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"
27 #include "llvm/Support/raw_ostream.h"
31 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
32 /// reusing an existing cast if a suitable one exists, moving an existing
33 /// cast if a suitable one exists but isn't in the right place, or
34 /// creating a new one.
35 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
36 Instruction::CastOps Op,
37 BasicBlock::iterator IP) {
38 // This function must be called with the builder having a valid insertion
39 // point. It doesn't need to be the actual IP where the uses of the returned
40 // cast will be added, but it must dominate such IP.
41 // We use this precondition to produce a cast that will dominate all its
42 // uses. In particular, this is crucial for the case where the builder's
43 // insertion point *is* the point where we were asked to put the cast.
44 // Since we don't know the builder's insertion point is actually
45 // where the uses will be added (only that it dominates it), we are
46 // not allowed to move it.
47 BasicBlock::iterator BIP = Builder.GetInsertPoint();
49 Instruction *Ret = nullptr;
51 // Check to see if there is already a cast!
52 for (User *U : V->users())
53 if (U->getType() == Ty)
54 if (CastInst *CI = dyn_cast<CastInst>(U))
55 if (CI->getOpcode() == Op) {
56 // If the cast isn't where we want it, create a new cast at IP.
57 // Likewise, do not reuse a cast at BIP because it must dominate
58 // instructions that might be inserted before BIP.
59 if (BasicBlock::iterator(CI) != IP || BIP == IP) {
60 // Create a new cast, and leave the old cast in place in case
61 // it is being used as an insert point. Clear its operand
62 // so that it doesn't hold anything live.
63 Ret = CastInst::Create(Op, V, Ty, "", IP);
65 CI->replaceAllUsesWith(Ret);
66 CI->setOperand(0, UndefValue::get(V->getType()));
75 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
77 // We assert at the end of the function since IP might point to an
78 // instruction with different dominance properties than a cast
79 // (an invoke for example) and not dominate BIP (but the cast does).
80 assert(SE.DT->dominates(Ret, BIP));
82 rememberInstruction(Ret);
86 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
87 /// which must be possible with a noop cast, doing what we can to share
89 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
90 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
91 assert((Op == Instruction::BitCast ||
92 Op == Instruction::PtrToInt ||
93 Op == Instruction::IntToPtr) &&
94 "InsertNoopCastOfTo cannot perform non-noop casts!");
95 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
96 "InsertNoopCastOfTo cannot change sizes!");
98 // Short-circuit unnecessary bitcasts.
99 if (Op == Instruction::BitCast) {
100 if (V->getType() == Ty)
102 if (CastInst *CI = dyn_cast<CastInst>(V)) {
103 if (CI->getOperand(0)->getType() == Ty)
104 return CI->getOperand(0);
107 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
108 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
109 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
110 if (CastInst *CI = dyn_cast<CastInst>(V))
111 if ((CI->getOpcode() == Instruction::PtrToInt ||
112 CI->getOpcode() == Instruction::IntToPtr) &&
113 SE.getTypeSizeInBits(CI->getType()) ==
114 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
115 return CI->getOperand(0);
116 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
117 if ((CE->getOpcode() == Instruction::PtrToInt ||
118 CE->getOpcode() == Instruction::IntToPtr) &&
119 SE.getTypeSizeInBits(CE->getType()) ==
120 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
121 return CE->getOperand(0);
124 // Fold a cast of a constant.
125 if (Constant *C = dyn_cast<Constant>(V))
126 return ConstantExpr::getCast(Op, C, Ty);
128 // Cast the argument at the beginning of the entry block, after
129 // any bitcasts of other arguments.
130 if (Argument *A = dyn_cast<Argument>(V)) {
131 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
132 while ((isa<BitCastInst>(IP) &&
133 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
134 cast<BitCastInst>(IP)->getOperand(0) != A) ||
135 isa<DbgInfoIntrinsic>(IP) ||
136 isa<LandingPadInst>(IP))
138 return ReuseOrCreateCast(A, Ty, Op, IP);
141 // Cast the instruction immediately after the instruction.
142 Instruction *I = cast<Instruction>(V);
143 BasicBlock::iterator IP = I; ++IP;
144 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
145 IP = II->getNormalDest()->begin();
146 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
148 return ReuseOrCreateCast(I, Ty, Op, IP);
151 /// InsertBinop - Insert the specified binary operator, doing a small amount
152 /// of work to avoid inserting an obviously redundant operation.
153 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
154 Value *LHS, Value *RHS) {
155 // Fold a binop with constant operands.
156 if (Constant *CLHS = dyn_cast<Constant>(LHS))
157 if (Constant *CRHS = dyn_cast<Constant>(RHS))
158 return ConstantExpr::get(Opcode, CLHS, CRHS);
160 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
161 unsigned ScanLimit = 6;
162 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
163 // Scanning starts from the last instruction before the insertion point.
164 BasicBlock::iterator IP = Builder.GetInsertPoint();
165 if (IP != BlockBegin) {
167 for (; ScanLimit; --IP, --ScanLimit) {
168 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
170 if (isa<DbgInfoIntrinsic>(IP))
172 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
173 IP->getOperand(1) == RHS)
175 if (IP == BlockBegin) break;
179 // Save the original insertion point so we can restore it when we're done.
180 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
181 BuilderType::InsertPointGuard Guard(Builder);
183 // Move the insertion point out of as many loops as we can.
184 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
185 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
186 BasicBlock *Preheader = L->getLoopPreheader();
187 if (!Preheader) break;
189 // Ok, move up a level.
190 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
193 // If we haven't found this binop, insert it.
194 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
195 BO->setDebugLoc(Loc);
196 rememberInstruction(BO);
201 /// FactorOutConstant - Test if S is divisible by Factor, using signed
202 /// division. If so, update S with Factor divided out and return true.
203 /// S need not be evenly divisible if a reasonable remainder can be
205 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
206 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
207 /// check to see if the divide was folded.
208 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
209 const SCEV *Factor, ScalarEvolution &SE,
210 const DataLayout &DL) {
211 // Everything is divisible by one.
217 S = SE.getConstant(S->getType(), 1);
221 // For a Constant, check for a multiple of the given factor.
222 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
226 // Check for divisibility.
227 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
229 ConstantInt::get(SE.getContext(),
230 C->getValue()->getValue().sdiv(
231 FC->getValue()->getValue()));
232 // If the quotient is zero and the remainder is non-zero, reject
233 // the value at this scale. It will be considered for subsequent
236 const SCEV *Div = SE.getConstant(CI);
239 SE.getAddExpr(Remainder,
240 SE.getConstant(C->getValue()->getValue().srem(
241 FC->getValue()->getValue())));
247 // In a Mul, check if there is a constant operand which is a multiple
248 // of the given factor.
249 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
250 // Size is known, check if there is a constant operand which is a multiple
251 // of the given factor. If so, we can factor it.
252 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
253 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
254 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
255 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
256 NewMulOps[0] = SE.getConstant(
257 C->getValue()->getValue().sdiv(FC->getValue()->getValue()));
258 S = SE.getMulExpr(NewMulOps);
263 // In an AddRec, check if both start and step are divisible.
264 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
265 const SCEV *Step = A->getStepRecurrence(SE);
266 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
267 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
269 if (!StepRem->isZero())
271 const SCEV *Start = A->getStart();
272 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
274 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
275 A->getNoWrapFlags(SCEV::FlagNW));
282 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
283 /// is the number of SCEVAddRecExprs present, which are kept at the end of
286 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
288 ScalarEvolution &SE) {
289 unsigned NumAddRecs = 0;
290 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
292 // Group Ops into non-addrecs and addrecs.
293 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
294 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
295 // Let ScalarEvolution sort and simplify the non-addrecs list.
296 const SCEV *Sum = NoAddRecs.empty() ?
297 SE.getConstant(Ty, 0) :
298 SE.getAddExpr(NoAddRecs);
299 // If it returned an add, use the operands. Otherwise it simplified
300 // the sum into a single value, so just use that.
302 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
303 Ops.append(Add->op_begin(), Add->op_end());
304 else if (!Sum->isZero())
306 // Then append the addrecs.
307 Ops.append(AddRecs.begin(), AddRecs.end());
310 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
311 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
312 /// This helps expose more opportunities for folding parts of the expressions
313 /// into GEP indices.
315 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
317 ScalarEvolution &SE) {
319 SmallVector<const SCEV *, 8> AddRecs;
320 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
321 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
322 const SCEV *Start = A->getStart();
323 if (Start->isZero()) break;
324 const SCEV *Zero = SE.getConstant(Ty, 0);
325 AddRecs.push_back(SE.getAddRecExpr(Zero,
326 A->getStepRecurrence(SE),
328 A->getNoWrapFlags(SCEV::FlagNW)));
329 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
331 Ops.append(Add->op_begin(), Add->op_end());
332 e += Add->getNumOperands();
337 if (!AddRecs.empty()) {
338 // Add the addrecs onto the end of the list.
339 Ops.append(AddRecs.begin(), AddRecs.end());
340 // Resort the operand list, moving any constants to the front.
341 SimplifyAddOperands(Ops, Ty, SE);
345 /// expandAddToGEP - Expand an addition expression with a pointer type into
346 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
347 /// BasicAliasAnalysis and other passes analyze the result. See the rules
348 /// for getelementptr vs. inttoptr in
349 /// http://llvm.org/docs/LangRef.html#pointeraliasing
352 /// Design note: The correctness of using getelementptr here depends on
353 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
354 /// they may introduce pointer arithmetic which may not be safely converted
355 /// into getelementptr.
357 /// Design note: It might seem desirable for this function to be more
358 /// loop-aware. If some of the indices are loop-invariant while others
359 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
360 /// loop-invariant portions of the overall computation outside the loop.
361 /// However, there are a few reasons this is not done here. Hoisting simple
362 /// arithmetic is a low-level optimization that often isn't very
363 /// important until late in the optimization process. In fact, passes
364 /// like InstructionCombining will combine GEPs, even if it means
365 /// pushing loop-invariant computation down into loops, so even if the
366 /// GEPs were split here, the work would quickly be undone. The
367 /// LoopStrengthReduction pass, which is usually run quite late (and
368 /// after the last InstructionCombining pass), takes care of hoisting
369 /// loop-invariant portions of expressions, after considering what
370 /// can be folded using target addressing modes.
372 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
373 const SCEV *const *op_end,
377 Type *ElTy = PTy->getElementType();
378 SmallVector<Value *, 4> GepIndices;
379 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
380 bool AnyNonZeroIndices = false;
382 // Split AddRecs up into parts as either of the parts may be usable
383 // without the other.
384 SplitAddRecs(Ops, Ty, SE);
386 Type *IntPtrTy = DL.getIntPtrType(PTy);
388 // Descend down the pointer's type and attempt to convert the other
389 // operands into GEP indices, at each level. The first index in a GEP
390 // indexes into the array implied by the pointer operand; the rest of
391 // the indices index into the element or field type selected by the
394 // If the scale size is not 0, attempt to factor out a scale for
396 SmallVector<const SCEV *, 8> ScaledOps;
397 if (ElTy->isSized()) {
398 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
399 if (!ElSize->isZero()) {
400 SmallVector<const SCEV *, 8> NewOps;
401 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
402 const SCEV *Op = Ops[i];
403 const SCEV *Remainder = SE.getConstant(Ty, 0);
404 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
405 // Op now has ElSize factored out.
406 ScaledOps.push_back(Op);
407 if (!Remainder->isZero())
408 NewOps.push_back(Remainder);
409 AnyNonZeroIndices = true;
411 // The operand was not divisible, so add it to the list of operands
412 // we'll scan next iteration.
413 NewOps.push_back(Ops[i]);
416 // If we made any changes, update Ops.
417 if (!ScaledOps.empty()) {
419 SimplifyAddOperands(Ops, Ty, SE);
424 // Record the scaled array index for this level of the type. If
425 // we didn't find any operands that could be factored, tentatively
426 // assume that element zero was selected (since the zero offset
427 // would obviously be folded away).
428 Value *Scaled = ScaledOps.empty() ?
429 Constant::getNullValue(Ty) :
430 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
431 GepIndices.push_back(Scaled);
433 // Collect struct field index operands.
434 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
435 bool FoundFieldNo = false;
436 // An empty struct has no fields.
437 if (STy->getNumElements() == 0) break;
438 // Field offsets are known. See if a constant offset falls within any of
439 // the struct fields.
442 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
443 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
444 const StructLayout &SL = *DL.getStructLayout(STy);
445 uint64_t FullOffset = C->getValue()->getZExtValue();
446 if (FullOffset < SL.getSizeInBytes()) {
447 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
448 GepIndices.push_back(
449 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
450 ElTy = STy->getTypeAtIndex(ElIdx);
452 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
453 AnyNonZeroIndices = true;
457 // If no struct field offsets were found, tentatively assume that
458 // field zero was selected (since the zero offset would obviously
461 ElTy = STy->getTypeAtIndex(0u);
462 GepIndices.push_back(
463 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
467 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
468 ElTy = ATy->getElementType();
473 // If none of the operands were convertible to proper GEP indices, cast
474 // the base to i8* and do an ugly getelementptr with that. It's still
475 // better than ptrtoint+arithmetic+inttoptr at least.
476 if (!AnyNonZeroIndices) {
477 // Cast the base to i8*.
478 V = InsertNoopCastOfTo(V,
479 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
481 assert(!isa<Instruction>(V) ||
482 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
484 // Expand the operands for a plain byte offset.
485 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
487 // Fold a GEP with constant operands.
488 if (Constant *CLHS = dyn_cast<Constant>(V))
489 if (Constant *CRHS = dyn_cast<Constant>(Idx))
490 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
492 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
493 unsigned ScanLimit = 6;
494 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
495 // Scanning starts from the last instruction before the insertion point.
496 BasicBlock::iterator IP = Builder.GetInsertPoint();
497 if (IP != BlockBegin) {
499 for (; ScanLimit; --IP, --ScanLimit) {
500 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
502 if (isa<DbgInfoIntrinsic>(IP))
504 if (IP->getOpcode() == Instruction::GetElementPtr &&
505 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
507 if (IP == BlockBegin) break;
511 // Save the original insertion point so we can restore it when we're done.
512 BuilderType::InsertPointGuard Guard(Builder);
514 // Move the insertion point out of as many loops as we can.
515 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
516 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
517 BasicBlock *Preheader = L->getLoopPreheader();
518 if (!Preheader) break;
520 // Ok, move up a level.
521 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
525 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
526 rememberInstruction(GEP);
531 // Save the original insertion point so we can restore it when we're done.
532 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP();
534 // Move the insertion point out of as many loops as we can.
535 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
536 if (!L->isLoopInvariant(V)) break;
538 bool AnyIndexNotLoopInvariant = false;
539 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
540 E = GepIndices.end(); I != E; ++I)
541 if (!L->isLoopInvariant(*I)) {
542 AnyIndexNotLoopInvariant = true;
545 if (AnyIndexNotLoopInvariant)
548 BasicBlock *Preheader = L->getLoopPreheader();
549 if (!Preheader) break;
551 // Ok, move up a level.
552 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
555 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
556 // because ScalarEvolution may have changed the address arithmetic to
557 // compute a value which is beyond the end of the allocated object.
559 if (V->getType() != PTy)
560 Casted = InsertNoopCastOfTo(Casted, PTy);
561 Value *GEP = Builder.CreateGEP(Casted,
564 Ops.push_back(SE.getUnknown(GEP));
565 rememberInstruction(GEP);
567 // Restore the original insert point.
568 Builder.restoreIP(SaveInsertPt);
570 return expand(SE.getAddExpr(Ops));
573 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
574 /// SCEV expansion. If they are nested, this is the most nested. If they are
575 /// neighboring, pick the later.
576 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
580 if (A->contains(B)) return B;
581 if (B->contains(A)) return A;
582 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
583 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
584 return A; // Arbitrarily break the tie.
587 /// getRelevantLoop - Get the most relevant loop associated with the given
588 /// expression, according to PickMostRelevantLoop.
589 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
590 // Test whether we've already computed the most relevant loop for this SCEV.
591 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
592 RelevantLoops.insert(std::make_pair(S, nullptr));
594 return Pair.first->second;
596 if (isa<SCEVConstant>(S))
597 // A constant has no relevant loops.
599 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
600 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
601 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
602 // A non-instruction has no relevant loops.
605 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
606 const Loop *L = nullptr;
607 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
609 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
611 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
612 return RelevantLoops[N] = L;
614 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
615 const Loop *Result = getRelevantLoop(C->getOperand());
616 return RelevantLoops[C] = Result;
618 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
620 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
621 getRelevantLoop(D->getRHS()),
623 return RelevantLoops[D] = Result;
625 llvm_unreachable("Unexpected SCEV type!");
630 /// LoopCompare - Compare loops by PickMostRelevantLoop.
634 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
636 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
637 std::pair<const Loop *, const SCEV *> RHS) const {
638 // Keep pointer operands sorted at the end.
639 if (LHS.second->getType()->isPointerTy() !=
640 RHS.second->getType()->isPointerTy())
641 return LHS.second->getType()->isPointerTy();
643 // Compare loops with PickMostRelevantLoop.
644 if (LHS.first != RHS.first)
645 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
647 // If one operand is a non-constant negative and the other is not,
648 // put the non-constant negative on the right so that a sub can
649 // be used instead of a negate and add.
650 if (LHS.second->isNonConstantNegative()) {
651 if (!RHS.second->isNonConstantNegative())
653 } else if (RHS.second->isNonConstantNegative())
656 // Otherwise they are equivalent according to this comparison.
663 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
664 Type *Ty = SE.getEffectiveSCEVType(S->getType());
666 // Collect all the add operands in a loop, along with their associated loops.
667 // Iterate in reverse so that constants are emitted last, all else equal, and
668 // so that pointer operands are inserted first, which the code below relies on
669 // to form more involved GEPs.
670 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
671 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
672 E(S->op_begin()); I != E; ++I)
673 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
675 // Sort by loop. Use a stable sort so that constants follow non-constants and
676 // pointer operands precede non-pointer operands.
677 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
679 // Emit instructions to add all the operands. Hoist as much as possible
680 // out of loops, and form meaningful getelementptrs where possible.
681 Value *Sum = nullptr;
682 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
683 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
684 const Loop *CurLoop = I->first;
685 const SCEV *Op = I->second;
687 // This is the first operand. Just expand it.
690 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
691 // The running sum expression is a pointer. Try to form a getelementptr
692 // at this level with that as the base.
693 SmallVector<const SCEV *, 4> NewOps;
694 for (; I != E && I->first == CurLoop; ++I) {
695 // If the operand is SCEVUnknown and not instructions, peek through
696 // it, to enable more of it to be folded into the GEP.
697 const SCEV *X = I->second;
698 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
699 if (!isa<Instruction>(U->getValue()))
700 X = SE.getSCEV(U->getValue());
703 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
704 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
705 // The running sum is an integer, and there's a pointer at this level.
706 // Try to form a getelementptr. If the running sum is instructions,
707 // use a SCEVUnknown to avoid re-analyzing them.
708 SmallVector<const SCEV *, 4> NewOps;
709 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
711 for (++I; I != E && I->first == CurLoop; ++I)
712 NewOps.push_back(I->second);
713 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
714 } else if (Op->isNonConstantNegative()) {
715 // Instead of doing a negate and add, just do a subtract.
716 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
717 Sum = InsertNoopCastOfTo(Sum, Ty);
718 Sum = InsertBinop(Instruction::Sub, Sum, W);
722 Value *W = expandCodeFor(Op, Ty);
723 Sum = InsertNoopCastOfTo(Sum, Ty);
724 // Canonicalize a constant to the RHS.
725 if (isa<Constant>(Sum)) std::swap(Sum, W);
726 Sum = InsertBinop(Instruction::Add, Sum, W);
734 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
735 Type *Ty = SE.getEffectiveSCEVType(S->getType());
737 // Collect all the mul operands in a loop, along with their associated loops.
738 // Iterate in reverse so that constants are emitted last, all else equal.
739 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
740 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
741 E(S->op_begin()); I != E; ++I)
742 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
744 // Sort by loop. Use a stable sort so that constants follow non-constants.
745 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
747 // Emit instructions to mul all the operands. Hoist as much as possible
749 Value *Prod = nullptr;
750 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
751 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
752 const SCEV *Op = I->second;
754 // This is the first operand. Just expand it.
757 } else if (Op->isAllOnesValue()) {
758 // Instead of doing a multiply by negative one, just do a negate.
759 Prod = InsertNoopCastOfTo(Prod, Ty);
760 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
764 Value *W = expandCodeFor(Op, Ty);
765 Prod = InsertNoopCastOfTo(Prod, Ty);
766 // Canonicalize a constant to the RHS.
767 if (isa<Constant>(Prod)) std::swap(Prod, W);
768 Prod = InsertBinop(Instruction::Mul, Prod, W);
776 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
777 Type *Ty = SE.getEffectiveSCEVType(S->getType());
779 Value *LHS = expandCodeFor(S->getLHS(), Ty);
780 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
781 const APInt &RHS = SC->getValue()->getValue();
782 if (RHS.isPowerOf2())
783 return InsertBinop(Instruction::LShr, LHS,
784 ConstantInt::get(Ty, RHS.logBase2()));
787 Value *RHS = expandCodeFor(S->getRHS(), Ty);
788 return InsertBinop(Instruction::UDiv, LHS, RHS);
791 /// Move parts of Base into Rest to leave Base with the minimal
792 /// expression that provides a pointer operand suitable for a
794 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
795 ScalarEvolution &SE) {
796 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
797 Base = A->getStart();
798 Rest = SE.getAddExpr(Rest,
799 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
800 A->getStepRecurrence(SE),
802 A->getNoWrapFlags(SCEV::FlagNW)));
804 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
805 Base = A->getOperand(A->getNumOperands()-1);
806 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
807 NewAddOps.back() = Rest;
808 Rest = SE.getAddExpr(NewAddOps);
809 ExposePointerBase(Base, Rest, SE);
813 /// Determine if this is a well-behaved chain of instructions leading back to
814 /// the PHI. If so, it may be reused by expanded expressions.
815 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
817 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
818 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
820 // If any of the operands don't dominate the insert position, bail.
821 // Addrec operands are always loop-invariant, so this can only happen
822 // if there are instructions which haven't been hoisted.
823 if (L == IVIncInsertLoop) {
824 for (User::op_iterator OI = IncV->op_begin()+1,
825 OE = IncV->op_end(); OI != OE; ++OI)
826 if (Instruction *OInst = dyn_cast<Instruction>(OI))
827 if (!SE.DT->dominates(OInst, IVIncInsertPos))
830 // Advance to the next instruction.
831 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
835 if (IncV->mayHaveSideEffects())
841 return isNormalAddRecExprPHI(PN, IncV, L);
844 /// getIVIncOperand returns an induction variable increment's induction
845 /// variable operand.
847 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
848 /// operands dominate InsertPos.
850 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
851 /// simple patterns generated by getAddRecExprPHILiterally and
852 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
853 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
854 Instruction *InsertPos,
856 if (IncV == InsertPos)
859 switch (IncV->getOpcode()) {
862 // Check for a simple Add/Sub or GEP of a loop invariant step.
863 case Instruction::Add:
864 case Instruction::Sub: {
865 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
866 if (!OInst || SE.DT->dominates(OInst, InsertPos))
867 return dyn_cast<Instruction>(IncV->getOperand(0));
870 case Instruction::BitCast:
871 return dyn_cast<Instruction>(IncV->getOperand(0));
872 case Instruction::GetElementPtr:
873 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
875 if (isa<Constant>(*I))
877 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
878 if (!SE.DT->dominates(OInst, InsertPos))
882 // allow any kind of GEP as long as it can be hoisted.
885 // This must be a pointer addition of constants (pretty), which is already
886 // handled, or some number of address-size elements (ugly). Ugly geps
887 // have 2 operands. i1* is used by the expander to represent an
888 // address-size element.
889 if (IncV->getNumOperands() != 2)
891 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
892 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
893 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
897 return dyn_cast<Instruction>(IncV->getOperand(0));
901 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
902 /// it available to other uses in this loop. Recursively hoist any operands,
903 /// until we reach a value that dominates InsertPos.
904 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
905 if (SE.DT->dominates(IncV, InsertPos))
908 // InsertPos must itself dominate IncV so that IncV's new position satisfies
909 // its existing users.
910 if (isa<PHINode>(InsertPos)
911 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
914 // Check that the chain of IV operands leading back to Phi can be hoisted.
915 SmallVector<Instruction*, 4> IVIncs;
917 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
920 // IncV is safe to hoist.
921 IVIncs.push_back(IncV);
923 if (SE.DT->dominates(IncV, InsertPos))
926 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
927 E = IVIncs.rend(); I != E; ++I) {
928 (*I)->moveBefore(InsertPos);
933 /// Determine if this cyclic phi is in a form that would have been generated by
934 /// LSR. We don't care if the phi was actually expanded in this pass, as long
935 /// as it is in a low-cost form, for example, no implied multiplication. This
936 /// should match any patterns generated by getAddRecExprPHILiterally and
938 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
940 for(Instruction *IVOper = IncV;
941 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
942 /*allowScale=*/false));) {
949 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
950 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
951 /// need to materialize IV increments elsewhere to handle difficult situations.
952 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
953 Type *ExpandTy, Type *IntTy,
956 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
957 if (ExpandTy->isPointerTy()) {
958 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
959 // If the step isn't constant, don't use an implicitly scaled GEP, because
960 // that would require a multiply inside the loop.
961 if (!isa<ConstantInt>(StepV))
962 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
963 GEPPtrTy->getAddressSpace());
964 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
965 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
966 if (IncV->getType() != PN->getType()) {
967 IncV = Builder.CreateBitCast(IncV, PN->getType());
968 rememberInstruction(IncV);
972 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
973 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
974 rememberInstruction(IncV);
979 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
980 /// position. This routine assumes that this is possible (has been checked).
981 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
982 Instruction *Pos, PHINode *LoopPhi) {
984 if (DT->dominates(InstToHoist, Pos))
986 // Make sure the increment is where we want it. But don't move it
987 // down past a potential existing post-inc user.
988 InstToHoist->moveBefore(Pos);
990 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
991 } while (InstToHoist != LoopPhi);
994 /// \brief Check whether we can cheaply express the requested SCEV in terms of
995 /// the available PHI SCEV by truncation and/or invertion of the step.
996 static bool canBeCheaplyTransformed(ScalarEvolution &SE,
997 const SCEVAddRecExpr *Phi,
998 const SCEVAddRecExpr *Requested,
1000 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1001 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1003 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1006 // Try truncate it if necessary.
1007 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1011 // Check whether truncation will help.
1012 if (Phi == Requested) {
1017 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1018 if (SE.getAddExpr(Requested->getStart(),
1019 SE.getNegativeSCEV(Requested)) == Phi) {
1027 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1028 if (!isa<IntegerType>(AR->getType()))
1031 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1032 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1033 const SCEV *Step = AR->getStepRecurrence(SE);
1034 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1035 SE.getSignExtendExpr(AR, WideTy));
1036 const SCEV *ExtendAfterOp =
1037 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1038 return ExtendAfterOp == OpAfterExtend;
1041 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1042 if (!isa<IntegerType>(AR->getType()))
1045 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1046 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1047 const SCEV *Step = AR->getStepRecurrence(SE);
1048 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1049 SE.getZeroExtendExpr(AR, WideTy));
1050 const SCEV *ExtendAfterOp =
1051 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1052 return ExtendAfterOp == OpAfterExtend;
1055 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1056 /// the base addrec, which is the addrec without any non-loop-dominating
1057 /// values, and return the PHI.
1059 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1065 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1067 // Reuse a previously-inserted PHI, if present.
1068 BasicBlock *LatchBlock = L->getLoopLatch();
1070 PHINode *AddRecPhiMatch = nullptr;
1071 Instruction *IncV = nullptr;
1075 // Only try partially matching scevs that need truncation and/or
1076 // step-inversion if we know this loop is outside the current loop.
1077 bool TryNonMatchingSCEV = IVIncInsertLoop &&
1078 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1080 for (BasicBlock::iterator I = L->getHeader()->begin();
1081 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1082 if (!SE.isSCEVable(PN->getType()))
1085 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1089 bool IsMatchingSCEV = PhiSCEV == Normalized;
1090 // We only handle truncation and inversion of phi recurrences for the
1091 // expanded expression if the expanded expression's loop dominates the
1092 // loop we insert to. Check now, so we can bail out early.
1093 if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1096 Instruction *TempIncV =
1097 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1099 // Check whether we can reuse this PHI node.
1101 if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1103 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1106 if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1110 // Stop if we have found an exact match SCEV.
1111 if (IsMatchingSCEV) {
1115 AddRecPhiMatch = PN;
1119 // Try whether the phi can be translated into the requested form
1120 // (truncated and/or offset by a constant).
1121 if ((!TruncTy || InvertStep) &&
1122 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1123 // Record the phi node. But don't stop we might find an exact match
1125 AddRecPhiMatch = PN;
1127 TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1131 if (AddRecPhiMatch) {
1132 // Potentially, move the increment. We have made sure in
1133 // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1134 if (L == IVIncInsertLoop)
1135 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1137 // Ok, the add recurrence looks usable.
1138 // Remember this PHI, even in post-inc mode.
1139 InsertedValues.insert(AddRecPhiMatch);
1140 // Remember the increment.
1141 rememberInstruction(IncV);
1142 return AddRecPhiMatch;
1146 // Save the original insertion point so we can restore it when we're done.
1147 BuilderType::InsertPointGuard Guard(Builder);
1149 // Another AddRec may need to be recursively expanded below. For example, if
1150 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1151 // loop. Remove this loop from the PostIncLoops set before expanding such
1152 // AddRecs. Otherwise, we cannot find a valid position for the step
1153 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1154 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1155 // so it's not worth implementing SmallPtrSet::swap.
1156 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1157 PostIncLoops.clear();
1159 // Expand code for the start value.
1160 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1161 L->getHeader()->begin());
1163 // StartV must be hoisted into L's preheader to dominate the new phi.
1164 assert(!isa<Instruction>(StartV) ||
1165 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1168 // Expand code for the step value. Do this before creating the PHI so that PHI
1169 // reuse code doesn't see an incomplete PHI.
1170 const SCEV *Step = Normalized->getStepRecurrence(SE);
1171 // If the stride is negative, insert a sub instead of an add for the increment
1172 // (unless it's a constant, because subtracts of constants are canonicalized
1174 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1176 Step = SE.getNegativeSCEV(Step);
1177 // Expand the step somewhere that dominates the loop header.
1178 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1180 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1181 // we actually do emit an addition. It does not apply if we emit a
1183 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1184 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1187 BasicBlock *Header = L->getHeader();
1188 Builder.SetInsertPoint(Header, Header->begin());
1189 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1190 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1191 Twine(IVName) + ".iv");
1192 rememberInstruction(PN);
1194 // Create the step instructions and populate the PHI.
1195 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1196 BasicBlock *Pred = *HPI;
1198 // Add a start value.
1199 if (!L->contains(Pred)) {
1200 PN->addIncoming(StartV, Pred);
1204 // Create a step value and add it to the PHI.
1205 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1206 // instructions at IVIncInsertPos.
1207 Instruction *InsertPos = L == IVIncInsertLoop ?
1208 IVIncInsertPos : Pred->getTerminator();
1209 Builder.SetInsertPoint(InsertPos);
1210 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1212 if (isa<OverflowingBinaryOperator>(IncV)) {
1214 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1216 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1218 PN->addIncoming(IncV, Pred);
1221 // After expanding subexpressions, restore the PostIncLoops set so the caller
1222 // can ensure that IVIncrement dominates the current uses.
1223 PostIncLoops = SavedPostIncLoops;
1225 // Remember this PHI, even in post-inc mode.
1226 InsertedValues.insert(PN);
1231 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1232 Type *STy = S->getType();
1233 Type *IntTy = SE.getEffectiveSCEVType(STy);
1234 const Loop *L = S->getLoop();
1236 // Determine a normalized form of this expression, which is the expression
1237 // before any post-inc adjustment is made.
1238 const SCEVAddRecExpr *Normalized = S;
1239 if (PostIncLoops.count(L)) {
1240 PostIncLoopSet Loops;
1243 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, nullptr,
1244 nullptr, Loops, SE, *SE.DT));
1247 // Strip off any non-loop-dominating component from the addrec start.
1248 const SCEV *Start = Normalized->getStart();
1249 const SCEV *PostLoopOffset = nullptr;
1250 if (!SE.properlyDominates(Start, L->getHeader())) {
1251 PostLoopOffset = Start;
1252 Start = SE.getConstant(Normalized->getType(), 0);
1253 Normalized = cast<SCEVAddRecExpr>(
1254 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1255 Normalized->getLoop(),
1256 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1259 // Strip off any non-loop-dominating component from the addrec step.
1260 const SCEV *Step = Normalized->getStepRecurrence(SE);
1261 const SCEV *PostLoopScale = nullptr;
1262 if (!SE.dominates(Step, L->getHeader())) {
1263 PostLoopScale = Step;
1264 Step = SE.getConstant(Normalized->getType(), 1);
1266 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1267 Start, Step, Normalized->getLoop(),
1268 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1271 // Expand the core addrec. If we need post-loop scaling, force it to
1272 // expand to an integer type to avoid the need for additional casting.
1273 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1274 // In some cases, we decide to reuse an existing phi node but need to truncate
1275 // it and/or invert the step.
1276 Type *TruncTy = nullptr;
1277 bool InvertStep = false;
1278 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy,
1279 TruncTy, InvertStep);
1281 // Accommodate post-inc mode, if necessary.
1283 if (!PostIncLoops.count(L))
1286 // In PostInc mode, use the post-incremented value.
1287 BasicBlock *LatchBlock = L->getLoopLatch();
1288 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1289 Result = PN->getIncomingValueForBlock(LatchBlock);
1291 // For an expansion to use the postinc form, the client must call
1292 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1293 // or dominated by IVIncInsertPos.
1294 if (isa<Instruction>(Result)
1295 && !SE.DT->dominates(cast<Instruction>(Result),
1296 Builder.GetInsertPoint())) {
1297 // The induction variable's postinc expansion does not dominate this use.
1298 // IVUsers tries to prevent this case, so it is rare. However, it can
1299 // happen when an IVUser outside the loop is not dominated by the latch
1300 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1301 // all cases. Consider a phi outide whose operand is replaced during
1302 // expansion with the value of the postinc user. Without fundamentally
1303 // changing the way postinc users are tracked, the only remedy is
1304 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1305 // but hopefully expandCodeFor handles that.
1307 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1309 Step = SE.getNegativeSCEV(Step);
1312 // Expand the step somewhere that dominates the loop header.
1313 BuilderType::InsertPointGuard Guard(Builder);
1314 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1316 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1320 // We have decided to reuse an induction variable of a dominating loop. Apply
1321 // truncation and/or invertion of the step.
1323 Type *ResTy = Result->getType();
1324 // Normalize the result type.
1325 if (ResTy != SE.getEffectiveSCEVType(ResTy))
1326 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1327 // Truncate the result.
1328 if (TruncTy != Result->getType()) {
1329 Result = Builder.CreateTrunc(Result, TruncTy);
1330 rememberInstruction(Result);
1332 // Invert the result.
1334 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1336 rememberInstruction(Result);
1340 // Re-apply any non-loop-dominating scale.
1341 if (PostLoopScale) {
1342 assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1343 Result = InsertNoopCastOfTo(Result, IntTy);
1344 Result = Builder.CreateMul(Result,
1345 expandCodeFor(PostLoopScale, IntTy));
1346 rememberInstruction(Result);
1349 // Re-apply any non-loop-dominating offset.
1350 if (PostLoopOffset) {
1351 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1352 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1353 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1355 Result = InsertNoopCastOfTo(Result, IntTy);
1356 Result = Builder.CreateAdd(Result,
1357 expandCodeFor(PostLoopOffset, IntTy));
1358 rememberInstruction(Result);
1365 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1366 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1368 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1369 const Loop *L = S->getLoop();
1371 // First check for an existing canonical IV in a suitable type.
1372 PHINode *CanonicalIV = nullptr;
1373 if (PHINode *PN = L->getCanonicalInductionVariable())
1374 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1377 // Rewrite an AddRec in terms of the canonical induction variable, if
1378 // its type is more narrow.
1380 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1381 SE.getTypeSizeInBits(Ty)) {
1382 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1383 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1384 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1385 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1386 S->getNoWrapFlags(SCEV::FlagNW)));
1387 BasicBlock::iterator NewInsertPt =
1388 std::next(BasicBlock::iterator(cast<Instruction>(V)));
1389 BuilderType::InsertPointGuard Guard(Builder);
1390 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1391 isa<LandingPadInst>(NewInsertPt))
1393 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1398 // {X,+,F} --> X + {0,+,F}
1399 if (!S->getStart()->isZero()) {
1400 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1401 NewOps[0] = SE.getConstant(Ty, 0);
1402 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1403 S->getNoWrapFlags(SCEV::FlagNW));
1405 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1406 // comments on expandAddToGEP for details.
1407 const SCEV *Base = S->getStart();
1408 const SCEV *RestArray[1] = { Rest };
1409 // Dig into the expression to find the pointer base for a GEP.
1410 ExposePointerBase(Base, RestArray[0], SE);
1411 // If we found a pointer, expand the AddRec with a GEP.
1412 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1413 // Make sure the Base isn't something exotic, such as a multiplied
1414 // or divided pointer value. In those cases, the result type isn't
1415 // actually a pointer type.
1416 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1417 Value *StartV = expand(Base);
1418 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1419 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1423 // Just do a normal add. Pre-expand the operands to suppress folding.
1424 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1425 SE.getUnknown(expand(Rest))));
1428 // If we don't yet have a canonical IV, create one.
1430 // Create and insert the PHI node for the induction variable in the
1432 BasicBlock *Header = L->getHeader();
1433 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1434 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1436 rememberInstruction(CanonicalIV);
1438 SmallSet<BasicBlock *, 4> PredSeen;
1439 Constant *One = ConstantInt::get(Ty, 1);
1440 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1441 BasicBlock *HP = *HPI;
1442 if (!PredSeen.insert(HP).second) {
1443 // There must be an incoming value for each predecessor, even the
1445 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1449 if (L->contains(HP)) {
1450 // Insert a unit add instruction right before the terminator
1451 // corresponding to the back-edge.
1452 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1454 HP->getTerminator());
1455 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1456 rememberInstruction(Add);
1457 CanonicalIV->addIncoming(Add, HP);
1459 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1464 // {0,+,1} --> Insert a canonical induction variable into the loop!
1465 if (S->isAffine() && S->getOperand(1)->isOne()) {
1466 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1467 "IVs with types different from the canonical IV should "
1468 "already have been handled!");
1472 // {0,+,F} --> {0,+,1} * F
1474 // If this is a simple linear addrec, emit it now as a special case.
1475 if (S->isAffine()) // {0,+,F} --> i*F
1477 expand(SE.getTruncateOrNoop(
1478 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1479 SE.getNoopOrAnyExtend(S->getOperand(1),
1480 CanonicalIV->getType())),
1483 // If this is a chain of recurrences, turn it into a closed form, using the
1484 // folders, then expandCodeFor the closed form. This allows the folders to
1485 // simplify the expression without having to build a bunch of special code
1486 // into this folder.
1487 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1489 // Promote S up to the canonical IV type, if the cast is foldable.
1490 const SCEV *NewS = S;
1491 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1492 if (isa<SCEVAddRecExpr>(Ext))
1495 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1496 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1498 // Truncate the result down to the original type, if needed.
1499 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1503 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1504 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1505 Value *V = expandCodeFor(S->getOperand(),
1506 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1507 Value *I = Builder.CreateTrunc(V, Ty);
1508 rememberInstruction(I);
1512 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1513 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1514 Value *V = expandCodeFor(S->getOperand(),
1515 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1516 Value *I = Builder.CreateZExt(V, Ty);
1517 rememberInstruction(I);
1521 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1522 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1523 Value *V = expandCodeFor(S->getOperand(),
1524 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1525 Value *I = Builder.CreateSExt(V, Ty);
1526 rememberInstruction(I);
1530 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1531 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1532 Type *Ty = LHS->getType();
1533 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1534 // In the case of mixed integer and pointer types, do the
1535 // rest of the comparisons as integer.
1536 if (S->getOperand(i)->getType() != Ty) {
1537 Ty = SE.getEffectiveSCEVType(Ty);
1538 LHS = InsertNoopCastOfTo(LHS, Ty);
1540 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1541 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1542 rememberInstruction(ICmp);
1543 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1544 rememberInstruction(Sel);
1547 // In the case of mixed integer and pointer types, cast the
1548 // final result back to the pointer type.
1549 if (LHS->getType() != S->getType())
1550 LHS = InsertNoopCastOfTo(LHS, S->getType());
1554 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1555 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1556 Type *Ty = LHS->getType();
1557 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1558 // In the case of mixed integer and pointer types, do the
1559 // rest of the comparisons as integer.
1560 if (S->getOperand(i)->getType() != Ty) {
1561 Ty = SE.getEffectiveSCEVType(Ty);
1562 LHS = InsertNoopCastOfTo(LHS, Ty);
1564 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1565 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1566 rememberInstruction(ICmp);
1567 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1568 rememberInstruction(Sel);
1571 // In the case of mixed integer and pointer types, cast the
1572 // final result back to the pointer type.
1573 if (LHS->getType() != S->getType())
1574 LHS = InsertNoopCastOfTo(LHS, S->getType());
1578 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1580 Builder.SetInsertPoint(IP->getParent(), IP);
1581 return expandCodeFor(SH, Ty);
1584 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1585 // Expand the code for this SCEV.
1586 Value *V = expand(SH);
1588 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1589 "non-trivial casts should be done with the SCEVs directly!");
1590 V = InsertNoopCastOfTo(V, Ty);
1595 Value *SCEVExpander::expand(const SCEV *S) {
1596 // Compute an insertion point for this SCEV object. Hoist the instructions
1597 // as far out in the loop nest as possible.
1598 Instruction *InsertPt = Builder.GetInsertPoint();
1599 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1600 L = L->getParentLoop())
1601 if (SE.isLoopInvariant(S, L)) {
1603 if (BasicBlock *Preheader = L->getLoopPreheader())
1604 InsertPt = Preheader->getTerminator();
1606 // LSR sets the insertion point for AddRec start/step values to the
1607 // block start to simplify value reuse, even though it's an invalid
1608 // position. SCEVExpander must correct for this in all cases.
1609 InsertPt = L->getHeader()->getFirstInsertionPt();
1612 // If the SCEV is computable at this level, insert it into the header
1613 // after the PHIs (and after any other instructions that we've inserted
1614 // there) so that it is guaranteed to dominate any user inside the loop.
1615 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1616 InsertPt = L->getHeader()->getFirstInsertionPt();
1617 while (InsertPt != Builder.GetInsertPoint()
1618 && (isInsertedInstruction(InsertPt)
1619 || isa<DbgInfoIntrinsic>(InsertPt))) {
1620 InsertPt = std::next(BasicBlock::iterator(InsertPt));
1625 // Check to see if we already expanded this here.
1626 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1627 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1628 if (I != InsertedExpressions.end())
1631 BuilderType::InsertPointGuard Guard(Builder);
1632 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1634 // Expand the expression into instructions.
1635 Value *V = visit(S);
1637 // Remember the expanded value for this SCEV at this location.
1639 // This is independent of PostIncLoops. The mapped value simply materializes
1640 // the expression at this insertion point. If the mapped value happened to be
1641 // a postinc expansion, it could be reused by a non-postinc user, but only if
1642 // its insertion point was already at the head of the loop.
1643 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1647 void SCEVExpander::rememberInstruction(Value *I) {
1648 if (!PostIncLoops.empty())
1649 InsertedPostIncValues.insert(I);
1651 InsertedValues.insert(I);
1654 /// getOrInsertCanonicalInductionVariable - This method returns the
1655 /// canonical induction variable of the specified type for the specified
1656 /// loop (inserting one if there is none). A canonical induction variable
1657 /// starts at zero and steps by one on each iteration.
1659 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1661 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1663 // Build a SCEV for {0,+,1}<L>.
1664 // Conservatively use FlagAnyWrap for now.
1665 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1666 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1668 // Emit code for it.
1669 BuilderType::InsertPointGuard Guard(Builder);
1670 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr,
1671 L->getHeader()->begin()));
1676 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1677 /// replace them with their most canonical representative. Return the number of
1678 /// phis eliminated.
1680 /// This does not depend on any SCEVExpander state but should be used in
1681 /// the same context that SCEVExpander is used.
1682 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1683 SmallVectorImpl<WeakVH> &DeadInsts,
1684 const TargetTransformInfo *TTI) {
1685 // Find integer phis in order of increasing width.
1686 SmallVector<PHINode*, 8> Phis;
1687 for (BasicBlock::iterator I = L->getHeader()->begin();
1688 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1689 Phis.push_back(Phi);
1692 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) {
1693 // Put pointers at the back and make sure pointer < pointer = false.
1694 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1695 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1696 return RHS->getType()->getPrimitiveSizeInBits() <
1697 LHS->getType()->getPrimitiveSizeInBits();
1700 unsigned NumElim = 0;
1701 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1702 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1703 // so narrow phis can reuse them.
1704 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1705 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1706 PHINode *Phi = *PIter;
1708 // Fold constant phis. They may be congruent to other constant phis and
1709 // would confuse the logic below that expects proper IVs.
1710 if (Value *V = SimplifyInstruction(Phi, DL, SE.TLI, SE.DT, SE.AC)) {
1711 Phi->replaceAllUsesWith(V);
1712 DeadInsts.push_back(Phi);
1714 DEBUG_WITH_TYPE(DebugType, dbgs()
1715 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1719 if (!SE.isSCEVable(Phi->getType()))
1722 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1725 if (Phi->getType()->isIntegerTy() && TTI
1726 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1727 // This phi can be freely truncated to the narrowest phi type. Map the
1728 // truncated expression to it so it will be reused for narrow types.
1729 const SCEV *TruncExpr =
1730 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1731 ExprToIVMap[TruncExpr] = Phi;
1736 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1738 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1741 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1742 Instruction *OrigInc =
1743 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1744 Instruction *IsomorphicInc =
1745 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1747 // If this phi has the same width but is more canonical, replace the
1748 // original with it. As part of the "more canonical" determination,
1749 // respect a prior decision to use an IV chain.
1750 if (OrigPhiRef->getType() == Phi->getType()
1751 && !(ChainedPhis.count(Phi)
1752 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1753 && (ChainedPhis.count(Phi)
1754 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1755 std::swap(OrigPhiRef, Phi);
1756 std::swap(OrigInc, IsomorphicInc);
1758 // Replacing the congruent phi is sufficient because acyclic redundancy
1759 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1760 // that a phi is congruent, it's often the head of an IV user cycle that
1761 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1762 // common case of a single IV increment so that DeleteDeadPHIs can remove
1763 // cycles that had postinc uses.
1764 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1765 IsomorphicInc->getType());
1766 if (OrigInc != IsomorphicInc
1767 && TruncExpr == SE.getSCEV(IsomorphicInc)
1768 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1769 || hoistIVInc(OrigInc, IsomorphicInc))) {
1770 DEBUG_WITH_TYPE(DebugType, dbgs()
1771 << "INDVARS: Eliminated congruent iv.inc: "
1772 << *IsomorphicInc << '\n');
1773 Value *NewInc = OrigInc;
1774 if (OrigInc->getType() != IsomorphicInc->getType()) {
1775 Instruction *IP = nullptr;
1776 if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1777 IP = PN->getParent()->getFirstInsertionPt();
1779 IP = OrigInc->getNextNode();
1781 IRBuilder<> Builder(IP);
1782 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1784 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1786 IsomorphicInc->replaceAllUsesWith(NewInc);
1787 DeadInsts.push_back(IsomorphicInc);
1790 DEBUG_WITH_TYPE(DebugType, dbgs()
1791 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1793 Value *NewIV = OrigPhiRef;
1794 if (OrigPhiRef->getType() != Phi->getType()) {
1795 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1796 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1797 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1799 Phi->replaceAllUsesWith(NewIV);
1800 DeadInsts.push_back(Phi);
1806 // Search for a SCEV subexpression that is not safe to expand. Any expression
1807 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1808 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1809 // instruction, but the important thing is that we prove the denominator is
1810 // nonzero before expansion.
1812 // IVUsers already checks that IV-derived expressions are safe. So this check is
1813 // only needed when the expression includes some subexpression that is not IV
1816 // Currently, we only allow division by a nonzero constant here. If this is
1817 // inadequate, we could easily allow division by SCEVUnknown by using
1818 // ValueTracking to check isKnownNonZero().
1820 // We cannot generally expand recurrences unless the step dominates the loop
1821 // header. The expander handles the special case of affine recurrences by
1822 // scaling the recurrence outside the loop, but this technique isn't generally
1823 // applicable. Expanding a nested recurrence outside a loop requires computing
1824 // binomial coefficients. This could be done, but the recurrence has to be in a
1825 // perfectly reduced form, which can't be guaranteed.
1826 struct SCEVFindUnsafe {
1827 ScalarEvolution &SE;
1830 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
1832 bool follow(const SCEV *S) {
1833 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1834 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1835 if (!SC || SC->getValue()->isZero()) {
1840 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1841 const SCEV *Step = AR->getStepRecurrence(SE);
1842 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
1849 bool isDone() const { return IsUnsafe; }
1854 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
1855 SCEVFindUnsafe Search(SE);
1856 visitAll(S, Search);
1857 return !Search.IsUnsafe;