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