1 | //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis ------------===// |
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 file contains the implementation of the scalar evolution expander, |
10 | // which is used to generate the code corresponding to a given scalar evolution |
11 | // expression. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
16 | #include "llvm/ADT/STLExtras.h" |
17 | #include "llvm/ADT/ScopeExit.h" |
18 | #include "llvm/ADT/SmallSet.h" |
19 | #include "llvm/Analysis/InstructionSimplify.h" |
20 | #include "llvm/Analysis/LoopInfo.h" |
21 | #include "llvm/Analysis/TargetTransformInfo.h" |
22 | #include "llvm/Analysis/ValueTracking.h" |
23 | #include "llvm/IR/DataLayout.h" |
24 | #include "llvm/IR/Dominators.h" |
25 | #include "llvm/IR/IntrinsicInst.h" |
26 | #include "llvm/IR/PatternMatch.h" |
27 | #include "llvm/Support/CommandLine.h" |
28 | #include "llvm/Support/raw_ostream.h" |
29 | #include "llvm/Transforms/Utils/LoopUtils.h" |
30 | |
31 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
32 | #define SCEV_DEBUG_WITH_TYPE(TYPE, X) DEBUG_WITH_TYPE(TYPE, X) |
33 | #else |
34 | #define SCEV_DEBUG_WITH_TYPE(TYPE, X) |
35 | #endif |
36 | |
37 | using namespace llvm; |
38 | |
39 | cl::opt<unsigned> llvm::SCEVCheapExpansionBudget( |
40 | "scev-cheap-expansion-budget" , cl::Hidden, cl::init(Val: 4), |
41 | cl::desc("When performing SCEV expansion only if it is cheap to do, this " |
42 | "controls the budget that is considered cheap (default = 4)" )); |
43 | |
44 | using namespace PatternMatch; |
45 | |
46 | PoisonFlags::PoisonFlags(const Instruction *I) { |
47 | NUW = false; |
48 | NSW = false; |
49 | Exact = false; |
50 | Disjoint = false; |
51 | NNeg = false; |
52 | SameSign = false; |
53 | GEPNW = GEPNoWrapFlags::none(); |
54 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: I)) { |
55 | NUW = OBO->hasNoUnsignedWrap(); |
56 | NSW = OBO->hasNoSignedWrap(); |
57 | } |
58 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(Val: I)) |
59 | Exact = PEO->isExact(); |
60 | if (auto *PDI = dyn_cast<PossiblyDisjointInst>(Val: I)) |
61 | Disjoint = PDI->isDisjoint(); |
62 | if (auto *PNI = dyn_cast<PossiblyNonNegInst>(Val: I)) |
63 | NNeg = PNI->hasNonNeg(); |
64 | if (auto *TI = dyn_cast<TruncInst>(Val: I)) { |
65 | NUW = TI->hasNoUnsignedWrap(); |
66 | NSW = TI->hasNoSignedWrap(); |
67 | } |
68 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I)) |
69 | GEPNW = GEP->getNoWrapFlags(); |
70 | if (auto *ICmp = dyn_cast<ICmpInst>(Val: I)) |
71 | SameSign = ICmp->hasSameSign(); |
72 | } |
73 | |
74 | void PoisonFlags::apply(Instruction *I) { |
75 | if (isa<OverflowingBinaryOperator>(Val: I)) { |
76 | I->setHasNoUnsignedWrap(NUW); |
77 | I->setHasNoSignedWrap(NSW); |
78 | } |
79 | if (isa<PossiblyExactOperator>(Val: I)) |
80 | I->setIsExact(Exact); |
81 | if (auto *PDI = dyn_cast<PossiblyDisjointInst>(Val: I)) |
82 | PDI->setIsDisjoint(Disjoint); |
83 | if (auto *PNI = dyn_cast<PossiblyNonNegInst>(Val: I)) |
84 | PNI->setNonNeg(NNeg); |
85 | if (isa<TruncInst>(Val: I)) { |
86 | I->setHasNoUnsignedWrap(NUW); |
87 | I->setHasNoSignedWrap(NSW); |
88 | } |
89 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: I)) |
90 | GEP->setNoWrapFlags(GEPNW); |
91 | if (auto *ICmp = dyn_cast<ICmpInst>(Val: I)) |
92 | ICmp->setSameSign(SameSign); |
93 | } |
94 | |
95 | /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, |
96 | /// reusing an existing cast if a suitable one (= dominating IP) exists, or |
97 | /// creating a new one. |
98 | Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, |
99 | Instruction::CastOps Op, |
100 | BasicBlock::iterator IP) { |
101 | // This function must be called with the builder having a valid insertion |
102 | // point. It doesn't need to be the actual IP where the uses of the returned |
103 | // cast will be added, but it must dominate such IP. |
104 | // We use this precondition to produce a cast that will dominate all its |
105 | // uses. In particular, this is crucial for the case where the builder's |
106 | // insertion point *is* the point where we were asked to put the cast. |
107 | // Since we don't know the builder's insertion point is actually |
108 | // where the uses will be added (only that it dominates it), we are |
109 | // not allowed to move it. |
110 | BasicBlock::iterator BIP = Builder.GetInsertPoint(); |
111 | |
112 | Value *Ret = nullptr; |
113 | |
114 | if (!isa<Constant>(Val: V)) { |
115 | // Check to see if there is already a cast! |
116 | for (User *U : V->users()) { |
117 | if (U->getType() != Ty) |
118 | continue; |
119 | CastInst *CI = dyn_cast<CastInst>(Val: U); |
120 | if (!CI || CI->getOpcode() != Op) |
121 | continue; |
122 | |
123 | // Found a suitable cast that is at IP or comes before IP. Use it. Note |
124 | // that the cast must also properly dominate the Builder's insertion |
125 | // point. |
126 | if (IP->getParent() == CI->getParent() && &*BIP != CI && |
127 | (&*IP == CI || CI->comesBefore(Other: &*IP))) { |
128 | Ret = CI; |
129 | break; |
130 | } |
131 | } |
132 | } |
133 | |
134 | // Create a new cast. |
135 | if (!Ret) { |
136 | SCEVInsertPointGuard Guard(Builder, this); |
137 | Builder.SetInsertPoint(&*IP); |
138 | Ret = Builder.CreateCast(Op, V, DestTy: Ty, Name: V->getName()); |
139 | } |
140 | |
141 | // We assert at the end of the function since IP might point to an |
142 | // instruction with different dominance properties than a cast |
143 | // (an invoke for example) and not dominate BIP (but the cast does). |
144 | assert(!isa<Instruction>(Ret) || |
145 | SE.DT.dominates(cast<Instruction>(Ret), &*BIP)); |
146 | |
147 | return Ret; |
148 | } |
149 | |
150 | BasicBlock::iterator |
151 | SCEVExpander::findInsertPointAfter(Instruction *I, |
152 | Instruction *MustDominate) const { |
153 | BasicBlock::iterator IP = ++I->getIterator(); |
154 | if (auto *II = dyn_cast<InvokeInst>(Val: I)) |
155 | IP = II->getNormalDest()->begin(); |
156 | |
157 | while (isa<PHINode>(Val: IP)) |
158 | ++IP; |
159 | |
160 | if (isa<FuncletPadInst>(Val: IP) || isa<LandingPadInst>(Val: IP)) { |
161 | ++IP; |
162 | } else if (isa<CatchSwitchInst>(Val: IP)) { |
163 | IP = MustDominate->getParent()->getFirstInsertionPt(); |
164 | } else { |
165 | assert(!IP->isEHPad() && "unexpected eh pad!" ); |
166 | } |
167 | |
168 | // Adjust insert point to be after instructions inserted by the expander, so |
169 | // we can re-use already inserted instructions. Avoid skipping past the |
170 | // original \p MustDominate, in case it is an inserted instruction. |
171 | while (isInsertedInstruction(I: &*IP) && &*IP != MustDominate) |
172 | ++IP; |
173 | |
174 | return IP; |
175 | } |
176 | |
177 | BasicBlock::iterator |
178 | SCEVExpander::GetOptimalInsertionPointForCastOf(Value *V) const { |
179 | // Cast the argument at the beginning of the entry block, after |
180 | // any bitcasts of other arguments. |
181 | if (Argument *A = dyn_cast<Argument>(Val: V)) { |
182 | BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); |
183 | while ((isa<BitCastInst>(Val: IP) && |
184 | isa<Argument>(Val: cast<BitCastInst>(Val&: IP)->getOperand(i_nocapture: 0)) && |
185 | cast<BitCastInst>(Val&: IP)->getOperand(i_nocapture: 0) != A)) |
186 | ++IP; |
187 | return IP; |
188 | } |
189 | |
190 | // Cast the instruction immediately after the instruction. |
191 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
192 | return findInsertPointAfter(I, MustDominate: &*Builder.GetInsertPoint()); |
193 | |
194 | // Otherwise, this must be some kind of a constant, |
195 | // so let's plop this cast into the function's entry block. |
196 | assert(isa<Constant>(V) && |
197 | "Expected the cast argument to be a global/constant" ); |
198 | return Builder.GetInsertBlock() |
199 | ->getParent() |
200 | ->getEntryBlock() |
201 | .getFirstInsertionPt(); |
202 | } |
203 | |
204 | /// InsertNoopCastOfTo - Insert a cast of V to the specified type, |
205 | /// which must be possible with a noop cast, doing what we can to share |
206 | /// the casts. |
207 | Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { |
208 | Instruction::CastOps Op = CastInst::getCastOpcode(Val: V, SrcIsSigned: false, Ty, DstIsSigned: false); |
209 | assert((Op == Instruction::BitCast || |
210 | Op == Instruction::PtrToInt || |
211 | Op == Instruction::IntToPtr) && |
212 | "InsertNoopCastOfTo cannot perform non-noop casts!" ); |
213 | assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && |
214 | "InsertNoopCastOfTo cannot change sizes!" ); |
215 | |
216 | // inttoptr only works for integral pointers. For non-integral pointers, we |
217 | // can create a GEP on null with the integral value as index. Note that |
218 | // it is safe to use GEP of null instead of inttoptr here, because only |
219 | // expressions already based on a GEP of null should be converted to pointers |
220 | // during expansion. |
221 | if (Op == Instruction::IntToPtr) { |
222 | auto *PtrTy = cast<PointerType>(Val: Ty); |
223 | if (DL.isNonIntegralPointerType(PT: PtrTy)) |
224 | return Builder.CreatePtrAdd(Ptr: Constant::getNullValue(Ty: PtrTy), Offset: V, Name: "scevgep" ); |
225 | } |
226 | // Short-circuit unnecessary bitcasts. |
227 | if (Op == Instruction::BitCast) { |
228 | if (V->getType() == Ty) |
229 | return V; |
230 | if (CastInst *CI = dyn_cast<CastInst>(Val: V)) { |
231 | if (CI->getOperand(i_nocapture: 0)->getType() == Ty) |
232 | return CI->getOperand(i_nocapture: 0); |
233 | } |
234 | } |
235 | // Short-circuit unnecessary inttoptr<->ptrtoint casts. |
236 | if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && |
237 | SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(Ty: V->getType())) { |
238 | if (CastInst *CI = dyn_cast<CastInst>(Val: V)) |
239 | if ((CI->getOpcode() == Instruction::PtrToInt || |
240 | CI->getOpcode() == Instruction::IntToPtr) && |
241 | SE.getTypeSizeInBits(Ty: CI->getType()) == |
242 | SE.getTypeSizeInBits(Ty: CI->getOperand(i_nocapture: 0)->getType())) |
243 | return CI->getOperand(i_nocapture: 0); |
244 | if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: V)) |
245 | if ((CE->getOpcode() == Instruction::PtrToInt || |
246 | CE->getOpcode() == Instruction::IntToPtr) && |
247 | SE.getTypeSizeInBits(Ty: CE->getType()) == |
248 | SE.getTypeSizeInBits(Ty: CE->getOperand(i_nocapture: 0)->getType())) |
249 | return CE->getOperand(i_nocapture: 0); |
250 | } |
251 | |
252 | // Fold a cast of a constant. |
253 | if (Constant *C = dyn_cast<Constant>(Val: V)) |
254 | return ConstantExpr::getCast(ops: Op, C, Ty); |
255 | |
256 | // Try to reuse existing cast, or insert one. |
257 | return ReuseOrCreateCast(V, Ty, Op, IP: GetOptimalInsertionPointForCastOf(V)); |
258 | } |
259 | |
260 | /// InsertBinop - Insert the specified binary operator, doing a small amount |
261 | /// of work to avoid inserting an obviously redundant operation, and hoisting |
262 | /// to an outer loop when the opportunity is there and it is safe. |
263 | Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, |
264 | Value *LHS, Value *RHS, |
265 | SCEV::NoWrapFlags Flags, bool IsSafeToHoist) { |
266 | // Fold a binop with constant operands. |
267 | if (Constant *CLHS = dyn_cast<Constant>(Val: LHS)) |
268 | if (Constant *CRHS = dyn_cast<Constant>(Val: RHS)) |
269 | if (Constant *Res = ConstantFoldBinaryOpOperands(Opcode, LHS: CLHS, RHS: CRHS, DL)) |
270 | return Res; |
271 | |
272 | // Do a quick scan to see if we have this binop nearby. If so, reuse it. |
273 | unsigned ScanLimit = 6; |
274 | BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); |
275 | // Scanning starts from the last instruction before the insertion point. |
276 | BasicBlock::iterator IP = Builder.GetInsertPoint(); |
277 | if (IP != BlockBegin) { |
278 | --IP; |
279 | for (; ScanLimit; --IP, --ScanLimit) { |
280 | auto canGenerateIncompatiblePoison = [&Flags](Instruction *I) { |
281 | // Ensure that no-wrap flags match. |
282 | if (isa<OverflowingBinaryOperator>(Val: I)) { |
283 | if (I->hasNoSignedWrap() != (Flags & SCEV::FlagNSW)) |
284 | return true; |
285 | if (I->hasNoUnsignedWrap() != (Flags & SCEV::FlagNUW)) |
286 | return true; |
287 | } |
288 | // Conservatively, do not use any instruction which has any of exact |
289 | // flags installed. |
290 | if (isa<PossiblyExactOperator>(Val: I) && I->isExact()) |
291 | return true; |
292 | return false; |
293 | }; |
294 | if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(i: 0) == LHS && |
295 | IP->getOperand(i: 1) == RHS && !canGenerateIncompatiblePoison(&*IP)) |
296 | return &*IP; |
297 | if (IP == BlockBegin) break; |
298 | } |
299 | } |
300 | |
301 | // Save the original insertion point so we can restore it when we're done. |
302 | DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); |
303 | SCEVInsertPointGuard Guard(Builder, this); |
304 | |
305 | if (IsSafeToHoist) { |
306 | // Move the insertion point out of as many loops as we can. |
307 | while (const Loop *L = SE.LI.getLoopFor(BB: Builder.GetInsertBlock())) { |
308 | if (!L->isLoopInvariant(V: LHS) || !L->isLoopInvariant(V: RHS)) break; |
309 | BasicBlock * = L->getLoopPreheader(); |
310 | if (!Preheader) break; |
311 | |
312 | // Ok, move up a level. |
313 | Builder.SetInsertPoint(Preheader->getTerminator()); |
314 | } |
315 | } |
316 | |
317 | // If we haven't found this binop, insert it. |
318 | // TODO: Use the Builder, which will make CreateBinOp below fold with |
319 | // InstSimplifyFolder. |
320 | Instruction *BO = Builder.Insert(I: BinaryOperator::Create(Op: Opcode, S1: LHS, S2: RHS)); |
321 | BO->setDebugLoc(Loc); |
322 | if (Flags & SCEV::FlagNUW) |
323 | BO->setHasNoUnsignedWrap(); |
324 | if (Flags & SCEV::FlagNSW) |
325 | BO->setHasNoSignedWrap(); |
326 | |
327 | return BO; |
328 | } |
329 | |
330 | /// expandAddToGEP - Expand an addition expression with a pointer type into |
331 | /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps |
332 | /// BasicAliasAnalysis and other passes analyze the result. See the rules |
333 | /// for getelementptr vs. inttoptr in |
334 | /// http://llvm.org/docs/LangRef.html#pointeraliasing |
335 | /// for details. |
336 | /// |
337 | /// Design note: The correctness of using getelementptr here depends on |
338 | /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as |
339 | /// they may introduce pointer arithmetic which may not be safely converted |
340 | /// into getelementptr. |
341 | /// |
342 | /// Design note: It might seem desirable for this function to be more |
343 | /// loop-aware. If some of the indices are loop-invariant while others |
344 | /// aren't, it might seem desirable to emit multiple GEPs, keeping the |
345 | /// loop-invariant portions of the overall computation outside the loop. |
346 | /// However, there are a few reasons this is not done here. Hoisting simple |
347 | /// arithmetic is a low-level optimization that often isn't very |
348 | /// important until late in the optimization process. In fact, passes |
349 | /// like InstructionCombining will combine GEPs, even if it means |
350 | /// pushing loop-invariant computation down into loops, so even if the |
351 | /// GEPs were split here, the work would quickly be undone. The |
352 | /// LoopStrengthReduction pass, which is usually run quite late (and |
353 | /// after the last InstructionCombining pass), takes care of hoisting |
354 | /// loop-invariant portions of expressions, after considering what |
355 | /// can be folded using target addressing modes. |
356 | /// |
357 | Value *SCEVExpander::expandAddToGEP(const SCEV *Offset, Value *V, |
358 | SCEV::NoWrapFlags Flags) { |
359 | assert(!isa<Instruction>(V) || |
360 | SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint())); |
361 | |
362 | Value *Idx = expand(S: Offset); |
363 | GEPNoWrapFlags NW = (Flags & SCEV::FlagNUW) ? GEPNoWrapFlags::noUnsignedWrap() |
364 | : GEPNoWrapFlags::none(); |
365 | |
366 | // Fold a GEP with constant operands. |
367 | if (Constant *CLHS = dyn_cast<Constant>(Val: V)) |
368 | if (Constant *CRHS = dyn_cast<Constant>(Val: Idx)) |
369 | return Builder.CreatePtrAdd(Ptr: CLHS, Offset: CRHS, Name: "" , NW); |
370 | |
371 | // Do a quick scan to see if we have this GEP nearby. If so, reuse it. |
372 | unsigned ScanLimit = 6; |
373 | BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); |
374 | // Scanning starts from the last instruction before the insertion point. |
375 | BasicBlock::iterator IP = Builder.GetInsertPoint(); |
376 | if (IP != BlockBegin) { |
377 | --IP; |
378 | for (; ScanLimit; --IP, --ScanLimit) { |
379 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val&: IP)) { |
380 | if (GEP->getPointerOperand() == V && |
381 | GEP->getSourceElementType() == Builder.getInt8Ty() && |
382 | GEP->getOperand(i_nocapture: 1) == Idx) { |
383 | rememberFlags(I: GEP); |
384 | GEP->setNoWrapFlags(GEP->getNoWrapFlags() & NW); |
385 | return &*IP; |
386 | } |
387 | } |
388 | if (IP == BlockBegin) break; |
389 | } |
390 | } |
391 | |
392 | // Save the original insertion point so we can restore it when we're done. |
393 | SCEVInsertPointGuard Guard(Builder, this); |
394 | |
395 | // Move the insertion point out of as many loops as we can. |
396 | while (const Loop *L = SE.LI.getLoopFor(BB: Builder.GetInsertBlock())) { |
397 | if (!L->isLoopInvariant(V) || !L->isLoopInvariant(V: Idx)) break; |
398 | BasicBlock * = L->getLoopPreheader(); |
399 | if (!Preheader) break; |
400 | |
401 | // Ok, move up a level. |
402 | Builder.SetInsertPoint(Preheader->getTerminator()); |
403 | } |
404 | |
405 | // Emit a GEP. |
406 | return Builder.CreatePtrAdd(Ptr: V, Offset: Idx, Name: "scevgep" , NW); |
407 | } |
408 | |
409 | /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for |
410 | /// SCEV expansion. If they are nested, this is the most nested. If they are |
411 | /// neighboring, pick the later. |
412 | static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, |
413 | DominatorTree &DT) { |
414 | if (!A) return B; |
415 | if (!B) return A; |
416 | if (A->contains(L: B)) return B; |
417 | if (B->contains(L: A)) return A; |
418 | if (DT.dominates(A: A->getHeader(), B: B->getHeader())) return B; |
419 | if (DT.dominates(A: B->getHeader(), B: A->getHeader())) return A; |
420 | return A; // Arbitrarily break the tie. |
421 | } |
422 | |
423 | /// getRelevantLoop - Get the most relevant loop associated with the given |
424 | /// expression, according to PickMostRelevantLoop. |
425 | const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { |
426 | // Test whether we've already computed the most relevant loop for this SCEV. |
427 | auto Pair = RelevantLoops.try_emplace(Key: S); |
428 | if (!Pair.second) |
429 | return Pair.first->second; |
430 | |
431 | switch (S->getSCEVType()) { |
432 | case scConstant: |
433 | case scVScale: |
434 | return nullptr; // A constant has no relevant loops. |
435 | case scTruncate: |
436 | case scZeroExtend: |
437 | case scSignExtend: |
438 | case scPtrToInt: |
439 | case scAddExpr: |
440 | case scMulExpr: |
441 | case scUDivExpr: |
442 | case scAddRecExpr: |
443 | case scUMaxExpr: |
444 | case scSMaxExpr: |
445 | case scUMinExpr: |
446 | case scSMinExpr: |
447 | case scSequentialUMinExpr: { |
448 | const Loop *L = nullptr; |
449 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) |
450 | L = AR->getLoop(); |
451 | for (const SCEV *Op : S->operands()) |
452 | L = PickMostRelevantLoop(A: L, B: getRelevantLoop(S: Op), DT&: SE.DT); |
453 | return RelevantLoops[S] = L; |
454 | } |
455 | case scUnknown: { |
456 | const SCEVUnknown *U = cast<SCEVUnknown>(Val: S); |
457 | if (const Instruction *I = dyn_cast<Instruction>(Val: U->getValue())) |
458 | return Pair.first->second = SE.LI.getLoopFor(BB: I->getParent()); |
459 | // A non-instruction has no relevant loops. |
460 | return nullptr; |
461 | } |
462 | case scCouldNotCompute: |
463 | llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!" ); |
464 | } |
465 | llvm_unreachable("Unexpected SCEV type!" ); |
466 | } |
467 | |
468 | namespace { |
469 | |
470 | /// LoopCompare - Compare loops by PickMostRelevantLoop. |
471 | class LoopCompare { |
472 | DominatorTree &DT; |
473 | public: |
474 | explicit LoopCompare(DominatorTree &dt) : DT(dt) {} |
475 | |
476 | bool operator()(std::pair<const Loop *, const SCEV *> LHS, |
477 | std::pair<const Loop *, const SCEV *> RHS) const { |
478 | // Keep pointer operands sorted at the end. |
479 | if (LHS.second->getType()->isPointerTy() != |
480 | RHS.second->getType()->isPointerTy()) |
481 | return LHS.second->getType()->isPointerTy(); |
482 | |
483 | // Compare loops with PickMostRelevantLoop. |
484 | if (LHS.first != RHS.first) |
485 | return PickMostRelevantLoop(A: LHS.first, B: RHS.first, DT) != LHS.first; |
486 | |
487 | // If one operand is a non-constant negative and the other is not, |
488 | // put the non-constant negative on the right so that a sub can |
489 | // be used instead of a negate and add. |
490 | if (LHS.second->isNonConstantNegative()) { |
491 | if (!RHS.second->isNonConstantNegative()) |
492 | return false; |
493 | } else if (RHS.second->isNonConstantNegative()) |
494 | return true; |
495 | |
496 | // Otherwise they are equivalent according to this comparison. |
497 | return false; |
498 | } |
499 | }; |
500 | |
501 | } |
502 | |
503 | Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { |
504 | // Recognize the canonical representation of an unsimplifed urem. |
505 | const SCEV *URemLHS = nullptr; |
506 | const SCEV *URemRHS = nullptr; |
507 | if (SE.matchURem(Expr: S, LHS&: URemLHS, RHS&: URemRHS)) { |
508 | Value *LHS = expand(S: URemLHS); |
509 | Value *RHS = expand(S: URemRHS); |
510 | return InsertBinop(Opcode: Instruction::URem, LHS, RHS, Flags: SCEV::FlagAnyWrap, |
511 | /*IsSafeToHoist*/ false); |
512 | } |
513 | |
514 | // Collect all the add operands in a loop, along with their associated loops. |
515 | // Iterate in reverse so that constants are emitted last, all else equal, and |
516 | // so that pointer operands are inserted first, which the code below relies on |
517 | // to form more involved GEPs. |
518 | SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; |
519 | for (const SCEV *Op : reverse(C: S->operands())) |
520 | OpsAndLoops.push_back(Elt: std::make_pair(x: getRelevantLoop(S: Op), y&: Op)); |
521 | |
522 | // Sort by loop. Use a stable sort so that constants follow non-constants and |
523 | // pointer operands precede non-pointer operands. |
524 | llvm::stable_sort(Range&: OpsAndLoops, C: LoopCompare(SE.DT)); |
525 | |
526 | // Emit instructions to add all the operands. Hoist as much as possible |
527 | // out of loops, and form meaningful getelementptrs where possible. |
528 | Value *Sum = nullptr; |
529 | for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) { |
530 | const Loop *CurLoop = I->first; |
531 | const SCEV *Op = I->second; |
532 | if (!Sum) { |
533 | // This is the first operand. Just expand it. |
534 | Sum = expand(S: Op); |
535 | ++I; |
536 | continue; |
537 | } |
538 | |
539 | assert(!Op->getType()->isPointerTy() && "Only first op can be pointer" ); |
540 | if (isa<PointerType>(Val: Sum->getType())) { |
541 | // The running sum expression is a pointer. Try to form a getelementptr |
542 | // at this level with that as the base. |
543 | SmallVector<const SCEV *, 4> NewOps; |
544 | for (; I != E && I->first == CurLoop; ++I) { |
545 | // If the operand is SCEVUnknown and not instructions, peek through |
546 | // it, to enable more of it to be folded into the GEP. |
547 | const SCEV *X = I->second; |
548 | if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Val: X)) |
549 | if (!isa<Instruction>(Val: U->getValue())) |
550 | X = SE.getSCEV(V: U->getValue()); |
551 | NewOps.push_back(Elt: X); |
552 | } |
553 | Sum = expandAddToGEP(Offset: SE.getAddExpr(Ops&: NewOps), V: Sum, Flags: S->getNoWrapFlags()); |
554 | } else if (Op->isNonConstantNegative()) { |
555 | // Instead of doing a negate and add, just do a subtract. |
556 | Value *W = expand(S: SE.getNegativeSCEV(V: Op)); |
557 | Sum = InsertBinop(Opcode: Instruction::Sub, LHS: Sum, RHS: W, Flags: SCEV::FlagAnyWrap, |
558 | /*IsSafeToHoist*/ true); |
559 | ++I; |
560 | } else { |
561 | // A simple add. |
562 | Value *W = expand(S: Op); |
563 | // Canonicalize a constant to the RHS. |
564 | if (isa<Constant>(Val: Sum)) |
565 | std::swap(a&: Sum, b&: W); |
566 | Sum = InsertBinop(Opcode: Instruction::Add, LHS: Sum, RHS: W, Flags: S->getNoWrapFlags(), |
567 | /*IsSafeToHoist*/ true); |
568 | ++I; |
569 | } |
570 | } |
571 | |
572 | return Sum; |
573 | } |
574 | |
575 | Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { |
576 | Type *Ty = S->getType(); |
577 | |
578 | // Collect all the mul operands in a loop, along with their associated loops. |
579 | // Iterate in reverse so that constants are emitted last, all else equal. |
580 | SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; |
581 | for (const SCEV *Op : reverse(C: S->operands())) |
582 | OpsAndLoops.push_back(Elt: std::make_pair(x: getRelevantLoop(S: Op), y&: Op)); |
583 | |
584 | // Sort by loop. Use a stable sort so that constants follow non-constants. |
585 | llvm::stable_sort(Range&: OpsAndLoops, C: LoopCompare(SE.DT)); |
586 | |
587 | // Emit instructions to mul all the operands. Hoist as much as possible |
588 | // out of loops. |
589 | Value *Prod = nullptr; |
590 | auto I = OpsAndLoops.begin(); |
591 | |
592 | // Expand the calculation of X pow N in the following manner: |
593 | // Let N = P1 + P2 + ... + PK, where all P are powers of 2. Then: |
594 | // X pow N = (X pow P1) * (X pow P2) * ... * (X pow PK). |
595 | const auto ExpandOpBinPowN = [this, &I, &OpsAndLoops]() { |
596 | auto E = I; |
597 | // Calculate how many times the same operand from the same loop is included |
598 | // into this power. |
599 | uint64_t Exponent = 0; |
600 | const uint64_t MaxExponent = UINT64_MAX >> 1; |
601 | // No one sane will ever try to calculate such huge exponents, but if we |
602 | // need this, we stop on UINT64_MAX / 2 because we need to exit the loop |
603 | // below when the power of 2 exceeds our Exponent, and we want it to be |
604 | // 1u << 31 at most to not deal with unsigned overflow. |
605 | while (E != OpsAndLoops.end() && *I == *E && Exponent != MaxExponent) { |
606 | ++Exponent; |
607 | ++E; |
608 | } |
609 | assert(Exponent > 0 && "Trying to calculate a zeroth exponent of operand?" ); |
610 | |
611 | // Calculate powers with exponents 1, 2, 4, 8 etc. and include those of them |
612 | // that are needed into the result. |
613 | Value *P = expand(S: I->second); |
614 | Value *Result = nullptr; |
615 | if (Exponent & 1) |
616 | Result = P; |
617 | for (uint64_t BinExp = 2; BinExp <= Exponent; BinExp <<= 1) { |
618 | P = InsertBinop(Opcode: Instruction::Mul, LHS: P, RHS: P, Flags: SCEV::FlagAnyWrap, |
619 | /*IsSafeToHoist*/ true); |
620 | if (Exponent & BinExp) |
621 | Result = Result ? InsertBinop(Opcode: Instruction::Mul, LHS: Result, RHS: P, |
622 | Flags: SCEV::FlagAnyWrap, |
623 | /*IsSafeToHoist*/ true) |
624 | : P; |
625 | } |
626 | |
627 | I = E; |
628 | assert(Result && "Nothing was expanded?" ); |
629 | return Result; |
630 | }; |
631 | |
632 | while (I != OpsAndLoops.end()) { |
633 | if (!Prod) { |
634 | // This is the first operand. Just expand it. |
635 | Prod = ExpandOpBinPowN(); |
636 | } else if (I->second->isAllOnesValue()) { |
637 | // Instead of doing a multiply by negative one, just do a negate. |
638 | Prod = InsertBinop(Opcode: Instruction::Sub, LHS: Constant::getNullValue(Ty), RHS: Prod, |
639 | Flags: SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); |
640 | ++I; |
641 | } else { |
642 | // A simple mul. |
643 | Value *W = ExpandOpBinPowN(); |
644 | // Canonicalize a constant to the RHS. |
645 | if (isa<Constant>(Val: Prod)) std::swap(a&: Prod, b&: W); |
646 | const APInt *RHS; |
647 | if (match(V: W, P: m_Power2(V&: RHS))) { |
648 | // Canonicalize Prod*(1<<C) to Prod<<C. |
649 | assert(!Ty->isVectorTy() && "vector types are not SCEVable" ); |
650 | auto NWFlags = S->getNoWrapFlags(); |
651 | // clear nsw flag if shl will produce poison value. |
652 | if (RHS->logBase2() == RHS->getBitWidth() - 1) |
653 | NWFlags = ScalarEvolution::clearFlags(Flags: NWFlags, OffFlags: SCEV::FlagNSW); |
654 | Prod = InsertBinop(Opcode: Instruction::Shl, LHS: Prod, |
655 | RHS: ConstantInt::get(Ty, V: RHS->logBase2()), Flags: NWFlags, |
656 | /*IsSafeToHoist*/ true); |
657 | } else { |
658 | Prod = InsertBinop(Opcode: Instruction::Mul, LHS: Prod, RHS: W, Flags: S->getNoWrapFlags(), |
659 | /*IsSafeToHoist*/ true); |
660 | } |
661 | } |
662 | } |
663 | |
664 | return Prod; |
665 | } |
666 | |
667 | Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { |
668 | Value *LHS = expand(S: S->getLHS()); |
669 | if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Val: S->getRHS())) { |
670 | const APInt &RHS = SC->getAPInt(); |
671 | if (RHS.isPowerOf2()) |
672 | return InsertBinop(Opcode: Instruction::LShr, LHS, |
673 | RHS: ConstantInt::get(Ty: SC->getType(), V: RHS.logBase2()), |
674 | Flags: SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); |
675 | } |
676 | |
677 | const SCEV *RHSExpr = S->getRHS(); |
678 | Value *RHS = expand(S: RHSExpr); |
679 | if (SafeUDivMode) { |
680 | bool GuaranteedNotPoison = |
681 | ScalarEvolution::isGuaranteedNotToBePoison(Op: RHSExpr); |
682 | if (!GuaranteedNotPoison) |
683 | RHS = Builder.CreateFreeze(V: RHS); |
684 | |
685 | // We need an umax if either RHSExpr is not known to be zero, or if it is |
686 | // not guaranteed to be non-poison. In the later case, the frozen poison may |
687 | // be 0. |
688 | if (!SE.isKnownNonZero(S: RHSExpr) || !GuaranteedNotPoison) |
689 | RHS = Builder.CreateIntrinsic(RetTy: RHS->getType(), ID: Intrinsic::umax, |
690 | Args: {RHS, ConstantInt::get(Ty: RHS->getType(), V: 1)}); |
691 | } |
692 | return InsertBinop(Opcode: Instruction::UDiv, LHS, RHS, Flags: SCEV::FlagAnyWrap, |
693 | /*IsSafeToHoist*/ SE.isKnownNonZero(S: S->getRHS())); |
694 | } |
695 | |
696 | /// Determine if this is a well-behaved chain of instructions leading back to |
697 | /// the PHI. If so, it may be reused by expanded expressions. |
698 | bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, |
699 | const Loop *L) { |
700 | if (IncV->getNumOperands() == 0 || isa<PHINode>(Val: IncV) || |
701 | (isa<CastInst>(Val: IncV) && !isa<BitCastInst>(Val: IncV))) |
702 | return false; |
703 | // If any of the operands don't dominate the insert position, bail. |
704 | // Addrec operands are always loop-invariant, so this can only happen |
705 | // if there are instructions which haven't been hoisted. |
706 | if (L == IVIncInsertLoop) { |
707 | for (Use &Op : llvm::drop_begin(RangeOrContainer: IncV->operands())) |
708 | if (Instruction *OInst = dyn_cast<Instruction>(Val&: Op)) |
709 | if (!SE.DT.dominates(Def: OInst, User: IVIncInsertPos)) |
710 | return false; |
711 | } |
712 | // Advance to the next instruction. |
713 | IncV = dyn_cast<Instruction>(Val: IncV->getOperand(i: 0)); |
714 | if (!IncV) |
715 | return false; |
716 | |
717 | if (IncV->mayHaveSideEffects()) |
718 | return false; |
719 | |
720 | if (IncV == PN) |
721 | return true; |
722 | |
723 | return isNormalAddRecExprPHI(PN, IncV, L); |
724 | } |
725 | |
726 | /// getIVIncOperand returns an induction variable increment's induction |
727 | /// variable operand. |
728 | /// |
729 | /// If allowScale is set, any type of GEP is allowed as long as the nonIV |
730 | /// operands dominate InsertPos. |
731 | /// |
732 | /// If allowScale is not set, ensure that a GEP increment conforms to one of the |
733 | /// simple patterns generated by getAddRecExprPHILiterally and |
734 | /// expandAddtoGEP. If the pattern isn't recognized, return NULL. |
735 | Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, |
736 | Instruction *InsertPos, |
737 | bool allowScale) { |
738 | if (IncV == InsertPos) |
739 | return nullptr; |
740 | |
741 | switch (IncV->getOpcode()) { |
742 | default: |
743 | return nullptr; |
744 | // Check for a simple Add/Sub or GEP of a loop invariant step. |
745 | case Instruction::Add: |
746 | case Instruction::Sub: { |
747 | Instruction *OInst = dyn_cast<Instruction>(Val: IncV->getOperand(i: 1)); |
748 | if (!OInst || SE.DT.dominates(Def: OInst, User: InsertPos)) |
749 | return dyn_cast<Instruction>(Val: IncV->getOperand(i: 0)); |
750 | return nullptr; |
751 | } |
752 | case Instruction::BitCast: |
753 | return dyn_cast<Instruction>(Val: IncV->getOperand(i: 0)); |
754 | case Instruction::GetElementPtr: |
755 | for (Use &U : llvm::drop_begin(RangeOrContainer: IncV->operands())) { |
756 | if (isa<Constant>(Val: U)) |
757 | continue; |
758 | if (Instruction *OInst = dyn_cast<Instruction>(Val&: U)) { |
759 | if (!SE.DT.dominates(Def: OInst, User: InsertPos)) |
760 | return nullptr; |
761 | } |
762 | if (allowScale) { |
763 | // allow any kind of GEP as long as it can be hoisted. |
764 | continue; |
765 | } |
766 | // GEPs produced by SCEVExpander use i8 element type. |
767 | if (!cast<GEPOperator>(Val: IncV)->getSourceElementType()->isIntegerTy(Bitwidth: 8)) |
768 | return nullptr; |
769 | break; |
770 | } |
771 | return dyn_cast<Instruction>(Val: IncV->getOperand(i: 0)); |
772 | } |
773 | } |
774 | |
775 | /// If the insert point of the current builder or any of the builders on the |
776 | /// stack of saved builders has 'I' as its insert point, update it to point to |
777 | /// the instruction after 'I'. This is intended to be used when the instruction |
778 | /// 'I' is being moved. If this fixup is not done and 'I' is moved to a |
779 | /// different block, the inconsistent insert point (with a mismatched |
780 | /// Instruction and Block) can lead to an instruction being inserted in a block |
781 | /// other than its parent. |
782 | void SCEVExpander::fixupInsertPoints(Instruction *I) { |
783 | BasicBlock::iterator It(*I); |
784 | BasicBlock::iterator NewInsertPt = std::next(x: It); |
785 | if (Builder.GetInsertPoint() == It) |
786 | Builder.SetInsertPoint(&*NewInsertPt); |
787 | for (auto *InsertPtGuard : InsertPointGuards) |
788 | if (InsertPtGuard->GetInsertPoint() == It) |
789 | InsertPtGuard->SetInsertPoint(NewInsertPt); |
790 | } |
791 | |
792 | /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make |
793 | /// it available to other uses in this loop. Recursively hoist any operands, |
794 | /// until we reach a value that dominates InsertPos. |
795 | bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos, |
796 | bool RecomputePoisonFlags) { |
797 | auto FixupPoisonFlags = [this](Instruction *I) { |
798 | // Drop flags that are potentially inferred from old context and infer flags |
799 | // in new context. |
800 | rememberFlags(I); |
801 | I->dropPoisonGeneratingFlags(); |
802 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: I)) |
803 | if (auto Flags = SE.getStrengthenedNoWrapFlagsFromBinOp(OBO)) { |
804 | auto *BO = cast<BinaryOperator>(Val: I); |
805 | BO->setHasNoUnsignedWrap( |
806 | ScalarEvolution::maskFlags(Flags: *Flags, Mask: SCEV::FlagNUW) == SCEV::FlagNUW); |
807 | BO->setHasNoSignedWrap( |
808 | ScalarEvolution::maskFlags(Flags: *Flags, Mask: SCEV::FlagNSW) == SCEV::FlagNSW); |
809 | } |
810 | }; |
811 | |
812 | if (SE.DT.dominates(Def: IncV, User: InsertPos)) { |
813 | if (RecomputePoisonFlags) |
814 | FixupPoisonFlags(IncV); |
815 | return true; |
816 | } |
817 | |
818 | // InsertPos must itself dominate IncV so that IncV's new position satisfies |
819 | // its existing users. |
820 | if (isa<PHINode>(Val: InsertPos) || |
821 | !SE.DT.dominates(A: InsertPos->getParent(), B: IncV->getParent())) |
822 | return false; |
823 | |
824 | if (!SE.LI.movementPreservesLCSSAForm(Inst: IncV, NewLoc: InsertPos)) |
825 | return false; |
826 | |
827 | // Check that the chain of IV operands leading back to Phi can be hoisted. |
828 | SmallVector<Instruction*, 4> IVIncs; |
829 | for(;;) { |
830 | Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); |
831 | if (!Oper) |
832 | return false; |
833 | // IncV is safe to hoist. |
834 | IVIncs.push_back(Elt: IncV); |
835 | IncV = Oper; |
836 | if (SE.DT.dominates(Def: IncV, User: InsertPos)) |
837 | break; |
838 | } |
839 | for (Instruction *I : llvm::reverse(C&: IVIncs)) { |
840 | fixupInsertPoints(I); |
841 | I->moveBefore(InsertPos: InsertPos->getIterator()); |
842 | if (RecomputePoisonFlags) |
843 | FixupPoisonFlags(I); |
844 | } |
845 | return true; |
846 | } |
847 | |
848 | bool SCEVExpander::canReuseFlagsFromOriginalIVInc(PHINode *OrigPhi, |
849 | PHINode *WidePhi, |
850 | Instruction *OrigInc, |
851 | Instruction *WideInc) { |
852 | return match(V: OrigInc, P: m_c_BinOp(L: m_Specific(V: OrigPhi), R: m_Value())) && |
853 | match(V: WideInc, P: m_c_BinOp(L: m_Specific(V: WidePhi), R: m_Value())) && |
854 | OrigInc->getOpcode() == WideInc->getOpcode(); |
855 | } |
856 | |
857 | /// Determine if this cyclic phi is in a form that would have been generated by |
858 | /// LSR. We don't care if the phi was actually expanded in this pass, as long |
859 | /// as it is in a low-cost form, for example, no implied multiplication. This |
860 | /// should match any patterns generated by getAddRecExprPHILiterally and |
861 | /// expandAddtoGEP. |
862 | bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, |
863 | const Loop *L) { |
864 | for(Instruction *IVOper = IncV; |
865 | (IVOper = getIVIncOperand(IncV: IVOper, InsertPos: L->getLoopPreheader()->getTerminator(), |
866 | /*allowScale=*/false));) { |
867 | if (IVOper == PN) |
868 | return true; |
869 | } |
870 | return false; |
871 | } |
872 | |
873 | /// expandIVInc - Expand an IV increment at Builder's current InsertPos. |
874 | /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may |
875 | /// need to materialize IV increments elsewhere to handle difficult situations. |
876 | Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, |
877 | bool useSubtract) { |
878 | Value *IncV; |
879 | // If the PHI is a pointer, use a GEP, otherwise use an add or sub. |
880 | if (PN->getType()->isPointerTy()) { |
881 | // TODO: Change name to IVName.iv.next. |
882 | IncV = Builder.CreatePtrAdd(Ptr: PN, Offset: StepV, Name: "scevgep" ); |
883 | } else { |
884 | IncV = useSubtract ? |
885 | Builder.CreateSub(LHS: PN, RHS: StepV, Name: Twine(IVName) + ".iv.next" ) : |
886 | Builder.CreateAdd(LHS: PN, RHS: StepV, Name: Twine(IVName) + ".iv.next" ); |
887 | } |
888 | return IncV; |
889 | } |
890 | |
891 | /// Check whether we can cheaply express the requested SCEV in terms of |
892 | /// the available PHI SCEV by truncation and/or inversion of the step. |
893 | static bool canBeCheaplyTransformed(ScalarEvolution &SE, |
894 | const SCEVAddRecExpr *Phi, |
895 | const SCEVAddRecExpr *Requested, |
896 | bool &InvertStep) { |
897 | // We can't transform to match a pointer PHI. |
898 | Type *PhiTy = Phi->getType(); |
899 | Type *RequestedTy = Requested->getType(); |
900 | if (PhiTy->isPointerTy() || RequestedTy->isPointerTy()) |
901 | return false; |
902 | |
903 | if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) |
904 | return false; |
905 | |
906 | // Try truncate it if necessary. |
907 | Phi = dyn_cast<SCEVAddRecExpr>(Val: SE.getTruncateOrNoop(V: Phi, Ty: RequestedTy)); |
908 | if (!Phi) |
909 | return false; |
910 | |
911 | // Check whether truncation will help. |
912 | if (Phi == Requested) { |
913 | InvertStep = false; |
914 | return true; |
915 | } |
916 | |
917 | // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. |
918 | if (SE.getMinusSCEV(LHS: Requested->getStart(), RHS: Requested) == Phi) { |
919 | InvertStep = true; |
920 | return true; |
921 | } |
922 | |
923 | return false; |
924 | } |
925 | |
926 | static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { |
927 | if (!isa<IntegerType>(Val: AR->getType())) |
928 | return false; |
929 | |
930 | unsigned BitWidth = cast<IntegerType>(Val: AR->getType())->getBitWidth(); |
931 | Type *WideTy = IntegerType::get(C&: AR->getType()->getContext(), NumBits: BitWidth * 2); |
932 | const SCEV *Step = AR->getStepRecurrence(SE); |
933 | const SCEV *OpAfterExtend = SE.getAddExpr(LHS: SE.getSignExtendExpr(Op: Step, Ty: WideTy), |
934 | RHS: SE.getSignExtendExpr(Op: AR, Ty: WideTy)); |
935 | const SCEV *ExtendAfterOp = |
936 | SE.getSignExtendExpr(Op: SE.getAddExpr(LHS: AR, RHS: Step), Ty: WideTy); |
937 | return ExtendAfterOp == OpAfterExtend; |
938 | } |
939 | |
940 | static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { |
941 | if (!isa<IntegerType>(Val: AR->getType())) |
942 | return false; |
943 | |
944 | unsigned BitWidth = cast<IntegerType>(Val: AR->getType())->getBitWidth(); |
945 | Type *WideTy = IntegerType::get(C&: AR->getType()->getContext(), NumBits: BitWidth * 2); |
946 | const SCEV *Step = AR->getStepRecurrence(SE); |
947 | const SCEV *OpAfterExtend = SE.getAddExpr(LHS: SE.getZeroExtendExpr(Op: Step, Ty: WideTy), |
948 | RHS: SE.getZeroExtendExpr(Op: AR, Ty: WideTy)); |
949 | const SCEV *ExtendAfterOp = |
950 | SE.getZeroExtendExpr(Op: SE.getAddExpr(LHS: AR, RHS: Step), Ty: WideTy); |
951 | return ExtendAfterOp == OpAfterExtend; |
952 | } |
953 | |
954 | /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand |
955 | /// the base addrec, which is the addrec without any non-loop-dominating |
956 | /// values, and return the PHI. |
957 | PHINode * |
958 | SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, |
959 | const Loop *L, Type *&TruncTy, |
960 | bool &InvertStep) { |
961 | assert((!IVIncInsertLoop || IVIncInsertPos) && |
962 | "Uninitialized insert position" ); |
963 | |
964 | // Reuse a previously-inserted PHI, if present. |
965 | BasicBlock *LatchBlock = L->getLoopLatch(); |
966 | if (LatchBlock) { |
967 | PHINode *AddRecPhiMatch = nullptr; |
968 | Instruction *IncV = nullptr; |
969 | TruncTy = nullptr; |
970 | InvertStep = false; |
971 | |
972 | // Only try partially matching scevs that need truncation and/or |
973 | // step-inversion if we know this loop is outside the current loop. |
974 | bool TryNonMatchingSCEV = |
975 | IVIncInsertLoop && |
976 | SE.DT.properlyDominates(A: LatchBlock, B: IVIncInsertLoop->getHeader()); |
977 | |
978 | for (PHINode &PN : L->getHeader()->phis()) { |
979 | if (!SE.isSCEVable(Ty: PN.getType())) |
980 | continue; |
981 | |
982 | // We should not look for a incomplete PHI. Getting SCEV for a incomplete |
983 | // PHI has no meaning at all. |
984 | if (!PN.isComplete()) { |
985 | SCEV_DEBUG_WITH_TYPE( |
986 | DebugType, dbgs() << "One incomplete PHI is found: " << PN << "\n" ); |
987 | continue; |
988 | } |
989 | |
990 | const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: &PN)); |
991 | if (!PhiSCEV) |
992 | continue; |
993 | |
994 | bool IsMatchingSCEV = PhiSCEV == Normalized; |
995 | // We only handle truncation and inversion of phi recurrences for the |
996 | // expanded expression if the expanded expression's loop dominates the |
997 | // loop we insert to. Check now, so we can bail out early. |
998 | if (!IsMatchingSCEV && !TryNonMatchingSCEV) |
999 | continue; |
1000 | |
1001 | // TODO: this possibly can be reworked to avoid this cast at all. |
1002 | Instruction *TempIncV = |
1003 | dyn_cast<Instruction>(Val: PN.getIncomingValueForBlock(BB: LatchBlock)); |
1004 | if (!TempIncV) |
1005 | continue; |
1006 | |
1007 | // Check whether we can reuse this PHI node. |
1008 | if (LSRMode) { |
1009 | if (!isExpandedAddRecExprPHI(PN: &PN, IncV: TempIncV, L)) |
1010 | continue; |
1011 | } else { |
1012 | if (!isNormalAddRecExprPHI(PN: &PN, IncV: TempIncV, L)) |
1013 | continue; |
1014 | } |
1015 | |
1016 | // Stop if we have found an exact match SCEV. |
1017 | if (IsMatchingSCEV) { |
1018 | IncV = TempIncV; |
1019 | TruncTy = nullptr; |
1020 | InvertStep = false; |
1021 | AddRecPhiMatch = &PN; |
1022 | break; |
1023 | } |
1024 | |
1025 | // Try whether the phi can be translated into the requested form |
1026 | // (truncated and/or offset by a constant). |
1027 | if ((!TruncTy || InvertStep) && |
1028 | canBeCheaplyTransformed(SE, Phi: PhiSCEV, Requested: Normalized, InvertStep)) { |
1029 | // Record the phi node. But don't stop we might find an exact match |
1030 | // later. |
1031 | AddRecPhiMatch = &PN; |
1032 | IncV = TempIncV; |
1033 | TruncTy = Normalized->getType(); |
1034 | } |
1035 | } |
1036 | |
1037 | if (AddRecPhiMatch) { |
1038 | // Ok, the add recurrence looks usable. |
1039 | // Remember this PHI, even in post-inc mode. |
1040 | InsertedValues.insert(V: AddRecPhiMatch); |
1041 | // Remember the increment. |
1042 | rememberInstruction(I: IncV); |
1043 | // Those values were not actually inserted but re-used. |
1044 | ReusedValues.insert(Ptr: AddRecPhiMatch); |
1045 | ReusedValues.insert(Ptr: IncV); |
1046 | return AddRecPhiMatch; |
1047 | } |
1048 | } |
1049 | |
1050 | // Save the original insertion point so we can restore it when we're done. |
1051 | SCEVInsertPointGuard Guard(Builder, this); |
1052 | |
1053 | // Another AddRec may need to be recursively expanded below. For example, if |
1054 | // this AddRec is quadratic, the StepV may itself be an AddRec in this |
1055 | // loop. Remove this loop from the PostIncLoops set before expanding such |
1056 | // AddRecs. Otherwise, we cannot find a valid position for the step |
1057 | // (i.e. StepV can never dominate its loop header). Ideally, we could do |
1058 | // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, |
1059 | // so it's not worth implementing SmallPtrSet::swap. |
1060 | PostIncLoopSet SavedPostIncLoops = PostIncLoops; |
1061 | PostIncLoops.clear(); |
1062 | |
1063 | // Expand code for the start value into the loop preheader. |
1064 | assert(L->getLoopPreheader() && |
1065 | "Can't expand add recurrences without a loop preheader!" ); |
1066 | Value *StartV = |
1067 | expand(S: Normalized->getStart(), I: L->getLoopPreheader()->getTerminator()); |
1068 | |
1069 | // StartV must have been be inserted into L's preheader to dominate the new |
1070 | // phi. |
1071 | assert(!isa<Instruction>(StartV) || |
1072 | SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(), |
1073 | L->getHeader())); |
1074 | |
1075 | // Expand code for the step value. Do this before creating the PHI so that PHI |
1076 | // reuse code doesn't see an incomplete PHI. |
1077 | const SCEV *Step = Normalized->getStepRecurrence(SE); |
1078 | Type *ExpandTy = Normalized->getType(); |
1079 | // If the stride is negative, insert a sub instead of an add for the increment |
1080 | // (unless it's a constant, because subtracts of constants are canonicalized |
1081 | // to adds). |
1082 | bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); |
1083 | if (useSubtract) |
1084 | Step = SE.getNegativeSCEV(V: Step); |
1085 | // Expand the step somewhere that dominates the loop header. |
1086 | Value *StepV = expand(S: Step, I: L->getHeader()->getFirstInsertionPt()); |
1087 | |
1088 | // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if |
1089 | // we actually do emit an addition. It does not apply if we emit a |
1090 | // subtraction. |
1091 | bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, AR: Normalized); |
1092 | bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, AR: Normalized); |
1093 | |
1094 | // Create the PHI. |
1095 | BasicBlock * = L->getHeader(); |
1096 | Builder.SetInsertPoint(TheBB: Header, IP: Header->begin()); |
1097 | PHINode *PN = |
1098 | Builder.CreatePHI(Ty: ExpandTy, NumReservedValues: pred_size(BB: Header), Name: Twine(IVName) + ".iv" ); |
1099 | |
1100 | // Create the step instructions and populate the PHI. |
1101 | for (BasicBlock *Pred : predecessors(BB: Header)) { |
1102 | // Add a start value. |
1103 | if (!L->contains(BB: Pred)) { |
1104 | PN->addIncoming(V: StartV, BB: Pred); |
1105 | continue; |
1106 | } |
1107 | |
1108 | // Create a step value and add it to the PHI. |
1109 | // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the |
1110 | // instructions at IVIncInsertPos. |
1111 | Instruction *InsertPos = L == IVIncInsertLoop ? |
1112 | IVIncInsertPos : Pred->getTerminator(); |
1113 | Builder.SetInsertPoint(InsertPos); |
1114 | Value *IncV = expandIVInc(PN, StepV, L, useSubtract); |
1115 | |
1116 | if (isa<OverflowingBinaryOperator>(Val: IncV)) { |
1117 | if (IncrementIsNUW) |
1118 | cast<BinaryOperator>(Val: IncV)->setHasNoUnsignedWrap(); |
1119 | if (IncrementIsNSW) |
1120 | cast<BinaryOperator>(Val: IncV)->setHasNoSignedWrap(); |
1121 | } |
1122 | PN->addIncoming(V: IncV, BB: Pred); |
1123 | } |
1124 | |
1125 | // After expanding subexpressions, restore the PostIncLoops set so the caller |
1126 | // can ensure that IVIncrement dominates the current uses. |
1127 | PostIncLoops = SavedPostIncLoops; |
1128 | |
1129 | // Remember this PHI, even in post-inc mode. LSR SCEV-based salvaging is most |
1130 | // effective when we are able to use an IV inserted here, so record it. |
1131 | InsertedValues.insert(V: PN); |
1132 | InsertedIVs.push_back(Elt: PN); |
1133 | return PN; |
1134 | } |
1135 | |
1136 | Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { |
1137 | const Loop *L = S->getLoop(); |
1138 | |
1139 | // Determine a normalized form of this expression, which is the expression |
1140 | // before any post-inc adjustment is made. |
1141 | const SCEVAddRecExpr *Normalized = S; |
1142 | if (PostIncLoops.count(Ptr: L)) { |
1143 | PostIncLoopSet Loops; |
1144 | Loops.insert(Ptr: L); |
1145 | Normalized = cast<SCEVAddRecExpr>( |
1146 | Val: normalizeForPostIncUse(S, Loops, SE, /*CheckInvertible=*/false)); |
1147 | } |
1148 | |
1149 | [[maybe_unused]] const SCEV *Start = Normalized->getStart(); |
1150 | const SCEV *Step = Normalized->getStepRecurrence(SE); |
1151 | assert(SE.properlyDominates(Start, L->getHeader()) && |
1152 | "Start does not properly dominate loop header" ); |
1153 | assert(SE.dominates(Step, L->getHeader()) && "Step not dominate loop header" ); |
1154 | |
1155 | // In some cases, we decide to reuse an existing phi node but need to truncate |
1156 | // it and/or invert the step. |
1157 | Type *TruncTy = nullptr; |
1158 | bool InvertStep = false; |
1159 | PHINode *PN = getAddRecExprPHILiterally(Normalized, L, TruncTy, InvertStep); |
1160 | |
1161 | // Accommodate post-inc mode, if necessary. |
1162 | Value *Result; |
1163 | if (!PostIncLoops.count(Ptr: L)) |
1164 | Result = PN; |
1165 | else { |
1166 | // In PostInc mode, use the post-incremented value. |
1167 | BasicBlock *LatchBlock = L->getLoopLatch(); |
1168 | assert(LatchBlock && "PostInc mode requires a unique loop latch!" ); |
1169 | Result = PN->getIncomingValueForBlock(BB: LatchBlock); |
1170 | |
1171 | // We might be introducing a new use of the post-inc IV that is not poison |
1172 | // safe, in which case we should drop poison generating flags. Only keep |
1173 | // those flags for which SCEV has proven that they always hold. |
1174 | if (isa<OverflowingBinaryOperator>(Val: Result)) { |
1175 | auto *I = cast<Instruction>(Val: Result); |
1176 | if (!S->hasNoUnsignedWrap()) |
1177 | I->setHasNoUnsignedWrap(false); |
1178 | if (!S->hasNoSignedWrap()) |
1179 | I->setHasNoSignedWrap(false); |
1180 | } |
1181 | |
1182 | // For an expansion to use the postinc form, the client must call |
1183 | // expandCodeFor with an InsertPoint that is either outside the PostIncLoop |
1184 | // or dominated by IVIncInsertPos. |
1185 | if (isa<Instruction>(Val: Result) && |
1186 | !SE.DT.dominates(Def: cast<Instruction>(Val: Result), |
1187 | User: &*Builder.GetInsertPoint())) { |
1188 | // The induction variable's postinc expansion does not dominate this use. |
1189 | // IVUsers tries to prevent this case, so it is rare. However, it can |
1190 | // happen when an IVUser outside the loop is not dominated by the latch |
1191 | // block. Adjusting IVIncInsertPos before expansion begins cannot handle |
1192 | // all cases. Consider a phi outside whose operand is replaced during |
1193 | // expansion with the value of the postinc user. Without fundamentally |
1194 | // changing the way postinc users are tracked, the only remedy is |
1195 | // inserting an extra IV increment. StepV might fold into PostLoopOffset, |
1196 | // but hopefully expandCodeFor handles that. |
1197 | bool useSubtract = |
1198 | !S->getType()->isPointerTy() && Step->isNonConstantNegative(); |
1199 | if (useSubtract) |
1200 | Step = SE.getNegativeSCEV(V: Step); |
1201 | Value *StepV; |
1202 | { |
1203 | // Expand the step somewhere that dominates the loop header. |
1204 | SCEVInsertPointGuard Guard(Builder, this); |
1205 | StepV = expand(S: Step, I: L->getHeader()->getFirstInsertionPt()); |
1206 | } |
1207 | Result = expandIVInc(PN, StepV, L, useSubtract); |
1208 | } |
1209 | } |
1210 | |
1211 | // We have decided to reuse an induction variable of a dominating loop. Apply |
1212 | // truncation and/or inversion of the step. |
1213 | if (TruncTy) { |
1214 | // Truncate the result. |
1215 | if (TruncTy != Result->getType()) |
1216 | Result = Builder.CreateTrunc(V: Result, DestTy: TruncTy); |
1217 | |
1218 | // Invert the result. |
1219 | if (InvertStep) |
1220 | Result = Builder.CreateSub(LHS: expand(S: Normalized->getStart()), RHS: Result); |
1221 | } |
1222 | |
1223 | return Result; |
1224 | } |
1225 | |
1226 | Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { |
1227 | // In canonical mode we compute the addrec as an expression of a canonical IV |
1228 | // using evaluateAtIteration and expand the resulting SCEV expression. This |
1229 | // way we avoid introducing new IVs to carry on the computation of the addrec |
1230 | // throughout the loop. |
1231 | // |
1232 | // For nested addrecs evaluateAtIteration might need a canonical IV of a |
1233 | // type wider than the addrec itself. Emitting a canonical IV of the |
1234 | // proper type might produce non-legal types, for example expanding an i64 |
1235 | // {0,+,2,+,1} addrec would need an i65 canonical IV. To avoid this just fall |
1236 | // back to non-canonical mode for nested addrecs. |
1237 | if (!CanonicalMode || (S->getNumOperands() > 2)) |
1238 | return expandAddRecExprLiterally(S); |
1239 | |
1240 | Type *Ty = SE.getEffectiveSCEVType(Ty: S->getType()); |
1241 | const Loop *L = S->getLoop(); |
1242 | |
1243 | // First check for an existing canonical IV in a suitable type. |
1244 | PHINode *CanonicalIV = nullptr; |
1245 | if (PHINode *PN = L->getCanonicalInductionVariable()) |
1246 | if (SE.getTypeSizeInBits(Ty: PN->getType()) >= SE.getTypeSizeInBits(Ty)) |
1247 | CanonicalIV = PN; |
1248 | |
1249 | // Rewrite an AddRec in terms of the canonical induction variable, if |
1250 | // its type is more narrow. |
1251 | if (CanonicalIV && |
1252 | SE.getTypeSizeInBits(Ty: CanonicalIV->getType()) > SE.getTypeSizeInBits(Ty) && |
1253 | !S->getType()->isPointerTy()) { |
1254 | SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); |
1255 | for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) |
1256 | NewOps[i] = SE.getAnyExtendExpr(Op: S->getOperand(i), Ty: CanonicalIV->getType()); |
1257 | Value *V = expand(S: SE.getAddRecExpr(Operands&: NewOps, L: S->getLoop(), |
1258 | Flags: S->getNoWrapFlags(Mask: SCEV::FlagNW))); |
1259 | BasicBlock::iterator NewInsertPt = |
1260 | findInsertPointAfter(I: cast<Instruction>(Val: V), MustDominate: &*Builder.GetInsertPoint()); |
1261 | V = expand(S: SE.getTruncateExpr(Op: SE.getUnknown(V), Ty), I: NewInsertPt); |
1262 | return V; |
1263 | } |
1264 | |
1265 | // {X,+,F} --> X + {0,+,F} |
1266 | if (!S->getStart()->isZero()) { |
1267 | if (isa<PointerType>(Val: S->getType())) { |
1268 | Value *StartV = expand(S: SE.getPointerBase(V: S)); |
1269 | return expandAddToGEP(Offset: SE.removePointerBase(S), V: StartV, |
1270 | Flags: S->getNoWrapFlags(Mask: SCEV::FlagNUW)); |
1271 | } |
1272 | |
1273 | SmallVector<const SCEV *, 4> NewOps(S->operands()); |
1274 | NewOps[0] = SE.getConstant(Ty, V: 0); |
1275 | const SCEV *Rest = SE.getAddRecExpr(Operands&: NewOps, L, |
1276 | Flags: S->getNoWrapFlags(Mask: SCEV::FlagNW)); |
1277 | |
1278 | // Just do a normal add. Pre-expand the operands to suppress folding. |
1279 | // |
1280 | // The LHS and RHS values are factored out of the expand call to make the |
1281 | // output independent of the argument evaluation order. |
1282 | const SCEV *AddExprLHS = SE.getUnknown(V: expand(S: S->getStart())); |
1283 | const SCEV *AddExprRHS = SE.getUnknown(V: expand(S: Rest)); |
1284 | return expand(S: SE.getAddExpr(LHS: AddExprLHS, RHS: AddExprRHS)); |
1285 | } |
1286 | |
1287 | // If we don't yet have a canonical IV, create one. |
1288 | if (!CanonicalIV) { |
1289 | // Create and insert the PHI node for the induction variable in the |
1290 | // specified loop. |
1291 | BasicBlock * = L->getHeader(); |
1292 | pred_iterator HPB = pred_begin(BB: Header), HPE = pred_end(BB: Header); |
1293 | CanonicalIV = PHINode::Create(Ty, NumReservedValues: std::distance(first: HPB, last: HPE), NameStr: "indvar" ); |
1294 | CanonicalIV->insertBefore(InsertPos: Header->begin()); |
1295 | rememberInstruction(I: CanonicalIV); |
1296 | |
1297 | SmallSet<BasicBlock *, 4> PredSeen; |
1298 | Constant *One = ConstantInt::get(Ty, V: 1); |
1299 | for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { |
1300 | BasicBlock *HP = *HPI; |
1301 | if (!PredSeen.insert(Ptr: HP).second) { |
1302 | // There must be an incoming value for each predecessor, even the |
1303 | // duplicates! |
1304 | CanonicalIV->addIncoming(V: CanonicalIV->getIncomingValueForBlock(BB: HP), BB: HP); |
1305 | continue; |
1306 | } |
1307 | |
1308 | if (L->contains(BB: HP)) { |
1309 | // Insert a unit add instruction right before the terminator |
1310 | // corresponding to the back-edge. |
1311 | Instruction *Add = BinaryOperator::CreateAdd(V1: CanonicalIV, V2: One, |
1312 | Name: "indvar.next" , |
1313 | InsertBefore: HP->getTerminator()->getIterator()); |
1314 | Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); |
1315 | rememberInstruction(I: Add); |
1316 | CanonicalIV->addIncoming(V: Add, BB: HP); |
1317 | } else { |
1318 | CanonicalIV->addIncoming(V: Constant::getNullValue(Ty), BB: HP); |
1319 | } |
1320 | } |
1321 | } |
1322 | |
1323 | // {0,+,1} --> Insert a canonical induction variable into the loop! |
1324 | if (S->isAffine() && S->getOperand(i: 1)->isOne()) { |
1325 | assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && |
1326 | "IVs with types different from the canonical IV should " |
1327 | "already have been handled!" ); |
1328 | return CanonicalIV; |
1329 | } |
1330 | |
1331 | // {0,+,F} --> {0,+,1} * F |
1332 | |
1333 | // If this is a simple linear addrec, emit it now as a special case. |
1334 | if (S->isAffine()) // {0,+,F} --> i*F |
1335 | return |
1336 | expand(S: SE.getTruncateOrNoop( |
1337 | V: SE.getMulExpr(LHS: SE.getUnknown(V: CanonicalIV), |
1338 | RHS: SE.getNoopOrAnyExtend(V: S->getOperand(i: 1), |
1339 | Ty: CanonicalIV->getType())), |
1340 | Ty)); |
1341 | |
1342 | // If this is a chain of recurrences, turn it into a closed form, using the |
1343 | // folders, then expandCodeFor the closed form. This allows the folders to |
1344 | // simplify the expression without having to build a bunch of special code |
1345 | // into this folder. |
1346 | const SCEV *IH = SE.getUnknown(V: CanonicalIV); // Get I as a "symbolic" SCEV. |
1347 | |
1348 | // Promote S up to the canonical IV type, if the cast is foldable. |
1349 | const SCEV *NewS = S; |
1350 | const SCEV *Ext = SE.getNoopOrAnyExtend(V: S, Ty: CanonicalIV->getType()); |
1351 | if (isa<SCEVAddRecExpr>(Val: Ext)) |
1352 | NewS = Ext; |
1353 | |
1354 | const SCEV *V = cast<SCEVAddRecExpr>(Val: NewS)->evaluateAtIteration(It: IH, SE); |
1355 | |
1356 | // Truncate the result down to the original type, if needed. |
1357 | const SCEV *T = SE.getTruncateOrNoop(V, Ty); |
1358 | return expand(S: T); |
1359 | } |
1360 | |
1361 | Value *SCEVExpander::visitPtrToIntExpr(const SCEVPtrToIntExpr *S) { |
1362 | Value *V = expand(S: S->getOperand()); |
1363 | return ReuseOrCreateCast(V, Ty: S->getType(), Op: CastInst::PtrToInt, |
1364 | IP: GetOptimalInsertionPointForCastOf(V)); |
1365 | } |
1366 | |
1367 | Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { |
1368 | Value *V = expand(S: S->getOperand()); |
1369 | return Builder.CreateTrunc(V, DestTy: S->getType()); |
1370 | } |
1371 | |
1372 | Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { |
1373 | Value *V = expand(S: S->getOperand()); |
1374 | return Builder.CreateZExt(V, DestTy: S->getType(), Name: "" , |
1375 | IsNonNeg: SE.isKnownNonNegative(S: S->getOperand())); |
1376 | } |
1377 | |
1378 | Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { |
1379 | Value *V = expand(S: S->getOperand()); |
1380 | return Builder.CreateSExt(V, DestTy: S->getType()); |
1381 | } |
1382 | |
1383 | Value *SCEVExpander::expandMinMaxExpr(const SCEVNAryExpr *S, |
1384 | Intrinsic::ID IntrinID, Twine Name, |
1385 | bool IsSequential) { |
1386 | bool PrevSafeMode = SafeUDivMode; |
1387 | SafeUDivMode |= IsSequential; |
1388 | Value *LHS = expand(S: S->getOperand(i: S->getNumOperands() - 1)); |
1389 | Type *Ty = LHS->getType(); |
1390 | if (IsSequential) |
1391 | LHS = Builder.CreateFreeze(V: LHS); |
1392 | for (int i = S->getNumOperands() - 2; i >= 0; --i) { |
1393 | SafeUDivMode = (IsSequential && i != 0) || PrevSafeMode; |
1394 | Value *RHS = expand(S: S->getOperand(i)); |
1395 | if (IsSequential && i != 0) |
1396 | RHS = Builder.CreateFreeze(V: RHS); |
1397 | Value *Sel; |
1398 | if (Ty->isIntegerTy()) |
1399 | Sel = Builder.CreateIntrinsic(ID: IntrinID, Types: {Ty}, Args: {LHS, RHS}, |
1400 | /*FMFSource=*/nullptr, Name); |
1401 | else { |
1402 | Value *ICmp = |
1403 | Builder.CreateICmp(P: MinMaxIntrinsic::getPredicate(ID: IntrinID), LHS, RHS); |
1404 | Sel = Builder.CreateSelect(C: ICmp, True: LHS, False: RHS, Name); |
1405 | } |
1406 | LHS = Sel; |
1407 | } |
1408 | SafeUDivMode = PrevSafeMode; |
1409 | return LHS; |
1410 | } |
1411 | |
1412 | Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { |
1413 | return expandMinMaxExpr(S, IntrinID: Intrinsic::smax, Name: "smax" ); |
1414 | } |
1415 | |
1416 | Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { |
1417 | return expandMinMaxExpr(S, IntrinID: Intrinsic::umax, Name: "umax" ); |
1418 | } |
1419 | |
1420 | Value *SCEVExpander::visitSMinExpr(const SCEVSMinExpr *S) { |
1421 | return expandMinMaxExpr(S, IntrinID: Intrinsic::smin, Name: "smin" ); |
1422 | } |
1423 | |
1424 | Value *SCEVExpander::visitUMinExpr(const SCEVUMinExpr *S) { |
1425 | return expandMinMaxExpr(S, IntrinID: Intrinsic::umin, Name: "umin" ); |
1426 | } |
1427 | |
1428 | Value *SCEVExpander::visitSequentialUMinExpr(const SCEVSequentialUMinExpr *S) { |
1429 | return expandMinMaxExpr(S, IntrinID: Intrinsic::umin, Name: "umin" , /*IsSequential*/true); |
1430 | } |
1431 | |
1432 | Value *SCEVExpander::visitVScale(const SCEVVScale *S) { |
1433 | return Builder.CreateVScale(Ty: S->getType()); |
1434 | } |
1435 | |
1436 | Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, |
1437 | BasicBlock::iterator IP) { |
1438 | setInsertPoint(IP); |
1439 | Value *V = expandCodeFor(SH, Ty); |
1440 | return V; |
1441 | } |
1442 | |
1443 | Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { |
1444 | // Expand the code for this SCEV. |
1445 | Value *V = expand(S: SH); |
1446 | |
1447 | if (Ty && Ty != V->getType()) { |
1448 | assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && |
1449 | "non-trivial casts should be done with the SCEVs directly!" ); |
1450 | V = InsertNoopCastOfTo(V, Ty); |
1451 | } |
1452 | return V; |
1453 | } |
1454 | |
1455 | Value *SCEVExpander::FindValueInExprValueMap( |
1456 | const SCEV *S, const Instruction *InsertPt, |
1457 | SmallVectorImpl<Instruction *> &DropPoisonGeneratingInsts) { |
1458 | // If the expansion is not in CanonicalMode, and the SCEV contains any |
1459 | // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally. |
1460 | if (!CanonicalMode && SE.containsAddRecurrence(S)) |
1461 | return nullptr; |
1462 | |
1463 | // If S is a constant or unknown, it may be worse to reuse an existing Value. |
1464 | if (isa<SCEVConstant>(Val: S) || isa<SCEVUnknown>(Val: S)) |
1465 | return nullptr; |
1466 | |
1467 | for (Value *V : SE.getSCEVValues(S)) { |
1468 | Instruction *EntInst = dyn_cast<Instruction>(Val: V); |
1469 | if (!EntInst) |
1470 | continue; |
1471 | |
1472 | // Choose a Value from the set which dominates the InsertPt. |
1473 | // InsertPt should be inside the Value's parent loop so as not to break |
1474 | // the LCSSA form. |
1475 | assert(EntInst->getFunction() == InsertPt->getFunction()); |
1476 | if (S->getType() != V->getType() || !SE.DT.dominates(Def: EntInst, User: InsertPt) || |
1477 | !(SE.LI.getLoopFor(BB: EntInst->getParent()) == nullptr || |
1478 | SE.LI.getLoopFor(BB: EntInst->getParent())->contains(Inst: InsertPt))) |
1479 | continue; |
1480 | |
1481 | // Make sure reusing the instruction is poison-safe. |
1482 | if (SE.canReuseInstruction(S, I: EntInst, DropPoisonGeneratingInsts)) |
1483 | return V; |
1484 | DropPoisonGeneratingInsts.clear(); |
1485 | } |
1486 | return nullptr; |
1487 | } |
1488 | |
1489 | // The expansion of SCEV will either reuse a previous Value in ExprValueMap, |
1490 | // or expand the SCEV literally. Specifically, if the expansion is in LSRMode, |
1491 | // and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded |
1492 | // literally, to prevent LSR's transformed SCEV from being reverted. Otherwise, |
1493 | // the expansion will try to reuse Value from ExprValueMap, and only when it |
1494 | // fails, expand the SCEV literally. |
1495 | Value *SCEVExpander::expand(const SCEV *S) { |
1496 | // Compute an insertion point for this SCEV object. Hoist the instructions |
1497 | // as far out in the loop nest as possible. |
1498 | BasicBlock::iterator InsertPt = Builder.GetInsertPoint(); |
1499 | |
1500 | // We can move insertion point only if there is no div or rem operations |
1501 | // otherwise we are risky to move it over the check for zero denominator. |
1502 | auto SafeToHoist = [](const SCEV *S) { |
1503 | return !SCEVExprContains(Root: S, Pred: [](const SCEV *S) { |
1504 | if (const auto *D = dyn_cast<SCEVUDivExpr>(Val: S)) { |
1505 | if (const auto *SC = dyn_cast<SCEVConstant>(Val: D->getRHS())) |
1506 | // Division by non-zero constants can be hoisted. |
1507 | return SC->getValue()->isZero(); |
1508 | // All other divisions should not be moved as they may be |
1509 | // divisions by zero and should be kept within the |
1510 | // conditions of the surrounding loops that guard their |
1511 | // execution (see PR35406). |
1512 | return true; |
1513 | } |
1514 | return false; |
1515 | }); |
1516 | }; |
1517 | if (SafeToHoist(S)) { |
1518 | for (Loop *L = SE.LI.getLoopFor(BB: Builder.GetInsertBlock());; |
1519 | L = L->getParentLoop()) { |
1520 | if (SE.isLoopInvariant(S, L)) { |
1521 | if (!L) break; |
1522 | if (BasicBlock * = L->getLoopPreheader()) { |
1523 | InsertPt = Preheader->getTerminator()->getIterator(); |
1524 | } else { |
1525 | // LSR sets the insertion point for AddRec start/step values to the |
1526 | // block start to simplify value reuse, even though it's an invalid |
1527 | // position. SCEVExpander must correct for this in all cases. |
1528 | InsertPt = L->getHeader()->getFirstInsertionPt(); |
1529 | } |
1530 | } else { |
1531 | // If the SCEV is computable at this level, insert it into the header |
1532 | // after the PHIs (and after any other instructions that we've inserted |
1533 | // there) so that it is guaranteed to dominate any user inside the loop. |
1534 | if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(Ptr: L)) |
1535 | InsertPt = L->getHeader()->getFirstInsertionPt(); |
1536 | |
1537 | while (InsertPt != Builder.GetInsertPoint() && |
1538 | (isInsertedInstruction(I: &*InsertPt))) { |
1539 | InsertPt = std::next(x: InsertPt); |
1540 | } |
1541 | break; |
1542 | } |
1543 | } |
1544 | } |
1545 | |
1546 | // Check to see if we already expanded this here. |
1547 | auto I = InsertedExpressions.find(Val: std::make_pair(x&: S, y: &*InsertPt)); |
1548 | if (I != InsertedExpressions.end()) |
1549 | return I->second; |
1550 | |
1551 | SCEVInsertPointGuard Guard(Builder, this); |
1552 | Builder.SetInsertPoint(TheBB: InsertPt->getParent(), IP: InsertPt); |
1553 | |
1554 | // Expand the expression into instructions. |
1555 | SmallVector<Instruction *> DropPoisonGeneratingInsts; |
1556 | Value *V = FindValueInExprValueMap(S, InsertPt: &*InsertPt, DropPoisonGeneratingInsts); |
1557 | if (!V) { |
1558 | V = visit(S); |
1559 | V = fixupLCSSAFormFor(V); |
1560 | } else { |
1561 | for (Instruction *I : DropPoisonGeneratingInsts) { |
1562 | rememberFlags(I); |
1563 | I->dropPoisonGeneratingAnnotations(); |
1564 | // See if we can re-infer from first principles any of the flags we just |
1565 | // dropped. |
1566 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: I)) |
1567 | if (auto Flags = SE.getStrengthenedNoWrapFlagsFromBinOp(OBO)) { |
1568 | auto *BO = cast<BinaryOperator>(Val: I); |
1569 | BO->setHasNoUnsignedWrap( |
1570 | ScalarEvolution::maskFlags(Flags: *Flags, Mask: SCEV::FlagNUW) == SCEV::FlagNUW); |
1571 | BO->setHasNoSignedWrap( |
1572 | ScalarEvolution::maskFlags(Flags: *Flags, Mask: SCEV::FlagNSW) == SCEV::FlagNSW); |
1573 | } |
1574 | if (auto *NNI = dyn_cast<PossiblyNonNegInst>(Val: I)) { |
1575 | auto *Src = NNI->getOperand(i_nocapture: 0); |
1576 | if (isImpliedByDomCondition(Pred: ICmpInst::ICMP_SGE, LHS: Src, |
1577 | RHS: Constant::getNullValue(Ty: Src->getType()), ContextI: I, |
1578 | DL).value_or(u: false)) |
1579 | NNI->setNonNeg(true); |
1580 | } |
1581 | } |
1582 | } |
1583 | // Remember the expanded value for this SCEV at this location. |
1584 | // |
1585 | // This is independent of PostIncLoops. The mapped value simply materializes |
1586 | // the expression at this insertion point. If the mapped value happened to be |
1587 | // a postinc expansion, it could be reused by a non-postinc user, but only if |
1588 | // its insertion point was already at the head of the loop. |
1589 | InsertedExpressions[std::make_pair(x&: S, y: &*InsertPt)] = V; |
1590 | return V; |
1591 | } |
1592 | |
1593 | void SCEVExpander::rememberInstruction(Value *I) { |
1594 | auto DoInsert = [this](Value *V) { |
1595 | if (!PostIncLoops.empty()) |
1596 | InsertedPostIncValues.insert(V); |
1597 | else |
1598 | InsertedValues.insert(V); |
1599 | }; |
1600 | DoInsert(I); |
1601 | } |
1602 | |
1603 | void SCEVExpander::rememberFlags(Instruction *I) { |
1604 | // If we already have flags for the instruction, keep the existing ones. |
1605 | OrigFlags.try_emplace(Key: I, Args: PoisonFlags(I)); |
1606 | } |
1607 | |
1608 | void SCEVExpander::replaceCongruentIVInc( |
1609 | PHINode *&Phi, PHINode *&OrigPhi, Loop *L, const DominatorTree *DT, |
1610 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) { |
1611 | BasicBlock *LatchBlock = L->getLoopLatch(); |
1612 | if (!LatchBlock) |
1613 | return; |
1614 | |
1615 | Instruction *OrigInc = |
1616 | dyn_cast<Instruction>(Val: OrigPhi->getIncomingValueForBlock(BB: LatchBlock)); |
1617 | Instruction *IsomorphicInc = |
1618 | dyn_cast<Instruction>(Val: Phi->getIncomingValueForBlock(BB: LatchBlock)); |
1619 | if (!OrigInc || !IsomorphicInc) |
1620 | return; |
1621 | |
1622 | // If this phi has the same width but is more canonical, replace the |
1623 | // original with it. As part of the "more canonical" determination, |
1624 | // respect a prior decision to use an IV chain. |
1625 | if (OrigPhi->getType() == Phi->getType()) { |
1626 | bool Chained = ChainedPhis.contains(V: Phi); |
1627 | if (!(Chained || isExpandedAddRecExprPHI(PN: OrigPhi, IncV: OrigInc, L)) && |
1628 | (Chained || isExpandedAddRecExprPHI(PN: Phi, IncV: IsomorphicInc, L))) { |
1629 | std::swap(a&: OrigPhi, b&: Phi); |
1630 | std::swap(a&: OrigInc, b&: IsomorphicInc); |
1631 | } |
1632 | } |
1633 | |
1634 | // Replacing the congruent phi is sufficient because acyclic |
1635 | // redundancy elimination, CSE/GVN, should handle the |
1636 | // rest. However, once SCEV proves that a phi is congruent, |
1637 | // it's often the head of an IV user cycle that is isomorphic |
1638 | // with the original phi. It's worth eagerly cleaning up the |
1639 | // common case of a single IV increment so that DeleteDeadPHIs |
1640 | // can remove cycles that had postinc uses. |
1641 | // Because we may potentially introduce a new use of OrigIV that didn't |
1642 | // exist before at this point, its poison flags need readjustment. |
1643 | const SCEV *TruncExpr = |
1644 | SE.getTruncateOrNoop(V: SE.getSCEV(V: OrigInc), Ty: IsomorphicInc->getType()); |
1645 | if (OrigInc == IsomorphicInc || TruncExpr != SE.getSCEV(V: IsomorphicInc) || |
1646 | !SE.LI.replacementPreservesLCSSAForm(From: IsomorphicInc, To: OrigInc)) |
1647 | return; |
1648 | |
1649 | bool BothHaveNUW = false; |
1650 | bool BothHaveNSW = false; |
1651 | auto *OBOIncV = dyn_cast<OverflowingBinaryOperator>(Val: OrigInc); |
1652 | auto *OBOIsomorphic = dyn_cast<OverflowingBinaryOperator>(Val: IsomorphicInc); |
1653 | if (OBOIncV && OBOIsomorphic) { |
1654 | BothHaveNUW = |
1655 | OBOIncV->hasNoUnsignedWrap() && OBOIsomorphic->hasNoUnsignedWrap(); |
1656 | BothHaveNSW = |
1657 | OBOIncV->hasNoSignedWrap() && OBOIsomorphic->hasNoSignedWrap(); |
1658 | } |
1659 | |
1660 | if (!hoistIVInc(IncV: OrigInc, InsertPos: IsomorphicInc, |
1661 | /*RecomputePoisonFlags*/ true)) |
1662 | return; |
1663 | |
1664 | // We are replacing with a wider increment. If both OrigInc and IsomorphicInc |
1665 | // are NUW/NSW, then we can preserve them on the wider increment; the narrower |
1666 | // IsomorphicInc would wrap before the wider OrigInc, so the replacement won't |
1667 | // make IsomorphicInc's uses more poisonous. |
1668 | assert(OrigInc->getType()->getScalarSizeInBits() >= |
1669 | IsomorphicInc->getType()->getScalarSizeInBits() && |
1670 | "Should only replace an increment with a wider one." ); |
1671 | if (BothHaveNUW || BothHaveNSW) { |
1672 | OrigInc->setHasNoUnsignedWrap(OBOIncV->hasNoUnsignedWrap() || BothHaveNUW); |
1673 | OrigInc->setHasNoSignedWrap(OBOIncV->hasNoSignedWrap() || BothHaveNSW); |
1674 | } |
1675 | |
1676 | SCEV_DEBUG_WITH_TYPE(DebugType, |
1677 | dbgs() << "INDVARS: Eliminated congruent iv.inc: " |
1678 | << *IsomorphicInc << '\n'); |
1679 | Value *NewInc = OrigInc; |
1680 | if (OrigInc->getType() != IsomorphicInc->getType()) { |
1681 | BasicBlock::iterator IP; |
1682 | if (PHINode *PN = dyn_cast<PHINode>(Val: OrigInc)) |
1683 | IP = PN->getParent()->getFirstInsertionPt(); |
1684 | else |
1685 | IP = OrigInc->getNextNonDebugInstruction()->getIterator(); |
1686 | |
1687 | IRBuilder<> Builder(IP->getParent(), IP); |
1688 | Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); |
1689 | NewInc = |
1690 | Builder.CreateTruncOrBitCast(V: OrigInc, DestTy: IsomorphicInc->getType(), Name: IVName); |
1691 | } |
1692 | IsomorphicInc->replaceAllUsesWith(V: NewInc); |
1693 | DeadInsts.emplace_back(Args&: IsomorphicInc); |
1694 | } |
1695 | |
1696 | /// replaceCongruentIVs - Check for congruent phis in this loop header and |
1697 | /// replace them with their most canonical representative. Return the number of |
1698 | /// phis eliminated. |
1699 | /// |
1700 | /// This does not depend on any SCEVExpander state but should be used in |
1701 | /// the same context that SCEVExpander is used. |
1702 | unsigned |
1703 | SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, |
1704 | SmallVectorImpl<WeakTrackingVH> &DeadInsts, |
1705 | const TargetTransformInfo *TTI) { |
1706 | // Find integer phis in order of increasing width. |
1707 | SmallVector<PHINode *, 8> Phis( |
1708 | llvm::make_pointer_range(Range: L->getHeader()->phis())); |
1709 | |
1710 | if (TTI) |
1711 | // Use stable_sort to preserve order of equivalent PHIs, so the order |
1712 | // of the sorted Phis is the same from run to run on the same loop. |
1713 | llvm::stable_sort(Range&: Phis, C: [](Value *LHS, Value *RHS) { |
1714 | // Put pointers at the back and make sure pointer < pointer = false. |
1715 | if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) |
1716 | return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); |
1717 | return RHS->getType()->getPrimitiveSizeInBits().getFixedValue() < |
1718 | LHS->getType()->getPrimitiveSizeInBits().getFixedValue(); |
1719 | }); |
1720 | |
1721 | unsigned NumElim = 0; |
1722 | DenseMap<const SCEV *, PHINode *> ExprToIVMap; |
1723 | // Process phis from wide to narrow. Map wide phis to their truncation |
1724 | // so narrow phis can reuse them. |
1725 | for (PHINode *Phi : Phis) { |
1726 | auto SimplifyPHINode = [&](PHINode *PN) -> Value * { |
1727 | if (Value *V = simplifyInstruction(I: PN, Q: {DL, &SE.TLI, &SE.DT, &SE.AC})) |
1728 | return V; |
1729 | if (!SE.isSCEVable(Ty: PN->getType())) |
1730 | return nullptr; |
1731 | auto *Const = dyn_cast<SCEVConstant>(Val: SE.getSCEV(V: PN)); |
1732 | if (!Const) |
1733 | return nullptr; |
1734 | return Const->getValue(); |
1735 | }; |
1736 | |
1737 | // Fold constant phis. They may be congruent to other constant phis and |
1738 | // would confuse the logic below that expects proper IVs. |
1739 | if (Value *V = SimplifyPHINode(Phi)) { |
1740 | if (V->getType() != Phi->getType()) |
1741 | continue; |
1742 | SE.forgetValue(V: Phi); |
1743 | Phi->replaceAllUsesWith(V); |
1744 | DeadInsts.emplace_back(Args&: Phi); |
1745 | ++NumElim; |
1746 | SCEV_DEBUG_WITH_TYPE(DebugType, |
1747 | dbgs() << "INDVARS: Eliminated constant iv: " << *Phi |
1748 | << '\n'); |
1749 | continue; |
1750 | } |
1751 | |
1752 | if (!SE.isSCEVable(Ty: Phi->getType())) |
1753 | continue; |
1754 | |
1755 | PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(V: Phi)]; |
1756 | if (!OrigPhiRef) { |
1757 | OrigPhiRef = Phi; |
1758 | if (Phi->getType()->isIntegerTy() && TTI && |
1759 | TTI->isTruncateFree(Ty1: Phi->getType(), Ty2: Phis.back()->getType())) { |
1760 | // Make sure we only rewrite using simple induction variables; |
1761 | // otherwise, we can make the trip count of a loop unanalyzable |
1762 | // to SCEV. |
1763 | const SCEV *PhiExpr = SE.getSCEV(V: Phi); |
1764 | if (isa<SCEVAddRecExpr>(Val: PhiExpr)) { |
1765 | // This phi can be freely truncated to the narrowest phi type. Map the |
1766 | // truncated expression to it so it will be reused for narrow types. |
1767 | const SCEV *TruncExpr = |
1768 | SE.getTruncateExpr(Op: PhiExpr, Ty: Phis.back()->getType()); |
1769 | ExprToIVMap[TruncExpr] = Phi; |
1770 | } |
1771 | } |
1772 | continue; |
1773 | } |
1774 | |
1775 | // Replacing a pointer phi with an integer phi or vice-versa doesn't make |
1776 | // sense. |
1777 | if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) |
1778 | continue; |
1779 | |
1780 | replaceCongruentIVInc(Phi, OrigPhi&: OrigPhiRef, L, DT, DeadInsts); |
1781 | SCEV_DEBUG_WITH_TYPE(DebugType, |
1782 | dbgs() << "INDVARS: Eliminated congruent iv: " << *Phi |
1783 | << '\n'); |
1784 | SCEV_DEBUG_WITH_TYPE( |
1785 | DebugType, dbgs() << "INDVARS: Original iv: " << *OrigPhiRef << '\n'); |
1786 | ++NumElim; |
1787 | Value *NewIV = OrigPhiRef; |
1788 | if (OrigPhiRef->getType() != Phi->getType()) { |
1789 | IRBuilder<> Builder(L->getHeader(), |
1790 | L->getHeader()->getFirstInsertionPt()); |
1791 | Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); |
1792 | NewIV = Builder.CreateTruncOrBitCast(V: OrigPhiRef, DestTy: Phi->getType(), Name: IVName); |
1793 | } |
1794 | Phi->replaceAllUsesWith(V: NewIV); |
1795 | DeadInsts.emplace_back(Args&: Phi); |
1796 | } |
1797 | return NumElim; |
1798 | } |
1799 | |
1800 | bool SCEVExpander::hasRelatedExistingExpansion(const SCEV *S, |
1801 | const Instruction *At, |
1802 | Loop *L) { |
1803 | using namespace llvm::PatternMatch; |
1804 | |
1805 | SmallVector<BasicBlock *, 4> ExitingBlocks; |
1806 | L->getExitingBlocks(ExitingBlocks); |
1807 | |
1808 | // Look for suitable value in simple conditions at the loop exits. |
1809 | for (BasicBlock *BB : ExitingBlocks) { |
1810 | CmpPredicate Pred; |
1811 | Instruction *LHS, *RHS; |
1812 | |
1813 | if (!match(V: BB->getTerminator(), |
1814 | P: m_Br(C: m_ICmp(Pred, L: m_Instruction(I&: LHS), R: m_Instruction(I&: RHS)), |
1815 | T: m_BasicBlock(), F: m_BasicBlock()))) |
1816 | continue; |
1817 | |
1818 | if (SE.getSCEV(V: LHS) == S && SE.DT.dominates(Def: LHS, User: At)) |
1819 | return true; |
1820 | |
1821 | if (SE.getSCEV(V: RHS) == S && SE.DT.dominates(Def: RHS, User: At)) |
1822 | return true; |
1823 | } |
1824 | |
1825 | // Use expand's logic which is used for reusing a previous Value in |
1826 | // ExprValueMap. Note that we don't currently model the cost of |
1827 | // needing to drop poison generating flags on the instruction if we |
1828 | // want to reuse it. We effectively assume that has zero cost. |
1829 | SmallVector<Instruction *> DropPoisonGeneratingInsts; |
1830 | return FindValueInExprValueMap(S, InsertPt: At, DropPoisonGeneratingInsts) != nullptr; |
1831 | } |
1832 | |
1833 | template<typename T> static InstructionCost costAndCollectOperands( |
1834 | const SCEVOperand &WorkItem, const TargetTransformInfo &TTI, |
1835 | TargetTransformInfo::TargetCostKind CostKind, |
1836 | SmallVectorImpl<SCEVOperand> &Worklist) { |
1837 | |
1838 | const T *S = cast<T>(WorkItem.S); |
1839 | InstructionCost Cost = 0; |
1840 | // Object to help map SCEV operands to expanded IR instructions. |
1841 | struct OperationIndices { |
1842 | OperationIndices(unsigned Opc, size_t min, size_t max) : |
1843 | Opcode(Opc), MinIdx(min), MaxIdx(max) { } |
1844 | unsigned Opcode; |
1845 | size_t MinIdx; |
1846 | size_t MaxIdx; |
1847 | }; |
1848 | |
1849 | // Collect the operations of all the instructions that will be needed to |
1850 | // expand the SCEVExpr. This is so that when we come to cost the operands, |
1851 | // we know what the generated user(s) will be. |
1852 | SmallVector<OperationIndices, 2> Operations; |
1853 | |
1854 | auto CastCost = [&](unsigned Opcode) -> InstructionCost { |
1855 | Operations.emplace_back(Opcode, 0, 0); |
1856 | return TTI.getCastInstrCost(Opcode, Dst: S->getType(), |
1857 | Src: S->getOperand(0)->getType(), |
1858 | CCH: TTI::CastContextHint::None, CostKind); |
1859 | }; |
1860 | |
1861 | auto ArithCost = [&](unsigned Opcode, unsigned NumRequired, |
1862 | unsigned MinIdx = 0, |
1863 | unsigned MaxIdx = 1) -> InstructionCost { |
1864 | Operations.emplace_back(Opcode, MinIdx, MaxIdx); |
1865 | return NumRequired * |
1866 | TTI.getArithmeticInstrCost(Opcode, Ty: S->getType(), CostKind); |
1867 | }; |
1868 | |
1869 | auto CmpSelCost = [&](unsigned Opcode, unsigned NumRequired, unsigned MinIdx, |
1870 | unsigned MaxIdx) -> InstructionCost { |
1871 | Operations.emplace_back(Opcode, MinIdx, MaxIdx); |
1872 | Type *OpType = S->getType(); |
1873 | return NumRequired * TTI.getCmpSelInstrCost( |
1874 | Opcode, ValTy: OpType, CondTy: CmpInst::makeCmpResultType(opnd_type: OpType), |
1875 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
1876 | }; |
1877 | |
1878 | switch (S->getSCEVType()) { |
1879 | case scCouldNotCompute: |
1880 | llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!" ); |
1881 | case scUnknown: |
1882 | case scConstant: |
1883 | case scVScale: |
1884 | return 0; |
1885 | case scPtrToInt: |
1886 | Cost = CastCost(Instruction::PtrToInt); |
1887 | break; |
1888 | case scTruncate: |
1889 | Cost = CastCost(Instruction::Trunc); |
1890 | break; |
1891 | case scZeroExtend: |
1892 | Cost = CastCost(Instruction::ZExt); |
1893 | break; |
1894 | case scSignExtend: |
1895 | Cost = CastCost(Instruction::SExt); |
1896 | break; |
1897 | case scUDivExpr: { |
1898 | unsigned Opcode = Instruction::UDiv; |
1899 | if (auto *SC = dyn_cast<SCEVConstant>(S->getOperand(1))) |
1900 | if (SC->getAPInt().isPowerOf2()) |
1901 | Opcode = Instruction::LShr; |
1902 | Cost = ArithCost(Opcode, 1); |
1903 | break; |
1904 | } |
1905 | case scAddExpr: |
1906 | Cost = ArithCost(Instruction::Add, S->getNumOperands() - 1); |
1907 | break; |
1908 | case scMulExpr: |
1909 | // TODO: this is a very pessimistic cost modelling for Mul, |
1910 | // because of Bin Pow algorithm actually used by the expander, |
1911 | // see SCEVExpander::visitMulExpr(), ExpandOpBinPowN(). |
1912 | Cost = ArithCost(Instruction::Mul, S->getNumOperands() - 1); |
1913 | break; |
1914 | case scSMaxExpr: |
1915 | case scUMaxExpr: |
1916 | case scSMinExpr: |
1917 | case scUMinExpr: |
1918 | case scSequentialUMinExpr: { |
1919 | // FIXME: should this ask the cost for Intrinsic's? |
1920 | // The reduction tree. |
1921 | Cost += CmpSelCost(Instruction::ICmp, S->getNumOperands() - 1, 0, 1); |
1922 | Cost += CmpSelCost(Instruction::Select, S->getNumOperands() - 1, 0, 2); |
1923 | switch (S->getSCEVType()) { |
1924 | case scSequentialUMinExpr: { |
1925 | // The safety net against poison. |
1926 | // FIXME: this is broken. |
1927 | Cost += CmpSelCost(Instruction::ICmp, S->getNumOperands() - 1, 0, 0); |
1928 | Cost += ArithCost(Instruction::Or, |
1929 | S->getNumOperands() > 2 ? S->getNumOperands() - 2 : 0); |
1930 | Cost += CmpSelCost(Instruction::Select, 1, 0, 1); |
1931 | break; |
1932 | } |
1933 | default: |
1934 | assert(!isa<SCEVSequentialMinMaxExpr>(S) && |
1935 | "Unhandled SCEV expression type?" ); |
1936 | break; |
1937 | } |
1938 | break; |
1939 | } |
1940 | case scAddRecExpr: { |
1941 | // Addrec expands to a phi and add per recurrence. |
1942 | unsigned NumRecurrences = S->getNumOperands() - 1; |
1943 | Cost += TTI.getCFInstrCost(Opcode: Instruction::PHI, CostKind) * NumRecurrences; |
1944 | Cost += |
1945 | TTI.getArithmeticInstrCost(Opcode: Instruction::Add, Ty: S->getType(), CostKind) * |
1946 | NumRecurrences; |
1947 | // AR start is used in phi. |
1948 | Worklist.emplace_back(Instruction::PHI, 0, S->getOperand(0)); |
1949 | // Other operands are used in add. |
1950 | for (const SCEV *Op : S->operands().drop_front()) |
1951 | Worklist.emplace_back(Args: Instruction::Add, Args: 1, Args&: Op); |
1952 | break; |
1953 | } |
1954 | } |
1955 | |
1956 | for (auto &CostOp : Operations) { |
1957 | for (auto SCEVOp : enumerate(S->operands())) { |
1958 | // Clamp the index to account for multiple IR operations being chained. |
1959 | size_t MinIdx = std::max(SCEVOp.index(), CostOp.MinIdx); |
1960 | size_t OpIdx = std::min(MinIdx, CostOp.MaxIdx); |
1961 | Worklist.emplace_back(CostOp.Opcode, OpIdx, SCEVOp.value()); |
1962 | } |
1963 | } |
1964 | return Cost; |
1965 | } |
1966 | |
1967 | bool SCEVExpander::isHighCostExpansionHelper( |
1968 | const SCEVOperand &WorkItem, Loop *L, const Instruction &At, |
1969 | InstructionCost &Cost, unsigned Budget, const TargetTransformInfo &TTI, |
1970 | SmallPtrSetImpl<const SCEV *> &Processed, |
1971 | SmallVectorImpl<SCEVOperand> &Worklist) { |
1972 | if (Cost > Budget) |
1973 | return true; // Already run out of budget, give up. |
1974 | |
1975 | const SCEV *S = WorkItem.S; |
1976 | // Was the cost of expansion of this expression already accounted for? |
1977 | if (!isa<SCEVConstant>(Val: S) && !Processed.insert(Ptr: S).second) |
1978 | return false; // We have already accounted for this expression. |
1979 | |
1980 | // If we can find an existing value for this scev available at the point "At" |
1981 | // then consider the expression cheap. |
1982 | if (hasRelatedExistingExpansion(S, At: &At, L)) |
1983 | return false; // Consider the expression to be free. |
1984 | |
1985 | TargetTransformInfo::TargetCostKind CostKind = |
1986 | L->getHeader()->getParent()->hasMinSize() |
1987 | ? TargetTransformInfo::TCK_CodeSize |
1988 | : TargetTransformInfo::TCK_RecipThroughput; |
1989 | |
1990 | switch (S->getSCEVType()) { |
1991 | case scCouldNotCompute: |
1992 | llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!" ); |
1993 | case scUnknown: |
1994 | case scVScale: |
1995 | // Assume to be zero-cost. |
1996 | return false; |
1997 | case scConstant: { |
1998 | // Only evalulate the costs of constants when optimizing for size. |
1999 | if (CostKind != TargetTransformInfo::TCK_CodeSize) |
2000 | return false; |
2001 | const APInt &Imm = cast<SCEVConstant>(Val: S)->getAPInt(); |
2002 | Type *Ty = S->getType(); |
2003 | Cost += TTI.getIntImmCostInst( |
2004 | Opc: WorkItem.ParentOpcode, Idx: WorkItem.OperandIdx, Imm, Ty, CostKind); |
2005 | return Cost > Budget; |
2006 | } |
2007 | case scTruncate: |
2008 | case scPtrToInt: |
2009 | case scZeroExtend: |
2010 | case scSignExtend: { |
2011 | Cost += |
2012 | costAndCollectOperands<SCEVCastExpr>(WorkItem, TTI, CostKind, Worklist); |
2013 | return false; // Will answer upon next entry into this function. |
2014 | } |
2015 | case scUDivExpr: { |
2016 | // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or |
2017 | // HowManyLessThans produced to compute a precise expression, rather than a |
2018 | // UDiv from the user's code. If we can't find a UDiv in the code with some |
2019 | // simple searching, we need to account for it's cost. |
2020 | |
2021 | // At the beginning of this function we already tried to find existing |
2022 | // value for plain 'S'. Now try to lookup 'S + 1' since it is common |
2023 | // pattern involving division. This is just a simple search heuristic. |
2024 | if (hasRelatedExistingExpansion( |
2025 | S: SE.getAddExpr(LHS: S, RHS: SE.getConstant(Ty: S->getType(), V: 1)), At: &At, L)) |
2026 | return false; // Consider it to be free. |
2027 | |
2028 | Cost += |
2029 | costAndCollectOperands<SCEVUDivExpr>(WorkItem, TTI, CostKind, Worklist); |
2030 | return false; // Will answer upon next entry into this function. |
2031 | } |
2032 | case scAddExpr: |
2033 | case scMulExpr: |
2034 | case scUMaxExpr: |
2035 | case scSMaxExpr: |
2036 | case scUMinExpr: |
2037 | case scSMinExpr: |
2038 | case scSequentialUMinExpr: { |
2039 | assert(cast<SCEVNAryExpr>(S)->getNumOperands() > 1 && |
2040 | "Nary expr should have more than 1 operand." ); |
2041 | // The simple nary expr will require one less op (or pair of ops) |
2042 | // than the number of it's terms. |
2043 | Cost += |
2044 | costAndCollectOperands<SCEVNAryExpr>(WorkItem, TTI, CostKind, Worklist); |
2045 | return Cost > Budget; |
2046 | } |
2047 | case scAddRecExpr: { |
2048 | assert(cast<SCEVAddRecExpr>(S)->getNumOperands() >= 2 && |
2049 | "Polynomial should be at least linear" ); |
2050 | Cost += costAndCollectOperands<SCEVAddRecExpr>( |
2051 | WorkItem, TTI, CostKind, Worklist); |
2052 | return Cost > Budget; |
2053 | } |
2054 | } |
2055 | llvm_unreachable("Unknown SCEV kind!" ); |
2056 | } |
2057 | |
2058 | Value *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred, |
2059 | Instruction *IP) { |
2060 | assert(IP); |
2061 | switch (Pred->getKind()) { |
2062 | case SCEVPredicate::P_Union: |
2063 | return expandUnionPredicate(Pred: cast<SCEVUnionPredicate>(Val: Pred), Loc: IP); |
2064 | case SCEVPredicate::P_Compare: |
2065 | return expandComparePredicate(Pred: cast<SCEVComparePredicate>(Val: Pred), Loc: IP); |
2066 | case SCEVPredicate::P_Wrap: { |
2067 | auto *AddRecPred = cast<SCEVWrapPredicate>(Val: Pred); |
2068 | return expandWrapPredicate(P: AddRecPred, Loc: IP); |
2069 | } |
2070 | } |
2071 | llvm_unreachable("Unknown SCEV predicate type" ); |
2072 | } |
2073 | |
2074 | Value *SCEVExpander::expandComparePredicate(const SCEVComparePredicate *Pred, |
2075 | Instruction *IP) { |
2076 | Value *Expr0 = expand(S: Pred->getLHS(), I: IP); |
2077 | Value *Expr1 = expand(S: Pred->getRHS(), I: IP); |
2078 | |
2079 | Builder.SetInsertPoint(IP); |
2080 | auto InvPred = ICmpInst::getInversePredicate(pred: Pred->getPredicate()); |
2081 | auto *I = Builder.CreateICmp(P: InvPred, LHS: Expr0, RHS: Expr1, Name: "ident.check" ); |
2082 | return I; |
2083 | } |
2084 | |
2085 | Value *SCEVExpander::generateOverflowCheck(const SCEVAddRecExpr *AR, |
2086 | Instruction *Loc, bool Signed) { |
2087 | assert(AR->isAffine() && "Cannot generate RT check for " |
2088 | "non-affine expression" ); |
2089 | |
2090 | // FIXME: It is highly suspicious that we're ignoring the predicates here. |
2091 | SmallVector<const SCEVPredicate *, 4> Pred; |
2092 | const SCEV *ExitCount = |
2093 | SE.getPredicatedSymbolicMaxBackedgeTakenCount(L: AR->getLoop(), Predicates&: Pred); |
2094 | |
2095 | assert(!isa<SCEVCouldNotCompute>(ExitCount) && "Invalid loop count" ); |
2096 | |
2097 | const SCEV *Step = AR->getStepRecurrence(SE); |
2098 | const SCEV *Start = AR->getStart(); |
2099 | |
2100 | Type *ARTy = AR->getType(); |
2101 | unsigned SrcBits = SE.getTypeSizeInBits(Ty: ExitCount->getType()); |
2102 | unsigned DstBits = SE.getTypeSizeInBits(Ty: ARTy); |
2103 | |
2104 | // The expression {Start,+,Step} has nusw/nssw if |
2105 | // Step < 0, Start - |Step| * Backedge <= Start |
2106 | // Step >= 0, Start + |Step| * Backedge > Start |
2107 | // and |Step| * Backedge doesn't unsigned overflow. |
2108 | |
2109 | Builder.SetInsertPoint(Loc); |
2110 | Value *TripCountVal = expand(S: ExitCount, I: Loc); |
2111 | |
2112 | IntegerType *Ty = |
2113 | IntegerType::get(C&: Loc->getContext(), NumBits: SE.getTypeSizeInBits(Ty: ARTy)); |
2114 | |
2115 | Value *StepValue = expand(S: Step, I: Loc); |
2116 | Value *NegStepValue = expand(S: SE.getNegativeSCEV(V: Step), I: Loc); |
2117 | Value *StartValue = expand(S: Start, I: Loc); |
2118 | |
2119 | ConstantInt *Zero = |
2120 | ConstantInt::get(Context&: Loc->getContext(), V: APInt::getZero(numBits: DstBits)); |
2121 | |
2122 | Builder.SetInsertPoint(Loc); |
2123 | // Compute |Step| |
2124 | Value *StepCompare = Builder.CreateICmp(P: ICmpInst::ICMP_SLT, LHS: StepValue, RHS: Zero); |
2125 | Value *AbsStep = Builder.CreateSelect(C: StepCompare, True: NegStepValue, False: StepValue); |
2126 | |
2127 | // Compute |Step| * Backedge |
2128 | // Compute: |
2129 | // 1. Start + |Step| * Backedge < Start |
2130 | // 2. Start - |Step| * Backedge > Start |
2131 | // |
2132 | // And select either 1. or 2. depending on whether step is positive or |
2133 | // negative. If Step is known to be positive or negative, only create |
2134 | // either 1. or 2. |
2135 | auto ComputeEndCheck = [&]() -> Value * { |
2136 | // Checking <u 0 is always false. |
2137 | if (!Signed && Start->isZero() && SE.isKnownPositive(S: Step)) |
2138 | return ConstantInt::getFalse(Context&: Loc->getContext()); |
2139 | |
2140 | // Get the backedge taken count and truncate or extended to the AR type. |
2141 | Value *TruncTripCount = Builder.CreateZExtOrTrunc(V: TripCountVal, DestTy: Ty); |
2142 | |
2143 | Value *MulV, *OfMul; |
2144 | if (Step->isOne()) { |
2145 | // Special-case Step of one. Potentially-costly `umul_with_overflow` isn't |
2146 | // needed, there is never an overflow, so to avoid artificially inflating |
2147 | // the cost of the check, directly emit the optimized IR. |
2148 | MulV = TruncTripCount; |
2149 | OfMul = ConstantInt::getFalse(Context&: MulV->getContext()); |
2150 | } else { |
2151 | CallInst *Mul = Builder.CreateIntrinsic(ID: Intrinsic::umul_with_overflow, Types: Ty, |
2152 | Args: {AbsStep, TruncTripCount}, |
2153 | /*FMFSource=*/nullptr, Name: "mul" ); |
2154 | MulV = Builder.CreateExtractValue(Agg: Mul, Idxs: 0, Name: "mul.result" ); |
2155 | OfMul = Builder.CreateExtractValue(Agg: Mul, Idxs: 1, Name: "mul.overflow" ); |
2156 | } |
2157 | |
2158 | Value *Add = nullptr, *Sub = nullptr; |
2159 | bool NeedPosCheck = !SE.isKnownNegative(S: Step); |
2160 | bool NeedNegCheck = !SE.isKnownPositive(S: Step); |
2161 | |
2162 | if (isa<PointerType>(Val: ARTy)) { |
2163 | Value *NegMulV = Builder.CreateNeg(V: MulV); |
2164 | if (NeedPosCheck) |
2165 | Add = Builder.CreatePtrAdd(Ptr: StartValue, Offset: MulV); |
2166 | if (NeedNegCheck) |
2167 | Sub = Builder.CreatePtrAdd(Ptr: StartValue, Offset: NegMulV); |
2168 | } else { |
2169 | if (NeedPosCheck) |
2170 | Add = Builder.CreateAdd(LHS: StartValue, RHS: MulV); |
2171 | if (NeedNegCheck) |
2172 | Sub = Builder.CreateSub(LHS: StartValue, RHS: MulV); |
2173 | } |
2174 | |
2175 | Value *EndCompareLT = nullptr; |
2176 | Value *EndCompareGT = nullptr; |
2177 | Value *EndCheck = nullptr; |
2178 | if (NeedPosCheck) |
2179 | EndCheck = EndCompareLT = Builder.CreateICmp( |
2180 | P: Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, LHS: Add, RHS: StartValue); |
2181 | if (NeedNegCheck) |
2182 | EndCheck = EndCompareGT = Builder.CreateICmp( |
2183 | P: Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, LHS: Sub, RHS: StartValue); |
2184 | if (NeedPosCheck && NeedNegCheck) { |
2185 | // Select the answer based on the sign of Step. |
2186 | EndCheck = Builder.CreateSelect(C: StepCompare, True: EndCompareGT, False: EndCompareLT); |
2187 | } |
2188 | return Builder.CreateOr(LHS: EndCheck, RHS: OfMul); |
2189 | }; |
2190 | Value *EndCheck = ComputeEndCheck(); |
2191 | |
2192 | // If the backedge taken count type is larger than the AR type, |
2193 | // check that we don't drop any bits by truncating it. If we are |
2194 | // dropping bits, then we have overflow (unless the step is zero). |
2195 | if (SrcBits > DstBits) { |
2196 | auto MaxVal = APInt::getMaxValue(numBits: DstBits).zext(width: SrcBits); |
2197 | auto *BackedgeCheck = |
2198 | Builder.CreateICmp(P: ICmpInst::ICMP_UGT, LHS: TripCountVal, |
2199 | RHS: ConstantInt::get(Context&: Loc->getContext(), V: MaxVal)); |
2200 | BackedgeCheck = Builder.CreateAnd( |
2201 | LHS: BackedgeCheck, RHS: Builder.CreateICmp(P: ICmpInst::ICMP_NE, LHS: StepValue, RHS: Zero)); |
2202 | |
2203 | EndCheck = Builder.CreateOr(LHS: EndCheck, RHS: BackedgeCheck); |
2204 | } |
2205 | |
2206 | return EndCheck; |
2207 | } |
2208 | |
2209 | Value *SCEVExpander::expandWrapPredicate(const SCEVWrapPredicate *Pred, |
2210 | Instruction *IP) { |
2211 | const auto *A = cast<SCEVAddRecExpr>(Val: Pred->getExpr()); |
2212 | Value *NSSWCheck = nullptr, *NUSWCheck = nullptr; |
2213 | |
2214 | // Add a check for NUSW |
2215 | if (Pred->getFlags() & SCEVWrapPredicate::IncrementNUSW) |
2216 | NUSWCheck = generateOverflowCheck(AR: A, Loc: IP, Signed: false); |
2217 | |
2218 | // Add a check for NSSW |
2219 | if (Pred->getFlags() & SCEVWrapPredicate::IncrementNSSW) |
2220 | NSSWCheck = generateOverflowCheck(AR: A, Loc: IP, Signed: true); |
2221 | |
2222 | if (NUSWCheck && NSSWCheck) |
2223 | return Builder.CreateOr(LHS: NUSWCheck, RHS: NSSWCheck); |
2224 | |
2225 | if (NUSWCheck) |
2226 | return NUSWCheck; |
2227 | |
2228 | if (NSSWCheck) |
2229 | return NSSWCheck; |
2230 | |
2231 | return ConstantInt::getFalse(Context&: IP->getContext()); |
2232 | } |
2233 | |
2234 | Value *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union, |
2235 | Instruction *IP) { |
2236 | // Loop over all checks in this set. |
2237 | SmallVector<Value *> Checks; |
2238 | for (const auto *Pred : Union->getPredicates()) { |
2239 | Checks.push_back(Elt: expandCodeForPredicate(Pred, IP)); |
2240 | Builder.SetInsertPoint(IP); |
2241 | } |
2242 | |
2243 | if (Checks.empty()) |
2244 | return ConstantInt::getFalse(Context&: IP->getContext()); |
2245 | return Builder.CreateOr(Ops: Checks); |
2246 | } |
2247 | |
2248 | Value *SCEVExpander::fixupLCSSAFormFor(Value *V) { |
2249 | auto *DefI = dyn_cast<Instruction>(Val: V); |
2250 | if (!PreserveLCSSA || !DefI) |
2251 | return V; |
2252 | |
2253 | BasicBlock::iterator InsertPt = Builder.GetInsertPoint(); |
2254 | Loop *DefLoop = SE.LI.getLoopFor(BB: DefI->getParent()); |
2255 | Loop *UseLoop = SE.LI.getLoopFor(BB: InsertPt->getParent()); |
2256 | if (!DefLoop || UseLoop == DefLoop || DefLoop->contains(L: UseLoop)) |
2257 | return V; |
2258 | |
2259 | // Create a temporary instruction to at the current insertion point, so we |
2260 | // can hand it off to the helper to create LCSSA PHIs if required for the |
2261 | // new use. |
2262 | // FIXME: Ideally formLCSSAForInstructions (used in fixupLCSSAFormFor) |
2263 | // would accept a insertion point and return an LCSSA phi for that |
2264 | // insertion point, so there is no need to insert & remove the temporary |
2265 | // instruction. |
2266 | Type *ToTy; |
2267 | if (DefI->getType()->isIntegerTy()) |
2268 | ToTy = PointerType::get(C&: DefI->getContext(), AddressSpace: 0); |
2269 | else |
2270 | ToTy = Type::getInt32Ty(C&: DefI->getContext()); |
2271 | Instruction *User = |
2272 | CastInst::CreateBitOrPointerCast(S: DefI, Ty: ToTy, Name: "tmp.lcssa.user" , InsertBefore: InsertPt); |
2273 | auto RemoveUserOnExit = |
2274 | make_scope_exit(F: [User]() { User->eraseFromParent(); }); |
2275 | |
2276 | SmallVector<Instruction *, 1> ToUpdate; |
2277 | ToUpdate.push_back(Elt: DefI); |
2278 | SmallVector<PHINode *, 16> PHIsToRemove; |
2279 | SmallVector<PHINode *, 16> InsertedPHIs; |
2280 | formLCSSAForInstructions(Worklist&: ToUpdate, DT: SE.DT, LI: SE.LI, SE: &SE, PHIsToRemove: &PHIsToRemove, |
2281 | InsertedPHIs: &InsertedPHIs); |
2282 | for (PHINode *PN : InsertedPHIs) |
2283 | rememberInstruction(I: PN); |
2284 | for (PHINode *PN : PHIsToRemove) { |
2285 | if (!PN->use_empty()) |
2286 | continue; |
2287 | InsertedValues.erase(V: PN); |
2288 | InsertedPostIncValues.erase(V: PN); |
2289 | PN->eraseFromParent(); |
2290 | } |
2291 | |
2292 | return User->getOperand(i: 0); |
2293 | } |
2294 | |
2295 | namespace { |
2296 | // Search for a SCEV subexpression that is not safe to expand. Any expression |
2297 | // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely |
2298 | // UDiv expressions. We don't know if the UDiv is derived from an IR divide |
2299 | // instruction, but the important thing is that we prove the denominator is |
2300 | // nonzero before expansion. |
2301 | // |
2302 | // IVUsers already checks that IV-derived expressions are safe. So this check is |
2303 | // only needed when the expression includes some subexpression that is not IV |
2304 | // derived. |
2305 | // |
2306 | // Currently, we only allow division by a value provably non-zero here. |
2307 | // |
2308 | // We cannot generally expand recurrences unless the step dominates the loop |
2309 | // header. The expander handles the special case of affine recurrences by |
2310 | // scaling the recurrence outside the loop, but this technique isn't generally |
2311 | // applicable. Expanding a nested recurrence outside a loop requires computing |
2312 | // binomial coefficients. This could be done, but the recurrence has to be in a |
2313 | // perfectly reduced form, which can't be guaranteed. |
2314 | struct SCEVFindUnsafe { |
2315 | ScalarEvolution &SE; |
2316 | bool CanonicalMode; |
2317 | bool IsUnsafe = false; |
2318 | |
2319 | SCEVFindUnsafe(ScalarEvolution &SE, bool CanonicalMode) |
2320 | : SE(SE), CanonicalMode(CanonicalMode) {} |
2321 | |
2322 | bool follow(const SCEV *S) { |
2323 | if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(Val: S)) { |
2324 | if (!SE.isKnownNonZero(S: D->getRHS())) { |
2325 | IsUnsafe = true; |
2326 | return false; |
2327 | } |
2328 | } |
2329 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) { |
2330 | // For non-affine addrecs or in non-canonical mode we need a preheader |
2331 | // to insert into. |
2332 | if (!AR->getLoop()->getLoopPreheader() && |
2333 | (!CanonicalMode || !AR->isAffine())) { |
2334 | IsUnsafe = true; |
2335 | return false; |
2336 | } |
2337 | } |
2338 | return true; |
2339 | } |
2340 | bool isDone() const { return IsUnsafe; } |
2341 | }; |
2342 | } // namespace |
2343 | |
2344 | bool SCEVExpander::isSafeToExpand(const SCEV *S) const { |
2345 | SCEVFindUnsafe Search(SE, CanonicalMode); |
2346 | visitAll(Root: S, Visitor&: Search); |
2347 | return !Search.IsUnsafe; |
2348 | } |
2349 | |
2350 | bool SCEVExpander::isSafeToExpandAt(const SCEV *S, |
2351 | const Instruction *InsertionPoint) const { |
2352 | if (!isSafeToExpand(S)) |
2353 | return false; |
2354 | // We have to prove that the expanded site of S dominates InsertionPoint. |
2355 | // This is easy when not in the same block, but hard when S is an instruction |
2356 | // to be expanded somewhere inside the same block as our insertion point. |
2357 | // What we really need here is something analogous to an OrderedBasicBlock, |
2358 | // but for the moment, we paper over the problem by handling two common and |
2359 | // cheap to check cases. |
2360 | if (SE.properlyDominates(S, BB: InsertionPoint->getParent())) |
2361 | return true; |
2362 | if (SE.dominates(S, BB: InsertionPoint->getParent())) { |
2363 | if (InsertionPoint->getParent()->getTerminator() == InsertionPoint) |
2364 | return true; |
2365 | if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Val: S)) |
2366 | if (llvm::is_contained(Range: InsertionPoint->operand_values(), Element: U->getValue())) |
2367 | return true; |
2368 | } |
2369 | return false; |
2370 | } |
2371 | |
2372 | void SCEVExpanderCleaner::cleanup() { |
2373 | // Result is used, nothing to remove. |
2374 | if (ResultUsed) |
2375 | return; |
2376 | |
2377 | // Restore original poison flags. |
2378 | for (auto [I, Flags] : Expander.OrigFlags) |
2379 | Flags.apply(I); |
2380 | |
2381 | auto InsertedInstructions = Expander.getAllInsertedInstructions(); |
2382 | #ifndef NDEBUG |
2383 | SmallPtrSet<Instruction *, 8> InsertedSet(llvm::from_range, |
2384 | InsertedInstructions); |
2385 | (void)InsertedSet; |
2386 | #endif |
2387 | // Remove sets with value handles. |
2388 | Expander.clear(); |
2389 | |
2390 | // Remove all inserted instructions. |
2391 | for (Instruction *I : reverse(C&: InsertedInstructions)) { |
2392 | #ifndef NDEBUG |
2393 | assert(all_of(I->users(), |
2394 | [&InsertedSet](Value *U) { |
2395 | return InsertedSet.contains(cast<Instruction>(U)); |
2396 | }) && |
2397 | "removed instruction should only be used by instructions inserted " |
2398 | "during expansion" ); |
2399 | #endif |
2400 | assert(!I->getType()->isVoidTy() && |
2401 | "inserted instruction should have non-void types" ); |
2402 | I->replaceAllUsesWith(V: PoisonValue::get(T: I->getType())); |
2403 | I->eraseFromParent(); |
2404 | } |
2405 | } |
2406 | |