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