1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the implementation of the scalar evolution expander,
11 // which is used to generate the code corresponding to a given scalar evolution
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Analysis/ScalarEvolutionExpander.h"
17 #include "llvm/ADT/STLExtras.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/IR/DataLayout.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/Support/Debug.h"
30 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
31 /// reusing an existing cast if a suitable one exists, moving an existing
32 /// cast if a suitable one exists but isn't in the right place, or
33 /// creating a new one.
34 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
35 Instruction::CastOps Op,
36 BasicBlock::iterator IP) {
37 // This function must be called with the builder having a valid insertion
38 // point. It doesn't need to be the actual IP where the uses of the returned
39 // cast will be added, but it must dominate such IP.
40 // We use this precondition to produce a cast that will dominate all its
41 // uses. In particular, this is crucial for the case where the builder's
42 // insertion point *is* the point where we were asked to put the cast.
43 // Since we don't know the builder's insertion point is actually
44 // where the uses will be added (only that it dominates it), we are
45 // not allowed to move it.
46 BasicBlock::iterator BIP = Builder.GetInsertPoint();
48 Instruction *Ret = nullptr;
50 // Check to see if there is already a cast!
51 for (User *U : V->users())
52 if (U->getType() == Ty)
53 if (CastInst *CI = dyn_cast<CastInst>(U))
54 if (CI->getOpcode() == Op) {
55 // If the cast isn't where we want it, create a new cast at IP.
56 // Likewise, do not reuse a cast at BIP because it must dominate
57 // instructions that might be inserted before BIP.
58 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
59 // Create a new cast, and leave the old cast in place in case
60 // it is being used as an insert point. Clear its operand
61 // so that it doesn't hold anything live.
62 Ret = CastInst::Create(Op, V, Ty, "", IP);
64 CI->replaceAllUsesWith(Ret);
65 CI->setOperand(0, UndefValue::get(V->getType()));
74 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
76 // We assert at the end of the function since IP might point to an
77 // instruction with different dominance properties than a cast
78 // (an invoke for example) and not dominate BIP (but the cast does).
79 assert(SE.DT->dominates(Ret, BIP));
81 rememberInstruction(Ret);
85 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
86 /// which must be possible with a noop cast, doing what we can to share
88 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
89 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
90 assert((Op == Instruction::BitCast ||
91 Op == Instruction::PtrToInt ||
92 Op == Instruction::IntToPtr) &&
93 "InsertNoopCastOfTo cannot perform non-noop casts!");
94 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
95 "InsertNoopCastOfTo cannot change sizes!");
97 // Short-circuit unnecessary bitcasts.
98 if (Op == Instruction::BitCast) {
99 if (V->getType() == Ty)
101 if (CastInst *CI = dyn_cast<CastInst>(V)) {
102 if (CI->getOperand(0)->getType() == Ty)
103 return CI->getOperand(0);
106 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
107 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
108 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
109 if (CastInst *CI = dyn_cast<CastInst>(V))
110 if ((CI->getOpcode() == Instruction::PtrToInt ||
111 CI->getOpcode() == Instruction::IntToPtr) &&
112 SE.getTypeSizeInBits(CI->getType()) ==
113 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
114 return CI->getOperand(0);
115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
116 if ((CE->getOpcode() == Instruction::PtrToInt ||
117 CE->getOpcode() == Instruction::IntToPtr) &&
118 SE.getTypeSizeInBits(CE->getType()) ==
119 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
120 return CE->getOperand(0);
123 // Fold a cast of a constant.
124 if (Constant *C = dyn_cast<Constant>(V))
125 return ConstantExpr::getCast(Op, C, Ty);
127 // Cast the argument at the beginning of the entry block, after
128 // any bitcasts of other arguments.
129 if (Argument *A = dyn_cast<Argument>(V)) {
130 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
131 while ((isa<BitCastInst>(IP) &&
132 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
133 cast<BitCastInst>(IP)->getOperand(0) != A) ||
134 isa<DbgInfoIntrinsic>(IP) ||
135 isa<LandingPadInst>(IP))
137 return ReuseOrCreateCast(A, Ty, Op, IP);
140 // Cast the instruction immediately after the instruction.
141 Instruction *I = cast<Instruction>(V);
142 BasicBlock::iterator IP = I; ++IP;
143 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
144 IP = II->getNormalDest()->begin();
145 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
147 return ReuseOrCreateCast(I, Ty, Op, IP);
150 /// InsertBinop - Insert the specified binary operator, doing a small amount
151 /// of work to avoid inserting an obviously redundant operation.
152 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
153 Value *LHS, Value *RHS) {
154 // Fold a binop with constant operands.
155 if (Constant *CLHS = dyn_cast<Constant>(LHS))
156 if (Constant *CRHS = dyn_cast<Constant>(RHS))
157 return ConstantExpr::get(Opcode, CLHS, CRHS);
159 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
160 unsigned ScanLimit = 6;
161 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
162 // Scanning starts from the last instruction before the insertion point.
163 BasicBlock::iterator IP = Builder.GetInsertPoint();
164 if (IP != BlockBegin) {
166 for (; ScanLimit; --IP, --ScanLimit) {
167 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
169 if (isa<DbgInfoIntrinsic>(IP))
171 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
172 IP->getOperand(1) == RHS)
174 if (IP == BlockBegin) break;
178 // Save the original insertion point so we can restore it when we're done.
179 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
180 BuilderType::InsertPointGuard Guard(Builder);
182 // Move the insertion point out of as many loops as we can.
183 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
184 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
185 BasicBlock *Preheader = L->getLoopPreheader();
186 if (!Preheader) break;
188 // Ok, move up a level.
189 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
192 // If we haven't found this binop, insert it.
193 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
194 BO->setDebugLoc(Loc);
195 rememberInstruction(BO);
200 /// FactorOutConstant - Test if S is divisible by Factor, using signed
201 /// division. If so, update S with Factor divided out and return true.
202 /// S need not be evenly divisible if a reasonable remainder can be
204 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
205 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
206 /// check to see if the divide was folded.
207 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
208 const SCEV *Factor, ScalarEvolution &SE,
209 const DataLayout &DL) {
210 // Everything is divisible by one.
216 S = SE.getConstant(S->getType(), 1);
220 // For a Constant, check for a multiple of the given factor.
221 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
225 // Check for divisibility.
226 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
228 ConstantInt::get(SE.getContext(),
229 C->getValue()->getValue().sdiv(
230 FC->getValue()->getValue()));
231 // If the quotient is zero and the remainder is non-zero, reject
232 // the value at this scale. It will be considered for subsequent
235 const SCEV *Div = SE.getConstant(CI);
238 SE.getAddExpr(Remainder,
239 SE.getConstant(C->getValue()->getValue().srem(
240 FC->getValue()->getValue())));
246 // In a Mul, check if there is a constant operand which is a multiple
247 // of the given factor.
248 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
249 // Size is known, check if there is a constant operand which is a multiple
250 // of the given factor. If so, we can factor it.
251 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
252 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
253 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
254 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
255 NewMulOps[0] = SE.getConstant(
256 C->getValue()->getValue().sdiv(FC->getValue()->getValue()));
257 S = SE.getMulExpr(NewMulOps);
262 // In an AddRec, check if both start and step are divisible.
263 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
264 const SCEV *Step = A->getStepRecurrence(SE);
265 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
266 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
268 if (!StepRem->isZero())
270 const SCEV *Start = A->getStart();
271 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
273 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
274 A->getNoWrapFlags(SCEV::FlagNW));
281 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
282 /// is the number of SCEVAddRecExprs present, which are kept at the end of
285 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
287 ScalarEvolution &SE) {
288 unsigned NumAddRecs = 0;
289 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
291 // Group Ops into non-addrecs and addrecs.
292 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
293 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
294 // Let ScalarEvolution sort and simplify the non-addrecs list.
295 const SCEV *Sum = NoAddRecs.empty() ?
296 SE.getConstant(Ty, 0) :
297 SE.getAddExpr(NoAddRecs);
298 // If it returned an add, use the operands. Otherwise it simplified
299 // the sum into a single value, so just use that.
301 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
302 Ops.append(Add->op_begin(), Add->op_end());
303 else if (!Sum->isZero())
305 // Then append the addrecs.
306 Ops.append(AddRecs.begin(), AddRecs.end());
309 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
310 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
311 /// This helps expose more opportunities for folding parts of the expressions
312 /// into GEP indices.
314 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
316 ScalarEvolution &SE) {
318 SmallVector<const SCEV *, 8> AddRecs;
319 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
320 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
321 const SCEV *Start = A->getStart();
322 if (Start->isZero()) break;
323 const SCEV *Zero = SE.getConstant(Ty, 0);
324 AddRecs.push_back(SE.getAddRecExpr(Zero,
325 A->getStepRecurrence(SE),
327 A->getNoWrapFlags(SCEV::FlagNW)));
328 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
330 Ops.append(Add->op_begin(), Add->op_end());
331 e += Add->getNumOperands();
336 if (!AddRecs.empty()) {
337 // Add the addrecs onto the end of the list.
338 Ops.append(AddRecs.begin(), AddRecs.end());
339 // Resort the operand list, moving any constants to the front.
340 SimplifyAddOperands(Ops, Ty, SE);
344 /// expandAddToGEP - Expand an addition expression with a pointer type into
345 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
346 /// BasicAliasAnalysis and other passes analyze the result. See the rules
347 /// for getelementptr vs. inttoptr in
348 /// http://llvm.org/docs/LangRef.html#pointeraliasing
351 /// Design note: The correctness of using getelementptr here depends on
352 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
353 /// they may introduce pointer arithmetic which may not be safely converted
354 /// into getelementptr.
356 /// Design note: It might seem desirable for this function to be more
357 /// loop-aware. If some of the indices are loop-invariant while others
358 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
359 /// loop-invariant portions of the overall computation outside the loop.
360 /// However, there are a few reasons this is not done here. Hoisting simple
361 /// arithmetic is a low-level optimization that often isn't very
362 /// important until late in the optimization process. In fact, passes
363 /// like InstructionCombining will combine GEPs, even if it means
364 /// pushing loop-invariant computation down into loops, so even if the
365 /// GEPs were split here, the work would quickly be undone. The
366 /// LoopStrengthReduction pass, which is usually run quite late (and
367 /// after the last InstructionCombining pass), takes care of hoisting
368 /// loop-invariant portions of expressions, after considering what
369 /// can be folded using target addressing modes.
371 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
372 const SCEV *const *op_end,
376 Type *ElTy = PTy->getElementType();
377 SmallVector<Value *, 4> GepIndices;
378 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
379 bool AnyNonZeroIndices = false;
381 // Split AddRecs up into parts as either of the parts may be usable
382 // without the other.
383 SplitAddRecs(Ops, Ty, SE);
385 Type *IntPtrTy = DL.getIntPtrType(PTy);
387 // Descend down the pointer's type and attempt to convert the other
388 // operands into GEP indices, at each level. The first index in a GEP
389 // indexes into the array implied by the pointer operand; the rest of
390 // the indices index into the element or field type selected by the
393 // If the scale size is not 0, attempt to factor out a scale for
395 SmallVector<const SCEV *, 8> ScaledOps;
396 if (ElTy->isSized()) {
397 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
398 if (!ElSize->isZero()) {
399 SmallVector<const SCEV *, 8> NewOps;
400 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
401 const SCEV *Op = Ops[i];
402 const SCEV *Remainder = SE.getConstant(Ty, 0);
403 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
404 // Op now has ElSize factored out.
405 ScaledOps.push_back(Op);
406 if (!Remainder->isZero())
407 NewOps.push_back(Remainder);
408 AnyNonZeroIndices = true;
410 // The operand was not divisible, so add it to the list of operands
411 // we'll scan next iteration.
412 NewOps.push_back(Ops[i]);
415 // If we made any changes, update Ops.
416 if (!ScaledOps.empty()) {
418 SimplifyAddOperands(Ops, Ty, SE);
423 // Record the scaled array index for this level of the type. If
424 // we didn't find any operands that could be factored, tentatively
425 // assume that element zero was selected (since the zero offset
426 // would obviously be folded away).
427 Value *Scaled = ScaledOps.empty() ?
428 Constant::getNullValue(Ty) :
429 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
430 GepIndices.push_back(Scaled);
432 // Collect struct field index operands.
433 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
434 bool FoundFieldNo = false;
435 // An empty struct has no fields.
436 if (STy->getNumElements() == 0) break;
437 // Field offsets are known. See if a constant offset falls within any of
438 // the struct fields.
441 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
442 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
443 const StructLayout &SL = *DL.getStructLayout(STy);
444 uint64_t FullOffset = C->getValue()->getZExtValue();
445 if (FullOffset < SL.getSizeInBytes()) {
446 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
447 GepIndices.push_back(
448 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
449 ElTy = STy->getTypeAtIndex(ElIdx);
451 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
452 AnyNonZeroIndices = true;
456 // If no struct field offsets were found, tentatively assume that
457 // field zero was selected (since the zero offset would obviously
460 ElTy = STy->getTypeAtIndex(0u);
461 GepIndices.push_back(
462 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
466 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
467 ElTy = ATy->getElementType();
472 // If none of the operands were convertible to proper GEP indices, cast
473 // the base to i8* and do an ugly getelementptr with that. It's still
474 // better than ptrtoint+arithmetic+inttoptr at least.
475 if (!AnyNonZeroIndices) {
476 // Cast the base to i8*.
477 V = InsertNoopCastOfTo(V,
478 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
480 assert(!isa<Instruction>(V) ||
481 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
483 // Expand the operands for a plain byte offset.
484 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
486 // Fold a GEP with constant operands.
487 if (Constant *CLHS = dyn_cast<Constant>(V))
488 if (Constant *CRHS = dyn_cast<Constant>(Idx))
489 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
491 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
492 unsigned ScanLimit = 6;
493 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
494 // Scanning starts from the last instruction before the insertion point.
495 BasicBlock::iterator IP = Builder.GetInsertPoint();
496 if (IP != BlockBegin) {
498 for (; ScanLimit; --IP, --ScanLimit) {
499 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
501 if (isa<DbgInfoIntrinsic>(IP))
503 if (IP->getOpcode() == Instruction::GetElementPtr &&
504 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
506 if (IP == BlockBegin) break;
510 // Save the original insertion point so we can restore it when we're done.
511 BuilderType::InsertPointGuard Guard(Builder);
513 // Move the insertion point out of as many loops as we can.
514 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
515 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
516 BasicBlock *Preheader = L->getLoopPreheader();
517 if (!Preheader) break;
519 // Ok, move up a level.
520 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
524 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
525 rememberInstruction(GEP);
530 // Save the original insertion point so we can restore it when we're done.
531 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP();
533 // Move the insertion point out of as many loops as we can.
534 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
535 if (!L->isLoopInvariant(V)) break;
537 bool AnyIndexNotLoopInvariant = false;
538 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
539 E = GepIndices.end(); I != E; ++I)
540 if (!L->isLoopInvariant(*I)) {
541 AnyIndexNotLoopInvariant = true;
544 if (AnyIndexNotLoopInvariant)
547 BasicBlock *Preheader = L->getLoopPreheader();
548 if (!Preheader) break;
550 // Ok, move up a level.
551 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
554 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
555 // because ScalarEvolution may have changed the address arithmetic to
556 // compute a value which is beyond the end of the allocated object.
558 if (V->getType() != PTy)
559 Casted = InsertNoopCastOfTo(Casted, PTy);
560 Value *GEP = Builder.CreateGEP(Casted,
563 Ops.push_back(SE.getUnknown(GEP));
564 rememberInstruction(GEP);
566 // Restore the original insert point.
567 Builder.restoreIP(SaveInsertPt);
569 return expand(SE.getAddExpr(Ops));
572 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
573 /// SCEV expansion. If they are nested, this is the most nested. If they are
574 /// neighboring, pick the later.
575 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
579 if (A->contains(B)) return B;
580 if (B->contains(A)) return A;
581 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
582 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
583 return A; // Arbitrarily break the tie.
586 /// getRelevantLoop - Get the most relevant loop associated with the given
587 /// expression, according to PickMostRelevantLoop.
588 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
589 // Test whether we've already computed the most relevant loop for this SCEV.
590 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
591 RelevantLoops.insert(std::make_pair(S, nullptr));
593 return Pair.first->second;
595 if (isa<SCEVConstant>(S))
596 // A constant has no relevant loops.
598 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
599 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
600 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
601 // A non-instruction has no relevant loops.
604 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
605 const Loop *L = nullptr;
606 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
608 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
610 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
611 return RelevantLoops[N] = L;
613 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
614 const Loop *Result = getRelevantLoop(C->getOperand());
615 return RelevantLoops[C] = Result;
617 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
619 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
620 getRelevantLoop(D->getRHS()),
622 return RelevantLoops[D] = Result;
624 llvm_unreachable("Unexpected SCEV type!");
629 /// LoopCompare - Compare loops by PickMostRelevantLoop.
633 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
635 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
636 std::pair<const Loop *, const SCEV *> RHS) const {
637 // Keep pointer operands sorted at the end.
638 if (LHS.second->getType()->isPointerTy() !=
639 RHS.second->getType()->isPointerTy())
640 return LHS.second->getType()->isPointerTy();
642 // Compare loops with PickMostRelevantLoop.
643 if (LHS.first != RHS.first)
644 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
646 // If one operand is a non-constant negative and the other is not,
647 // put the non-constant negative on the right so that a sub can
648 // be used instead of a negate and add.
649 if (LHS.second->isNonConstantNegative()) {
650 if (!RHS.second->isNonConstantNegative())
652 } else if (RHS.second->isNonConstantNegative())
655 // Otherwise they are equivalent according to this comparison.
662 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
663 Type *Ty = SE.getEffectiveSCEVType(S->getType());
665 // Collect all the add operands in a loop, along with their associated loops.
666 // Iterate in reverse so that constants are emitted last, all else equal, and
667 // so that pointer operands are inserted first, which the code below relies on
668 // to form more involved GEPs.
669 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
670 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
671 E(S->op_begin()); I != E; ++I)
672 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
674 // Sort by loop. Use a stable sort so that constants follow non-constants and
675 // pointer operands precede non-pointer operands.
676 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
678 // Emit instructions to add all the operands. Hoist as much as possible
679 // out of loops, and form meaningful getelementptrs where possible.
680 Value *Sum = nullptr;
681 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
682 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
683 const Loop *CurLoop = I->first;
684 const SCEV *Op = I->second;
686 // This is the first operand. Just expand it.
689 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
690 // The running sum expression is a pointer. Try to form a getelementptr
691 // at this level with that as the base.
692 SmallVector<const SCEV *, 4> NewOps;
693 for (; I != E && I->first == CurLoop; ++I) {
694 // If the operand is SCEVUnknown and not instructions, peek through
695 // it, to enable more of it to be folded into the GEP.
696 const SCEV *X = I->second;
697 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
698 if (!isa<Instruction>(U->getValue()))
699 X = SE.getSCEV(U->getValue());
702 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
703 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
704 // The running sum is an integer, and there's a pointer at this level.
705 // Try to form a getelementptr. If the running sum is instructions,
706 // use a SCEVUnknown to avoid re-analyzing them.
707 SmallVector<const SCEV *, 4> NewOps;
708 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
710 for (++I; I != E && I->first == CurLoop; ++I)
711 NewOps.push_back(I->second);
712 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
713 } else if (Op->isNonConstantNegative()) {
714 // Instead of doing a negate and add, just do a subtract.
715 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
716 Sum = InsertNoopCastOfTo(Sum, Ty);
717 Sum = InsertBinop(Instruction::Sub, Sum, W);
721 Value *W = expandCodeFor(Op, Ty);
722 Sum = InsertNoopCastOfTo(Sum, Ty);
723 // Canonicalize a constant to the RHS.
724 if (isa<Constant>(Sum)) std::swap(Sum, W);
725 Sum = InsertBinop(Instruction::Add, Sum, W);
733 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
734 Type *Ty = SE.getEffectiveSCEVType(S->getType());
736 // Collect all the mul operands in a loop, along with their associated loops.
737 // Iterate in reverse so that constants are emitted last, all else equal.
738 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
739 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
740 E(S->op_begin()); I != E; ++I)
741 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
743 // Sort by loop. Use a stable sort so that constants follow non-constants.
744 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
746 // Emit instructions to mul all the operands. Hoist as much as possible
748 Value *Prod = nullptr;
749 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
750 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
751 const SCEV *Op = I->second;
753 // This is the first operand. Just expand it.
756 } else if (Op->isAllOnesValue()) {
757 // Instead of doing a multiply by negative one, just do a negate.
758 Prod = InsertNoopCastOfTo(Prod, Ty);
759 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
763 Value *W = expandCodeFor(Op, Ty);
764 Prod = InsertNoopCastOfTo(Prod, Ty);
765 // Canonicalize a constant to the RHS.
766 if (isa<Constant>(Prod)) std::swap(Prod, W);
767 Prod = InsertBinop(Instruction::Mul, Prod, W);
775 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
776 Type *Ty = SE.getEffectiveSCEVType(S->getType());
778 Value *LHS = expandCodeFor(S->getLHS(), Ty);
779 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
780 const APInt &RHS = SC->getValue()->getValue();
781 if (RHS.isPowerOf2())
782 return InsertBinop(Instruction::LShr, LHS,
783 ConstantInt::get(Ty, RHS.logBase2()));
786 Value *RHS = expandCodeFor(S->getRHS(), Ty);
787 return InsertBinop(Instruction::UDiv, LHS, RHS);
790 /// Move parts of Base into Rest to leave Base with the minimal
791 /// expression that provides a pointer operand suitable for a
793 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
794 ScalarEvolution &SE) {
795 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
796 Base = A->getStart();
797 Rest = SE.getAddExpr(Rest,
798 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
799 A->getStepRecurrence(SE),
801 A->getNoWrapFlags(SCEV::FlagNW)));
803 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
804 Base = A->getOperand(A->getNumOperands()-1);
805 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
806 NewAddOps.back() = Rest;
807 Rest = SE.getAddExpr(NewAddOps);
808 ExposePointerBase(Base, Rest, SE);
812 /// Determine if this is a well-behaved chain of instructions leading back to
813 /// the PHI. If so, it may be reused by expanded expressions.
814 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
816 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
817 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
819 // If any of the operands don't dominate the insert position, bail.
820 // Addrec operands are always loop-invariant, so this can only happen
821 // if there are instructions which haven't been hoisted.
822 if (L == IVIncInsertLoop) {
823 for (User::op_iterator OI = IncV->op_begin()+1,
824 OE = IncV->op_end(); OI != OE; ++OI)
825 if (Instruction *OInst = dyn_cast<Instruction>(OI))
826 if (!SE.DT->dominates(OInst, IVIncInsertPos))
829 // Advance to the next instruction.
830 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
834 if (IncV->mayHaveSideEffects())
840 return isNormalAddRecExprPHI(PN, IncV, L);
843 /// getIVIncOperand returns an induction variable increment's induction
844 /// variable operand.
846 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
847 /// operands dominate InsertPos.
849 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
850 /// simple patterns generated by getAddRecExprPHILiterally and
851 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
852 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
853 Instruction *InsertPos,
855 if (IncV == InsertPos)
858 switch (IncV->getOpcode()) {
861 // Check for a simple Add/Sub or GEP of a loop invariant step.
862 case Instruction::Add:
863 case Instruction::Sub: {
864 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
865 if (!OInst || SE.DT->dominates(OInst, InsertPos))
866 return dyn_cast<Instruction>(IncV->getOperand(0));
869 case Instruction::BitCast:
870 return dyn_cast<Instruction>(IncV->getOperand(0));
871 case Instruction::GetElementPtr:
872 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
874 if (isa<Constant>(*I))
876 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
877 if (!SE.DT->dominates(OInst, InsertPos))
881 // allow any kind of GEP as long as it can be hoisted.
884 // This must be a pointer addition of constants (pretty), which is already
885 // handled, or some number of address-size elements (ugly). Ugly geps
886 // have 2 operands. i1* is used by the expander to represent an
887 // address-size element.
888 if (IncV->getNumOperands() != 2)
890 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
891 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
892 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
896 return dyn_cast<Instruction>(IncV->getOperand(0));
900 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
901 /// it available to other uses in this loop. Recursively hoist any operands,
902 /// until we reach a value that dominates InsertPos.
903 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
904 if (SE.DT->dominates(IncV, InsertPos))
907 // InsertPos must itself dominate IncV so that IncV's new position satisfies
908 // its existing users.
909 if (isa<PHINode>(InsertPos)
910 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
913 // Check that the chain of IV operands leading back to Phi can be hoisted.
914 SmallVector<Instruction*, 4> IVIncs;
916 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
919 // IncV is safe to hoist.
920 IVIncs.push_back(IncV);
922 if (SE.DT->dominates(IncV, InsertPos))
925 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
926 E = IVIncs.rend(); I != E; ++I) {
927 (*I)->moveBefore(InsertPos);
932 /// Determine if this cyclic phi is in a form that would have been generated by
933 /// LSR. We don't care if the phi was actually expanded in this pass, as long
934 /// as it is in a low-cost form, for example, no implied multiplication. This
935 /// should match any patterns generated by getAddRecExprPHILiterally and
937 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
939 for(Instruction *IVOper = IncV;
940 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
941 /*allowScale=*/false));) {
948 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
949 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
950 /// need to materialize IV increments elsewhere to handle difficult situations.
951 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
952 Type *ExpandTy, Type *IntTy,
955 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
956 if (ExpandTy->isPointerTy()) {
957 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
958 // If the step isn't constant, don't use an implicitly scaled GEP, because
959 // that would require a multiply inside the loop.
960 if (!isa<ConstantInt>(StepV))
961 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
962 GEPPtrTy->getAddressSpace());
963 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
964 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
965 if (IncV->getType() != PN->getType()) {
966 IncV = Builder.CreateBitCast(IncV, PN->getType());
967 rememberInstruction(IncV);
971 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
972 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
973 rememberInstruction(IncV);
978 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
979 /// position. This routine assumes that this is possible (has been checked).
980 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
981 Instruction *Pos, PHINode *LoopPhi) {
983 if (DT->dominates(InstToHoist, Pos))
985 // Make sure the increment is where we want it. But don't move it
986 // down past a potential existing post-inc user.
987 InstToHoist->moveBefore(Pos);
989 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
990 } while (InstToHoist != LoopPhi);
993 /// \brief Check whether we can cheaply express the requested SCEV in terms of
994 /// the available PHI SCEV by truncation and/or invertion of the step.
995 static bool canBeCheaplyTransformed(ScalarEvolution &SE,
996 const SCEVAddRecExpr *Phi,
997 const SCEVAddRecExpr *Requested,
999 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1000 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1002 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1005 // Try truncate it if necessary.
1006 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1010 // Check whether truncation will help.
1011 if (Phi == Requested) {
1016 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1017 if (SE.getAddExpr(Requested->getStart(),
1018 SE.getNegativeSCEV(Requested)) == Phi) {
1026 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1027 if (!isa<IntegerType>(AR->getType()))
1030 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1031 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1032 const SCEV *Step = AR->getStepRecurrence(SE);
1033 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1034 SE.getSignExtendExpr(AR, WideTy));
1035 const SCEV *ExtendAfterOp =
1036 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1037 return ExtendAfterOp == OpAfterExtend;
1040 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1041 if (!isa<IntegerType>(AR->getType()))
1044 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1045 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1046 const SCEV *Step = AR->getStepRecurrence(SE);
1047 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1048 SE.getZeroExtendExpr(AR, WideTy));
1049 const SCEV *ExtendAfterOp =
1050 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1051 return ExtendAfterOp == OpAfterExtend;
1054 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1055 /// the base addrec, which is the addrec without any non-loop-dominating
1056 /// values, and return the PHI.
1058 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1064 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1066 // Reuse a previously-inserted PHI, if present.
1067 BasicBlock *LatchBlock = L->getLoopLatch();
1069 PHINode *AddRecPhiMatch = nullptr;
1070 Instruction *IncV = nullptr;
1074 // Only try partially matching scevs that need truncation and/or
1075 // step-inversion if we know this loop is outside the current loop.
1076 bool TryNonMatchingSCEV = IVIncInsertLoop &&
1077 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1079 for (BasicBlock::iterator I = L->getHeader()->begin();
1080 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1081 if (!SE.isSCEVable(PN->getType()))
1084 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1088 bool IsMatchingSCEV = PhiSCEV == Normalized;
1089 // We only handle truncation and inversion of phi recurrences for the
1090 // expanded expression if the expanded expression's loop dominates the
1091 // loop we insert to. Check now, so we can bail out early.
1092 if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1095 Instruction *TempIncV =
1096 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1098 // Check whether we can reuse this PHI node.
1100 if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1102 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1105 if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1109 // Stop if we have found an exact match SCEV.
1110 if (IsMatchingSCEV) {
1114 AddRecPhiMatch = PN;
1118 // Try whether the phi can be translated into the requested form
1119 // (truncated and/or offset by a constant).
1120 if ((!TruncTy || InvertStep) &&
1121 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1122 // Record the phi node. But don't stop we might find an exact match
1124 AddRecPhiMatch = PN;
1126 TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1130 if (AddRecPhiMatch) {
1131 // Potentially, move the increment. We have made sure in
1132 // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1133 if (L == IVIncInsertLoop)
1134 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1136 // Ok, the add recurrence looks usable.
1137 // Remember this PHI, even in post-inc mode.
1138 InsertedValues.insert(AddRecPhiMatch);
1139 // Remember the increment.
1140 rememberInstruction(IncV);
1141 return AddRecPhiMatch;
1145 // Save the original insertion point so we can restore it when we're done.
1146 BuilderType::InsertPointGuard Guard(Builder);
1148 // Another AddRec may need to be recursively expanded below. For example, if
1149 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1150 // loop. Remove this loop from the PostIncLoops set before expanding such
1151 // AddRecs. Otherwise, we cannot find a valid position for the step
1152 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1153 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1154 // so it's not worth implementing SmallPtrSet::swap.
1155 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1156 PostIncLoops.clear();
1158 // Expand code for the start value.
1159 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1160 L->getHeader()->begin());
1162 // StartV must be hoisted into L's preheader to dominate the new phi.
1163 assert(!isa<Instruction>(StartV) ||
1164 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1167 // Expand code for the step value. Do this before creating the PHI so that PHI
1168 // reuse code doesn't see an incomplete PHI.
1169 const SCEV *Step = Normalized->getStepRecurrence(SE);
1170 // If the stride is negative, insert a sub instead of an add for the increment
1171 // (unless it's a constant, because subtracts of constants are canonicalized
1173 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1175 Step = SE.getNegativeSCEV(Step);
1176 // Expand the step somewhere that dominates the loop header.
1177 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1179 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1180 // we actually do emit an addition. It does not apply if we emit a
1182 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1183 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1186 BasicBlock *Header = L->getHeader();
1187 Builder.SetInsertPoint(Header, Header->begin());
1188 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1189 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1190 Twine(IVName) + ".iv");
1191 rememberInstruction(PN);
1193 // Create the step instructions and populate the PHI.
1194 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1195 BasicBlock *Pred = *HPI;
1197 // Add a start value.
1198 if (!L->contains(Pred)) {
1199 PN->addIncoming(StartV, Pred);
1203 // Create a step value and add it to the PHI.
1204 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1205 // instructions at IVIncInsertPos.
1206 Instruction *InsertPos = L == IVIncInsertLoop ?
1207 IVIncInsertPos : Pred->getTerminator();
1208 Builder.SetInsertPoint(InsertPos);
1209 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1211 if (isa<OverflowingBinaryOperator>(IncV)) {
1213 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1215 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1217 PN->addIncoming(IncV, Pred);
1220 // After expanding subexpressions, restore the PostIncLoops set so the caller
1221 // can ensure that IVIncrement dominates the current uses.
1222 PostIncLoops = SavedPostIncLoops;
1224 // Remember this PHI, even in post-inc mode.
1225 InsertedValues.insert(PN);
1230 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1231 Type *STy = S->getType();
1232 Type *IntTy = SE.getEffectiveSCEVType(STy);
1233 const Loop *L = S->getLoop();
1235 // Determine a normalized form of this expression, which is the expression
1236 // before any post-inc adjustment is made.
1237 const SCEVAddRecExpr *Normalized = S;
1238 if (PostIncLoops.count(L)) {
1239 PostIncLoopSet Loops;
1242 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, nullptr,
1243 nullptr, Loops, SE, *SE.DT));
1246 // Strip off any non-loop-dominating component from the addrec start.
1247 const SCEV *Start = Normalized->getStart();
1248 const SCEV *PostLoopOffset = nullptr;
1249 if (!SE.properlyDominates(Start, L->getHeader())) {
1250 PostLoopOffset = Start;
1251 Start = SE.getConstant(Normalized->getType(), 0);
1252 Normalized = cast<SCEVAddRecExpr>(
1253 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1254 Normalized->getLoop(),
1255 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1258 // Strip off any non-loop-dominating component from the addrec step.
1259 const SCEV *Step = Normalized->getStepRecurrence(SE);
1260 const SCEV *PostLoopScale = nullptr;
1261 if (!SE.dominates(Step, L->getHeader())) {
1262 PostLoopScale = Step;
1263 Step = SE.getConstant(Normalized->getType(), 1);
1265 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1266 Start, Step, Normalized->getLoop(),
1267 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1270 // Expand the core addrec. If we need post-loop scaling, force it to
1271 // expand to an integer type to avoid the need for additional casting.
1272 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1273 // In some cases, we decide to reuse an existing phi node but need to truncate
1274 // it and/or invert the step.
1275 Type *TruncTy = nullptr;
1276 bool InvertStep = false;
1277 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy,
1278 TruncTy, InvertStep);
1280 // Accommodate post-inc mode, if necessary.
1282 if (!PostIncLoops.count(L))
1285 // In PostInc mode, use the post-incremented value.
1286 BasicBlock *LatchBlock = L->getLoopLatch();
1287 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1288 Result = PN->getIncomingValueForBlock(LatchBlock);
1290 // For an expansion to use the postinc form, the client must call
1291 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1292 // or dominated by IVIncInsertPos.
1293 if (isa<Instruction>(Result)
1294 && !SE.DT->dominates(cast<Instruction>(Result),
1295 Builder.GetInsertPoint())) {
1296 // The induction variable's postinc expansion does not dominate this use.
1297 // IVUsers tries to prevent this case, so it is rare. However, it can
1298 // happen when an IVUser outside the loop is not dominated by the latch
1299 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1300 // all cases. Consider a phi outide whose operand is replaced during
1301 // expansion with the value of the postinc user. Without fundamentally
1302 // changing the way postinc users are tracked, the only remedy is
1303 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1304 // but hopefully expandCodeFor handles that.
1306 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1308 Step = SE.getNegativeSCEV(Step);
1311 // Expand the step somewhere that dominates the loop header.
1312 BuilderType::InsertPointGuard Guard(Builder);
1313 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1315 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1319 // We have decided to reuse an induction variable of a dominating loop. Apply
1320 // truncation and/or invertion of the step.
1322 Type *ResTy = Result->getType();
1323 // Normalize the result type.
1324 if (ResTy != SE.getEffectiveSCEVType(ResTy))
1325 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1326 // Truncate the result.
1327 if (TruncTy != Result->getType()) {
1328 Result = Builder.CreateTrunc(Result, TruncTy);
1329 rememberInstruction(Result);
1331 // Invert the result.
1333 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1335 rememberInstruction(Result);
1339 // Re-apply any non-loop-dominating scale.
1340 if (PostLoopScale) {
1341 assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1342 Result = InsertNoopCastOfTo(Result, IntTy);
1343 Result = Builder.CreateMul(Result,
1344 expandCodeFor(PostLoopScale, IntTy));
1345 rememberInstruction(Result);
1348 // Re-apply any non-loop-dominating offset.
1349 if (PostLoopOffset) {
1350 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1351 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1352 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1354 Result = InsertNoopCastOfTo(Result, IntTy);
1355 Result = Builder.CreateAdd(Result,
1356 expandCodeFor(PostLoopOffset, IntTy));
1357 rememberInstruction(Result);
1364 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1365 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1367 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1368 const Loop *L = S->getLoop();
1370 // First check for an existing canonical IV in a suitable type.
1371 PHINode *CanonicalIV = nullptr;
1372 if (PHINode *PN = L->getCanonicalInductionVariable())
1373 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1376 // Rewrite an AddRec in terms of the canonical induction variable, if
1377 // its type is more narrow.
1379 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1380 SE.getTypeSizeInBits(Ty)) {
1381 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1382 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1383 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1384 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1385 S->getNoWrapFlags(SCEV::FlagNW)));
1386 BasicBlock::iterator NewInsertPt =
1387 std::next(BasicBlock::iterator(cast<Instruction>(V)));
1388 BuilderType::InsertPointGuard Guard(Builder);
1389 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1390 isa<LandingPadInst>(NewInsertPt))
1392 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1397 // {X,+,F} --> X + {0,+,F}
1398 if (!S->getStart()->isZero()) {
1399 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1400 NewOps[0] = SE.getConstant(Ty, 0);
1401 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1402 S->getNoWrapFlags(SCEV::FlagNW));
1404 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1405 // comments on expandAddToGEP for details.
1406 const SCEV *Base = S->getStart();
1407 const SCEV *RestArray[1] = { Rest };
1408 // Dig into the expression to find the pointer base for a GEP.
1409 ExposePointerBase(Base, RestArray[0], SE);
1410 // If we found a pointer, expand the AddRec with a GEP.
1411 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1412 // Make sure the Base isn't something exotic, such as a multiplied
1413 // or divided pointer value. In those cases, the result type isn't
1414 // actually a pointer type.
1415 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1416 Value *StartV = expand(Base);
1417 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1418 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1422 // Just do a normal add. Pre-expand the operands to suppress folding.
1423 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1424 SE.getUnknown(expand(Rest))));
1427 // If we don't yet have a canonical IV, create one.
1429 // Create and insert the PHI node for the induction variable in the
1431 BasicBlock *Header = L->getHeader();
1432 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1433 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1435 rememberInstruction(CanonicalIV);
1437 SmallSet<BasicBlock *, 4> PredSeen;
1438 Constant *One = ConstantInt::get(Ty, 1);
1439 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1440 BasicBlock *HP = *HPI;
1441 if (!PredSeen.insert(HP).second) {
1442 // There must be an incoming value for each predecessor, even the
1444 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1448 if (L->contains(HP)) {
1449 // Insert a unit add instruction right before the terminator
1450 // corresponding to the back-edge.
1451 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1453 HP->getTerminator());
1454 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1455 rememberInstruction(Add);
1456 CanonicalIV->addIncoming(Add, HP);
1458 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1463 // {0,+,1} --> Insert a canonical induction variable into the loop!
1464 if (S->isAffine() && S->getOperand(1)->isOne()) {
1465 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1466 "IVs with types different from the canonical IV should "
1467 "already have been handled!");
1471 // {0,+,F} --> {0,+,1} * F
1473 // If this is a simple linear addrec, emit it now as a special case.
1474 if (S->isAffine()) // {0,+,F} --> i*F
1476 expand(SE.getTruncateOrNoop(
1477 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1478 SE.getNoopOrAnyExtend(S->getOperand(1),
1479 CanonicalIV->getType())),
1482 // If this is a chain of recurrences, turn it into a closed form, using the
1483 // folders, then expandCodeFor the closed form. This allows the folders to
1484 // simplify the expression without having to build a bunch of special code
1485 // into this folder.
1486 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1488 // Promote S up to the canonical IV type, if the cast is foldable.
1489 const SCEV *NewS = S;
1490 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1491 if (isa<SCEVAddRecExpr>(Ext))
1494 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1495 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1497 // Truncate the result down to the original type, if needed.
1498 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1502 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1503 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1504 Value *V = expandCodeFor(S->getOperand(),
1505 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1506 Value *I = Builder.CreateTrunc(V, Ty);
1507 rememberInstruction(I);
1511 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1512 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1513 Value *V = expandCodeFor(S->getOperand(),
1514 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1515 Value *I = Builder.CreateZExt(V, Ty);
1516 rememberInstruction(I);
1520 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1521 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1522 Value *V = expandCodeFor(S->getOperand(),
1523 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1524 Value *I = Builder.CreateSExt(V, Ty);
1525 rememberInstruction(I);
1529 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1530 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1531 Type *Ty = LHS->getType();
1532 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1533 // In the case of mixed integer and pointer types, do the
1534 // rest of the comparisons as integer.
1535 if (S->getOperand(i)->getType() != Ty) {
1536 Ty = SE.getEffectiveSCEVType(Ty);
1537 LHS = InsertNoopCastOfTo(LHS, Ty);
1539 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1540 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1541 rememberInstruction(ICmp);
1542 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1543 rememberInstruction(Sel);
1546 // In the case of mixed integer and pointer types, cast the
1547 // final result back to the pointer type.
1548 if (LHS->getType() != S->getType())
1549 LHS = InsertNoopCastOfTo(LHS, S->getType());
1553 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1554 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1555 Type *Ty = LHS->getType();
1556 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1557 // In the case of mixed integer and pointer types, do the
1558 // rest of the comparisons as integer.
1559 if (S->getOperand(i)->getType() != Ty) {
1560 Ty = SE.getEffectiveSCEVType(Ty);
1561 LHS = InsertNoopCastOfTo(LHS, Ty);
1563 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1564 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1565 rememberInstruction(ICmp);
1566 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1567 rememberInstruction(Sel);
1570 // In the case of mixed integer and pointer types, cast the
1571 // final result back to the pointer type.
1572 if (LHS->getType() != S->getType())
1573 LHS = InsertNoopCastOfTo(LHS, S->getType());
1577 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1579 Builder.SetInsertPoint(IP->getParent(), IP);
1580 return expandCodeFor(SH, Ty);
1583 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1584 // Expand the code for this SCEV.
1585 Value *V = expand(SH);
1587 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1588 "non-trivial casts should be done with the SCEVs directly!");
1589 V = InsertNoopCastOfTo(V, Ty);
1594 Value *SCEVExpander::expand(const SCEV *S) {
1595 // Compute an insertion point for this SCEV object. Hoist the instructions
1596 // as far out in the loop nest as possible.
1597 Instruction *InsertPt = Builder.GetInsertPoint();
1598 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1599 L = L->getParentLoop())
1600 if (SE.isLoopInvariant(S, L)) {
1602 if (BasicBlock *Preheader = L->getLoopPreheader())
1603 InsertPt = Preheader->getTerminator();
1605 // LSR sets the insertion point for AddRec start/step values to the
1606 // block start to simplify value reuse, even though it's an invalid
1607 // position. SCEVExpander must correct for this in all cases.
1608 InsertPt = L->getHeader()->getFirstInsertionPt();
1611 // If the SCEV is computable at this level, insert it into the header
1612 // after the PHIs (and after any other instructions that we've inserted
1613 // there) so that it is guaranteed to dominate any user inside the loop.
1614 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1615 InsertPt = L->getHeader()->getFirstInsertionPt();
1616 while (InsertPt != Builder.GetInsertPoint()
1617 && (isInsertedInstruction(InsertPt)
1618 || isa<DbgInfoIntrinsic>(InsertPt))) {
1619 InsertPt = std::next(BasicBlock::iterator(InsertPt));
1624 // Check to see if we already expanded this here.
1625 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1626 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1627 if (I != InsertedExpressions.end())
1630 BuilderType::InsertPointGuard Guard(Builder);
1631 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1633 // Expand the expression into instructions.
1634 Value *V = visit(S);
1636 // Remember the expanded value for this SCEV at this location.
1638 // This is independent of PostIncLoops. The mapped value simply materializes
1639 // the expression at this insertion point. If the mapped value happened to be
1640 // a postinc expansion, it could be reused by a non-postinc user, but only if
1641 // its insertion point was already at the head of the loop.
1642 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1646 void SCEVExpander::rememberInstruction(Value *I) {
1647 if (!PostIncLoops.empty())
1648 InsertedPostIncValues.insert(I);
1650 InsertedValues.insert(I);
1653 /// getOrInsertCanonicalInductionVariable - This method returns the
1654 /// canonical induction variable of the specified type for the specified
1655 /// loop (inserting one if there is none). A canonical induction variable
1656 /// starts at zero and steps by one on each iteration.
1658 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1660 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1662 // Build a SCEV for {0,+,1}<L>.
1663 // Conservatively use FlagAnyWrap for now.
1664 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1665 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1667 // Emit code for it.
1668 BuilderType::InsertPointGuard Guard(Builder);
1669 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr,
1670 L->getHeader()->begin()));
1675 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1676 /// replace them with their most canonical representative. Return the number of
1677 /// phis eliminated.
1679 /// This does not depend on any SCEVExpander state but should be used in
1680 /// the same context that SCEVExpander is used.
1681 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1682 SmallVectorImpl<WeakVH> &DeadInsts,
1683 const TargetTransformInfo *TTI) {
1684 // Find integer phis in order of increasing width.
1685 SmallVector<PHINode*, 8> Phis;
1686 for (BasicBlock::iterator I = L->getHeader()->begin();
1687 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1688 Phis.push_back(Phi);
1691 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) {
1692 // Put pointers at the back and make sure pointer < pointer = false.
1693 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1694 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1695 return RHS->getType()->getPrimitiveSizeInBits() <
1696 LHS->getType()->getPrimitiveSizeInBits();
1699 unsigned NumElim = 0;
1700 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1701 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1702 // so narrow phis can reuse them.
1703 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1704 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1705 PHINode *Phi = *PIter;
1707 // Fold constant phis. They may be congruent to other constant phis and
1708 // would confuse the logic below that expects proper IVs.
1709 if (Value *V = SimplifyInstruction(Phi, DL, SE.TLI, SE.DT, SE.AC)) {
1710 Phi->replaceAllUsesWith(V);
1711 DeadInsts.push_back(Phi);
1713 DEBUG_WITH_TYPE(DebugType, dbgs()
1714 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1718 if (!SE.isSCEVable(Phi->getType()))
1721 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1724 if (Phi->getType()->isIntegerTy() && TTI
1725 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1726 // This phi can be freely truncated to the narrowest phi type. Map the
1727 // truncated expression to it so it will be reused for narrow types.
1728 const SCEV *TruncExpr =
1729 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1730 ExprToIVMap[TruncExpr] = Phi;
1735 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1737 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1740 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1741 Instruction *OrigInc =
1742 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1743 Instruction *IsomorphicInc =
1744 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1746 // If this phi has the same width but is more canonical, replace the
1747 // original with it. As part of the "more canonical" determination,
1748 // respect a prior decision to use an IV chain.
1749 if (OrigPhiRef->getType() == Phi->getType()
1750 && !(ChainedPhis.count(Phi)
1751 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1752 && (ChainedPhis.count(Phi)
1753 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1754 std::swap(OrigPhiRef, Phi);
1755 std::swap(OrigInc, IsomorphicInc);
1757 // Replacing the congruent phi is sufficient because acyclic redundancy
1758 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1759 // that a phi is congruent, it's often the head of an IV user cycle that
1760 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1761 // common case of a single IV increment so that DeleteDeadPHIs can remove
1762 // cycles that had postinc uses.
1763 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1764 IsomorphicInc->getType());
1765 if (OrigInc != IsomorphicInc
1766 && TruncExpr == SE.getSCEV(IsomorphicInc)
1767 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1768 || hoistIVInc(OrigInc, IsomorphicInc))) {
1769 DEBUG_WITH_TYPE(DebugType, dbgs()
1770 << "INDVARS: Eliminated congruent iv.inc: "
1771 << *IsomorphicInc << '\n');
1772 Value *NewInc = OrigInc;
1773 if (OrigInc->getType() != IsomorphicInc->getType()) {
1774 Instruction *IP = nullptr;
1775 if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1776 IP = PN->getParent()->getFirstInsertionPt();
1778 IP = OrigInc->getNextNode();
1780 IRBuilder<> Builder(IP);
1781 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1783 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1785 IsomorphicInc->replaceAllUsesWith(NewInc);
1786 DeadInsts.push_back(IsomorphicInc);
1789 DEBUG_WITH_TYPE(DebugType, dbgs()
1790 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1792 Value *NewIV = OrigPhiRef;
1793 if (OrigPhiRef->getType() != Phi->getType()) {
1794 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1795 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1796 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1798 Phi->replaceAllUsesWith(NewIV);
1799 DeadInsts.push_back(Phi);
1805 // Search for a SCEV subexpression that is not safe to expand. Any expression
1806 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1807 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1808 // instruction, but the important thing is that we prove the denominator is
1809 // nonzero before expansion.
1811 // IVUsers already checks that IV-derived expressions are safe. So this check is
1812 // only needed when the expression includes some subexpression that is not IV
1815 // Currently, we only allow division by a nonzero constant here. If this is
1816 // inadequate, we could easily allow division by SCEVUnknown by using
1817 // ValueTracking to check isKnownNonZero().
1819 // We cannot generally expand recurrences unless the step dominates the loop
1820 // header. The expander handles the special case of affine recurrences by
1821 // scaling the recurrence outside the loop, but this technique isn't generally
1822 // applicable. Expanding a nested recurrence outside a loop requires computing
1823 // binomial coefficients. This could be done, but the recurrence has to be in a
1824 // perfectly reduced form, which can't be guaranteed.
1825 struct SCEVFindUnsafe {
1826 ScalarEvolution &SE;
1829 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
1831 bool follow(const SCEV *S) {
1832 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1833 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1834 if (!SC || SC->getValue()->isZero()) {
1839 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1840 const SCEV *Step = AR->getStepRecurrence(SE);
1841 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
1848 bool isDone() const { return IsUnsafe; }
1853 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
1854 SCEVFindUnsafe Search(SE);
1855 visitAll(S, Search);
1856 return !Search.IsUnsafe;