(NewCount, Body);
1988 else
1989 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1990
1991 // step 4: Forget the "non-computable" trip-count SCEV associated with the
1992 // loop. The loop would otherwise not be deleted even if it becomes empty.
1993 SE->forgetLoop(CurLoop);
1994 }
1995
transformLoopToPopcount(BasicBlock * PreCondBB,Instruction * CntInst,PHINode * CntPhi,Value * Var)1996 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
1997 Instruction *CntInst,
1998 PHINode *CntPhi, Value *Var) {
1999 BasicBlock *PreHead = CurLoop->getLoopPreheader();
2000 auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator());
2001 const DebugLoc &DL = CntInst->getDebugLoc();
2002
2003 // Assuming before transformation, the loop is following:
2004 // if (x) // the precondition
2005 // do { cnt++; x &= x - 1; } while(x);
2006
2007 // Step 1: Insert the ctpop instruction at the end of the precondition block
2008 IRBuilder<> Builder(PreCondBr);
2009 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
2010 {
2011 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
2012 NewCount = PopCntZext =
2013 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
2014
2015 if (NewCount != PopCnt)
2016 (cast<Instruction>(NewCount))->setDebugLoc(DL);
2017
2018 // TripCnt is exactly the number of iterations the loop has
2019 TripCnt = NewCount;
2020
2021 // If the population counter's initial value is not zero, insert Add Inst.
2022 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
2023 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
2024 if (!InitConst || !InitConst->isZero()) {
2025 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
2026 (cast<Instruction>(NewCount))->setDebugLoc(DL);
2027 }
2028 }
2029
2030 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
2031 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
2032 // function would be partial dead code, and downstream passes will drag
2033 // it back from the precondition block to the preheader.
2034 {
2035 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
2036
2037 Value *Opnd0 = PopCntZext;
2038 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
2039 if (PreCond->getOperand(0) != Var)
2040 std::swap(Opnd0, Opnd1);
2041
2042 ICmpInst *NewPreCond = cast<ICmpInst>(
2043 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
2044 PreCondBr->setCondition(NewPreCond);
2045
2046 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
2047 }
2048
2049 // Step 3: Note that the population count is exactly the trip count of the
2050 // loop in question, which enable us to convert the loop from noncountable
2051 // loop into a countable one. The benefit is twofold:
2052 //
2053 // - If the loop only counts population, the entire loop becomes dead after
2054 // the transformation. It is a lot easier to prove a countable loop dead
2055 // than to prove a noncountable one. (In some C dialects, an infinite loop
2056 // isn't dead even if it computes nothing useful. In general, DCE needs
2057 // to prove a noncountable loop finite before safely delete it.)
2058 //
2059 // - If the loop also performs something else, it remains alive.
2060 // Since it is transformed to countable form, it can be aggressively
2061 // optimized by some optimizations which are in general not applicable
2062 // to a noncountable loop.
2063 //
2064 // After this step, this loop (conceptually) would look like following:
2065 // newcnt = __builtin_ctpop(x);
2066 // t = newcnt;
2067 // if (x)
2068 // do { cnt++; x &= x-1; t--) } while (t > 0);
2069 BasicBlock *Body = *(CurLoop->block_begin());
2070 {
2071 auto *LbBr = cast<BranchInst>(Body->getTerminator());
2072 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
2073 Type *Ty = TripCnt->getType();
2074
2075 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
2076
2077 Builder.SetInsertPoint(LbCond);
2078 Instruction *TcDec = cast<Instruction>(
2079 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
2080 "tcdec", false, true));
2081
2082 TcPhi->addIncoming(TripCnt, PreHead);
2083 TcPhi->addIncoming(TcDec, Body);
2084
2085 CmpInst::Predicate Pred =
2086 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
2087 LbCond->setPredicate(Pred);
2088 LbCond->setOperand(0, TcDec);
2089 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
2090 }
2091
2092 // Step 4: All the references to the original population counter outside
2093 // the loop are replaced with the NewCount -- the value returned from
2094 // __builtin_ctpop().
2095 CntInst->replaceUsesOutsideBlock(NewCount, Body);
2096
2097 // step 5: Forget the "non-computable" trip-count SCEV associated with the
2098 // loop. The loop would otherwise not be deleted even if it becomes empty.
2099 SE->forgetLoop(CurLoop);
2100 }
2101
2102 /// Match loop-invariant value.
2103 template <typename SubPattern_t> struct match_LoopInvariant {
2104 SubPattern_t SubPattern;
2105 const Loop *L;
2106
match_LoopInvariantmatch_LoopInvariant2107 match_LoopInvariant(const SubPattern_t &SP, const Loop *L)
2108 : SubPattern(SP), L(L) {}
2109
matchmatch_LoopInvariant2110 template <typename ITy> bool match(ITy *V) {
2111 return L->isLoopInvariant(V) && SubPattern.match(V);
2112 }
2113 };
2114
2115 /// Matches if the value is loop-invariant.
2116 template <typename Ty>
m_LoopInvariant(const Ty & M,const Loop * L)2117 inline match_LoopInvariant<Ty> m_LoopInvariant(const Ty &M, const Loop *L) {
2118 return match_LoopInvariant<Ty>(M, L);
2119 }
2120
2121 /// Return true if the idiom is detected in the loop.
2122 ///
2123 /// The core idiom we are trying to detect is:
2124 /// \code
2125 /// entry:
2126 /// <...>
2127 /// %bitmask = shl i32 1, %bitpos
2128 /// br label %loop
2129 ///
2130 /// loop:
2131 /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2132 /// %x.curr.bitmasked = and i32 %x.curr, %bitmask
2133 /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2134 /// %x.next = shl i32 %x.curr, 1
2135 /// <...>
2136 /// br i1 %x.curr.isbitunset, label %loop, label %end
2137 ///
2138 /// end:
2139 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2140 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2141 /// <...>
2142 /// \endcode
detectShiftUntilBitTestIdiom(Loop * CurLoop,Value * & BaseX,Value * & BitMask,Value * & BitPos,Value * & CurrX,Instruction * & NextX)2143 static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX,
2144 Value *&BitMask, Value *&BitPos,
2145 Value *&CurrX, Instruction *&NextX) {
2146 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2147 " Performing shift-until-bittest idiom detection.\n");
2148
2149 // Give up if the loop has multiple blocks or multiple backedges.
2150 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2151 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2152 return false;
2153 }
2154
2155 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2156 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2157 assert(LoopPreheaderBB && "There is always a loop preheader.");
2158
2159 using namespace PatternMatch;
2160
2161 // Step 1: Check if the loop backedge is in desirable form.
2162
2163 ICmpInst::Predicate Pred;
2164 Value *CmpLHS, *CmpRHS;
2165 BasicBlock *TrueBB, *FalseBB;
2166 if (!match(LoopHeaderBB->getTerminator(),
2167 m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)),
2168 m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) {
2169 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2170 return false;
2171 }
2172
2173 // Step 2: Check if the backedge's condition is in desirable form.
2174
2175 auto MatchVariableBitMask = [&]() {
2176 return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2177 match(CmpLHS,
2178 m_c_And(m_Value(CurrX),
2179 m_CombineAnd(
2180 m_Value(BitMask),
2181 m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)),
2182 CurLoop))));
2183 };
2184 auto MatchConstantBitMask = [&]() {
2185 return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) &&
2186 match(CmpLHS, m_And(m_Value(CurrX),
2187 m_CombineAnd(m_Value(BitMask), m_Power2()))) &&
2188 (BitPos = ConstantExpr::getExactLogBase2(cast<Constant>(BitMask)));
2189 };
2190 auto MatchDecomposableConstantBitMask = [&]() {
2191 APInt Mask;
2192 return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) &&
2193 ICmpInst::isEquality(Pred) && Mask.isPowerOf2() &&
2194 (BitMask = ConstantInt::get(CurrX->getType(), Mask)) &&
2195 (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2()));
2196 };
2197
2198 if (!MatchVariableBitMask() && !MatchConstantBitMask() &&
2199 !MatchDecomposableConstantBitMask()) {
2200 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n");
2201 return false;
2202 }
2203
2204 // Step 3: Check if the recurrence is in desirable form.
2205 auto *CurrXPN = dyn_cast<PHINode>(CurrX);
2206 if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) {
2207 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2208 return false;
2209 }
2210
2211 BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB);
2212 NextX =
2213 dyn_cast<Instruction>(CurrXPN->getIncomingValueForBlock(LoopHeaderBB));
2214
2215 assert(CurLoop->isLoopInvariant(BaseX) &&
2216 "Expected BaseX to be avaliable in the preheader!");
2217
2218 if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) {
2219 // FIXME: support right-shift?
2220 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2221 return false;
2222 }
2223
2224 // Step 4: Check if the backedge's destinations are in desirable form.
2225
2226 assert(ICmpInst::isEquality(Pred) &&
2227 "Should only get equality predicates here.");
2228
2229 // cmp-br is commutative, so canonicalize to a single variant.
2230 if (Pred != ICmpInst::Predicate::ICMP_EQ) {
2231 Pred = ICmpInst::getInversePredicate(Pred);
2232 std::swap(TrueBB, FalseBB);
2233 }
2234
2235 // We expect to exit loop when comparison yields false,
2236 // so when it yields true we should branch back to loop header.
2237 if (TrueBB != LoopHeaderBB) {
2238 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2239 return false;
2240 }
2241
2242 // Okay, idiom checks out.
2243 return true;
2244 }
2245
2246 /// Look for the following loop:
2247 /// \code
2248 /// entry:
2249 /// <...>
2250 /// %bitmask = shl i32 1, %bitpos
2251 /// br label %loop
2252 ///
2253 /// loop:
2254 /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ]
2255 /// %x.curr.bitmasked = and i32 %x.curr, %bitmask
2256 /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0
2257 /// %x.next = shl i32 %x.curr, 1
2258 /// <...>
2259 /// br i1 %x.curr.isbitunset, label %loop, label %end
2260 ///
2261 /// end:
2262 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2263 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2264 /// <...>
2265 /// \endcode
2266 ///
2267 /// And transform it into:
2268 /// \code
2269 /// entry:
2270 /// %bitmask = shl i32 1, %bitpos
2271 /// %lowbitmask = add i32 %bitmask, -1
2272 /// %mask = or i32 %lowbitmask, %bitmask
2273 /// %x.masked = and i32 %x, %mask
2274 /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked,
2275 /// i1 true)
2276 /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros
2277 /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1
2278 /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos
2279 /// %tripcount = add i32 %backedgetakencount, 1
2280 /// %x.curr = shl i32 %x, %backedgetakencount
2281 /// %x.next = shl i32 %x, %tripcount
2282 /// br label %loop
2283 ///
2284 /// loop:
2285 /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ]
2286 /// %loop.iv.next = add nuw i32 %loop.iv, 1
2287 /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount
2288 /// <...>
2289 /// br i1 %loop.ivcheck, label %end, label %loop
2290 ///
2291 /// end:
2292 /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...>
2293 /// %x.next.res = phi i32 [ %x.next, %loop ] <...>
2294 /// <...>
2295 /// \endcode
recognizeShiftUntilBitTest()2296 bool LoopIdiomRecognize::recognizeShiftUntilBitTest() {
2297 bool MadeChange = false;
2298
2299 Value *X, *BitMask, *BitPos, *XCurr;
2300 Instruction *XNext;
2301 if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr,
2302 XNext)) {
2303 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2304 " shift-until-bittest idiom detection failed.\n");
2305 return MadeChange;
2306 }
2307 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n");
2308
2309 // Ok, it is the idiom we were looking for, we *could* transform this loop,
2310 // but is it profitable to transform?
2311
2312 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2313 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2314 assert(LoopPreheaderBB && "There is always a loop preheader.");
2315
2316 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2317 assert(SuccessorBB && "There is only a single successor.");
2318
2319 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2320 Builder.SetCurrentDebugLocation(cast<Instruction>(XCurr)->getDebugLoc());
2321
2322 Intrinsic::ID IntrID = Intrinsic::ctlz;
2323 Type *Ty = X->getType();
2324 unsigned Bitwidth = Ty->getScalarSizeInBits();
2325
2326 TargetTransformInfo::TargetCostKind CostKind =
2327 TargetTransformInfo::TCK_SizeAndLatency;
2328
2329 // The rewrite is considered to be unprofitable iff and only iff the
2330 // intrinsic/shift we'll use are not cheap. Note that we are okay with *just*
2331 // making the loop countable, even if nothing else changes.
2332 IntrinsicCostAttributes Attrs(
2333 IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getTrue()});
2334 InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2335 if (Cost > TargetTransformInfo::TCC_Basic) {
2336 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2337 " Intrinsic is too costly, not beneficial\n");
2338 return MadeChange;
2339 }
2340 if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) >
2341 TargetTransformInfo::TCC_Basic) {
2342 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n");
2343 return MadeChange;
2344 }
2345
2346 // Ok, transform appears worthwhile.
2347 MadeChange = true;
2348
2349 // Step 1: Compute the loop trip count.
2350
2351 Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty),
2352 BitPos->getName() + ".lowbitmask");
2353 Value *Mask =
2354 Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask");
2355 Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked");
2356 CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic(
2357 IntrID, Ty, {XMasked, /*is_zero_undef=*/Builder.getTrue()},
2358 /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros");
2359 Value *XMaskedNumActiveBits = Builder.CreateSub(
2360 ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros,
2361 XMasked->getName() + ".numactivebits", /*HasNUW=*/true,
2362 /*HasNSW=*/Bitwidth != 2);
2363 Value *XMaskedLeadingOnePos =
2364 Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty),
2365 XMasked->getName() + ".leadingonepos", /*HasNUW=*/false,
2366 /*HasNSW=*/Bitwidth > 2);
2367
2368 Value *LoopBackedgeTakenCount = Builder.CreateSub(
2369 BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount",
2370 /*HasNUW=*/true, /*HasNSW=*/true);
2371 // We know loop's backedge-taken count, but what's loop's trip count?
2372 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2373 Value *LoopTripCount =
2374 Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
2375 CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
2376 /*HasNSW=*/Bitwidth != 2);
2377
2378 // Step 2: Compute the recurrence's final value without a loop.
2379
2380 // NewX is always safe to compute, because `LoopBackedgeTakenCount`
2381 // will always be smaller than `bitwidth(X)`, i.e. we never get poison.
2382 Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount);
2383 NewX->takeName(XCurr);
2384 if (auto *I = dyn_cast<Instruction>(NewX))
2385 I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2386
2387 Value *NewXNext;
2388 // Rewriting XNext is more complicated, however, because `X << LoopTripCount`
2389 // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen
2390 // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know
2391 // that isn't the case, we'll need to emit an alternative, safe IR.
2392 if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() ||
2393 PatternMatch::match(
2394 BitPos, PatternMatch::m_SpecificInt_ICMP(
2395 ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(),
2396 Ty->getScalarSizeInBits() - 1))))
2397 NewXNext = Builder.CreateShl(X, LoopTripCount);
2398 else {
2399 // Otherwise, just additionally shift by one. It's the smallest solution,
2400 // alternatively, we could check that NewX is INT_MIN (or BitPos is )
2401 // and select 0 instead.
2402 NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1));
2403 }
2404
2405 NewXNext->takeName(XNext);
2406 if (auto *I = dyn_cast<Instruction>(NewXNext))
2407 I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true);
2408
2409 // Step 3: Adjust the successor basic block to recieve the computed
2410 // recurrence's final value instead of the recurrence itself.
2411
2412 XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB);
2413 XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB);
2414
2415 // Step 4: Rewrite the loop into a countable form, with canonical IV.
2416
2417 // The new canonical induction variable.
2418 Builder.SetInsertPoint(&LoopHeaderBB->front());
2419 auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
2420
2421 // The induction itself.
2422 // Note that while NUW is always safe, while NSW is only for bitwidths != 2.
2423 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
2424 auto *IVNext =
2425 Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next",
2426 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
2427
2428 // The loop trip count check.
2429 auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount,
2430 CurLoop->getName() + ".ivcheck");
2431 Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB);
2432 LoopHeaderBB->getTerminator()->eraseFromParent();
2433
2434 // Populate the IV PHI.
2435 IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
2436 IV->addIncoming(IVNext, LoopHeaderBB);
2437
2438 // Step 5: Forget the "non-computable" trip-count SCEV associated with the
2439 // loop. The loop would otherwise not be deleted even if it becomes empty.
2440
2441 SE->forgetLoop(CurLoop);
2442
2443 // Other passes will take care of actually deleting the loop if possible.
2444
2445 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n");
2446
2447 ++NumShiftUntilBitTest;
2448 return MadeChange;
2449 }
2450
2451 /// Return true if the idiom is detected in the loop.
2452 ///
2453 /// The core idiom we are trying to detect is:
2454 /// \code
2455 /// entry:
2456 /// <...>
2457 /// %start = <...>
2458 /// %extraoffset = <...>
2459 /// <...>
2460 /// br label %for.cond
2461 ///
2462 /// loop:
2463 /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2464 /// %nbits = add nsw i8 %iv, %extraoffset
2465 /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2466 /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2467 /// %iv.next = add i8 %iv, 1
2468 /// <...>
2469 /// br i1 %val.shifted.iszero, label %end, label %loop
2470 ///
2471 /// end:
2472 /// %iv.res = phi i8 [ %iv, %loop ] <...>
2473 /// %nbits.res = phi i8 [ %nbits, %loop ] <...>
2474 /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2475 /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2476 /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2477 /// <...>
2478 /// \endcode
detectShiftUntilZeroIdiom(Loop * CurLoop,ScalarEvolution * SE,Instruction * & ValShiftedIsZero,Intrinsic::ID & IntrinID,Instruction * & IV,Value * & Start,Value * & Val,const SCEV * & ExtraOffsetExpr,bool & InvertedCond)2479 static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE,
2480 Instruction *&ValShiftedIsZero,
2481 Intrinsic::ID &IntrinID, Instruction *&IV,
2482 Value *&Start, Value *&Val,
2483 const SCEV *&ExtraOffsetExpr,
2484 bool &InvertedCond) {
2485 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2486 " Performing shift-until-zero idiom detection.\n");
2487
2488 // Give up if the loop has multiple blocks or multiple backedges.
2489 if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) {
2490 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n");
2491 return false;
2492 }
2493
2494 Instruction *ValShifted, *NBits, *IVNext;
2495 Value *ExtraOffset;
2496
2497 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2498 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2499 assert(LoopPreheaderBB && "There is always a loop preheader.");
2500
2501 using namespace PatternMatch;
2502
2503 // Step 1: Check if the loop backedge, condition is in desirable form.
2504
2505 ICmpInst::Predicate Pred;
2506 BasicBlock *TrueBB, *FalseBB;
2507 if (!match(LoopHeaderBB->getTerminator(),
2508 m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB),
2509 m_BasicBlock(FalseBB))) ||
2510 !match(ValShiftedIsZero,
2511 m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) ||
2512 !ICmpInst::isEquality(Pred)) {
2513 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n");
2514 return false;
2515 }
2516
2517 // Step 2: Check if the comparison's operand is in desirable form.
2518 // FIXME: Val could be a one-input PHI node, which we should look past.
2519 if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop),
2520 m_Instruction(NBits)))) {
2521 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n");
2522 return false;
2523 }
2524 IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz
2525 : Intrinsic::ctlz;
2526
2527 // Step 3: Check if the shift amount is in desirable form.
2528
2529 if (match(NBits, m_c_Add(m_Instruction(IV),
2530 m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2531 (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap()))
2532 ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset));
2533 else if (match(NBits,
2534 m_Sub(m_Instruction(IV),
2535 m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) &&
2536 NBits->hasNoSignedWrap())
2537 ExtraOffsetExpr = SE->getSCEV(ExtraOffset);
2538 else {
2539 IV = NBits;
2540 ExtraOffsetExpr = SE->getZero(NBits->getType());
2541 }
2542
2543 // Step 4: Check if the recurrence is in desirable form.
2544 auto *IVPN = dyn_cast<PHINode>(IV);
2545 if (!IVPN || IVPN->getParent() != LoopHeaderBB) {
2546 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n");
2547 return false;
2548 }
2549
2550 Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB);
2551 IVNext = dyn_cast<Instruction>(IVPN->getIncomingValueForBlock(LoopHeaderBB));
2552
2553 if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) {
2554 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n");
2555 return false;
2556 }
2557
2558 // Step 4: Check if the backedge's destinations are in desirable form.
2559
2560 assert(ICmpInst::isEquality(Pred) &&
2561 "Should only get equality predicates here.");
2562
2563 // cmp-br is commutative, so canonicalize to a single variant.
2564 InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ;
2565 if (InvertedCond) {
2566 Pred = ICmpInst::getInversePredicate(Pred);
2567 std::swap(TrueBB, FalseBB);
2568 }
2569
2570 // We expect to exit loop when comparison yields true,
2571 // so when it yields false we should branch back to loop header.
2572 if (FalseBB != LoopHeaderBB) {
2573 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n");
2574 return false;
2575 }
2576
2577 // The new, countable, loop will certainly only run a known number of
2578 // iterations, It won't be infinite. But the old loop might be infinite
2579 // under certain conditions. For logical shifts, the value will become zero
2580 // after at most bitwidth(%Val) loop iterations. However, for arithmetic
2581 // right-shift, iff the sign bit was set, the value will never become zero,
2582 // and the loop may never finish.
2583 if (ValShifted->getOpcode() == Instruction::AShr &&
2584 !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) {
2585 LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n");
2586 return false;
2587 }
2588
2589 // Okay, idiom checks out.
2590 return true;
2591 }
2592
2593 /// Look for the following loop:
2594 /// \code
2595 /// entry:
2596 /// <...>
2597 /// %start = <...>
2598 /// %extraoffset = <...>
2599 /// <...>
2600 /// br label %for.cond
2601 ///
2602 /// loop:
2603 /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ]
2604 /// %nbits = add nsw i8 %iv, %extraoffset
2605 /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits
2606 /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0
2607 /// %iv.next = add i8 %iv, 1
2608 /// <...>
2609 /// br i1 %val.shifted.iszero, label %end, label %loop
2610 ///
2611 /// end:
2612 /// %iv.res = phi i8 [ %iv, %loop ] <...>
2613 /// %nbits.res = phi i8 [ %nbits, %loop ] <...>
2614 /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...>
2615 /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...>
2616 /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...>
2617 /// <...>
2618 /// \endcode
2619 ///
2620 /// And transform it into:
2621 /// \code
2622 /// entry:
2623 /// <...>
2624 /// %start = <...>
2625 /// %extraoffset = <...>
2626 /// <...>
2627 /// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0)
2628 /// %val.numactivebits = sub i8 8, %val.numleadingzeros
2629 /// %extraoffset.neg = sub i8 0, %extraoffset
2630 /// %tmp = add i8 %val.numactivebits, %extraoffset.neg
2631 /// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start)
2632 /// %loop.tripcount = sub i8 %iv.final, %start
2633 /// br label %loop
2634 ///
2635 /// loop:
2636 /// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ]
2637 /// %loop.iv.next = add i8 %loop.iv, 1
2638 /// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount
2639 /// %iv = add i8 %loop.iv, %start
2640 /// <...>
2641 /// br i1 %loop.ivcheck, label %end, label %loop
2642 ///
2643 /// end:
2644 /// %iv.res = phi i8 [ %iv.final, %loop ] <...>
2645 /// <...>
2646 /// \endcode
recognizeShiftUntilZero()2647 bool LoopIdiomRecognize::recognizeShiftUntilZero() {
2648 bool MadeChange = false;
2649
2650 Instruction *ValShiftedIsZero;
2651 Intrinsic::ID IntrID;
2652 Instruction *IV;
2653 Value *Start, *Val;
2654 const SCEV *ExtraOffsetExpr;
2655 bool InvertedCond;
2656 if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV,
2657 Start, Val, ExtraOffsetExpr, InvertedCond)) {
2658 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2659 " shift-until-zero idiom detection failed.\n");
2660 return MadeChange;
2661 }
2662 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n");
2663
2664 // Ok, it is the idiom we were looking for, we *could* transform this loop,
2665 // but is it profitable to transform?
2666
2667 BasicBlock *LoopHeaderBB = CurLoop->getHeader();
2668 BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader();
2669 assert(LoopPreheaderBB && "There is always a loop preheader.");
2670
2671 BasicBlock *SuccessorBB = CurLoop->getExitBlock();
2672 assert(SuccessorBB && "There is only a single successor.");
2673
2674 IRBuilder<> Builder(LoopPreheaderBB->getTerminator());
2675 Builder.SetCurrentDebugLocation(IV->getDebugLoc());
2676
2677 Type *Ty = Val->getType();
2678 unsigned Bitwidth = Ty->getScalarSizeInBits();
2679
2680 TargetTransformInfo::TargetCostKind CostKind =
2681 TargetTransformInfo::TCK_SizeAndLatency;
2682
2683 // The rewrite is considered to be unprofitable iff and only iff the
2684 // intrinsic we'll use are not cheap. Note that we are okay with *just*
2685 // making the loop countable, even if nothing else changes.
2686 IntrinsicCostAttributes Attrs(
2687 IntrID, Ty, {UndefValue::get(Ty), /*is_zero_undef=*/Builder.getFalse()});
2688 InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind);
2689 if (Cost > TargetTransformInfo::TCC_Basic) {
2690 LLVM_DEBUG(dbgs() << DEBUG_TYPE
2691 " Intrinsic is too costly, not beneficial\n");
2692 return MadeChange;
2693 }
2694
2695 // Ok, transform appears worthwhile.
2696 MadeChange = true;
2697
2698 bool OffsetIsZero = false;
2699 if (auto *ExtraOffsetExprC = dyn_cast<SCEVConstant>(ExtraOffsetExpr))
2700 OffsetIsZero = ExtraOffsetExprC->isZero();
2701
2702 // Step 1: Compute the loop's final IV value / trip count.
2703
2704 CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic(
2705 IntrID, Ty, {Val, /*is_zero_undef=*/Builder.getFalse()},
2706 /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros");
2707 Value *ValNumActiveBits = Builder.CreateSub(
2708 ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros,
2709 Val->getName() + ".numactivebits", /*HasNUW=*/true,
2710 /*HasNSW=*/Bitwidth != 2);
2711
2712 SCEVExpander Expander(*SE, *DL, "loop-idiom");
2713 Expander.setInsertPoint(&*Builder.GetInsertPoint());
2714 Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr);
2715
2716 Value *ValNumActiveBitsOffset = Builder.CreateAdd(
2717 ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset",
2718 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true);
2719 Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty},
2720 {ValNumActiveBitsOffset, Start},
2721 /*FMFSource=*/nullptr, "iv.final");
2722
2723 auto *LoopBackedgeTakenCount = cast<Instruction>(Builder.CreateSub(
2724 IVFinal, Start, CurLoop->getName() + ".backedgetakencount",
2725 /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true));
2726 // FIXME: or when the offset was `add nuw`
2727
2728 // We know loop's backedge-taken count, but what's loop's trip count?
2729 Value *LoopTripCount =
2730 Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1),
2731 CurLoop->getName() + ".tripcount", /*HasNUW=*/true,
2732 /*HasNSW=*/Bitwidth != 2);
2733
2734 // Step 2: Adjust the successor basic block to recieve the original
2735 // induction variable's final value instead of the orig. IV itself.
2736
2737 IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB);
2738
2739 // Step 3: Rewrite the loop into a countable form, with canonical IV.
2740
2741 // The new canonical induction variable.
2742 Builder.SetInsertPoint(&LoopHeaderBB->front());
2743 auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv");
2744
2745 // The induction itself.
2746 Builder.SetInsertPoint(LoopHeaderBB->getFirstNonPHI());
2747 auto *CIVNext =
2748 Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next",
2749 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2);
2750
2751 // The loop trip count check.
2752 auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount,
2753 CurLoop->getName() + ".ivcheck");
2754 auto *NewIVCheck = CIVCheck;
2755 if (InvertedCond) {
2756 NewIVCheck = Builder.CreateNot(CIVCheck);
2757 NewIVCheck->takeName(ValShiftedIsZero);
2758 }
2759
2760 // The original IV, but rebased to be an offset to the CIV.
2761 auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false,
2762 /*HasNSW=*/true); // FIXME: what about NUW?
2763 IVDePHId->takeName(IV);
2764
2765 // The loop terminator.
2766 Builder.SetInsertPoint(LoopHeaderBB->getTerminator());
2767 Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB);
2768 LoopHeaderBB->getTerminator()->eraseFromParent();
2769
2770 // Populate the IV PHI.
2771 CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB);
2772 CIV->addIncoming(CIVNext, LoopHeaderBB);
2773
2774 // Step 4: Forget the "non-computable" trip-count SCEV associated with the
2775 // loop. The loop would otherwise not be deleted even if it becomes empty.
2776
2777 SE->forgetLoop(CurLoop);
2778
2779 // Step 5: Try to cleanup the loop's body somewhat.
2780 IV->replaceAllUsesWith(IVDePHId);
2781 IV->eraseFromParent();
2782
2783 ValShiftedIsZero->replaceAllUsesWith(NewIVCheck);
2784 ValShiftedIsZero->eraseFromParent();
2785
2786 // Other passes will take care of actually deleting the loop if possible.
2787
2788 LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n");
2789
2790 ++NumShiftUntilZero;
2791 return MadeChange;
2792 }
2793