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/LoopInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/LLVMContext.h"
25 #include "llvm/Support/Debug.h"
29 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
30 /// reusing an existing cast if a suitable one exists, moving an existing
31 /// cast if a suitable one exists but isn't in the right place, or
32 /// creating a new one.
33 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
34 Instruction::CastOps Op,
35 BasicBlock::iterator IP) {
36 // This function must be called with the builder having a valid insertion
37 // point. It doesn't need to be the actual IP where the uses of the returned
38 // cast will be added, but it must dominate such IP.
39 // We use this precondition to produce a cast that will dominate all its
40 // uses. In particular, this is crucial for the case where the builder's
41 // insertion point *is* the point where we were asked to put the cast.
42 // Since we don't know the builder's insertion point is actually
43 // where the uses will be added (only that it dominates it), we are
44 // not allowed to move it.
45 BasicBlock::iterator BIP = Builder.GetInsertPoint();
47 Instruction *Ret = NULL;
49 // Check to see if there is already a cast!
50 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
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()));
76 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP);
78 // We assert at the end of the function since IP might point to an
79 // instruction with different dominance properties than a cast
80 // (an invoke for example) and not dominate BIP (but the cast does).
81 assert(SE.DT->dominates(Ret, BIP));
83 rememberInstruction(Ret);
87 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
88 /// which must be possible with a noop cast, doing what we can to share
90 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
91 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
92 assert((Op == Instruction::BitCast ||
93 Op == Instruction::PtrToInt ||
94 Op == Instruction::IntToPtr) &&
95 "InsertNoopCastOfTo cannot perform non-noop casts!");
96 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
97 "InsertNoopCastOfTo cannot change sizes!");
99 // Short-circuit unnecessary bitcasts.
100 if (Op == Instruction::BitCast) {
101 if (V->getType() == Ty)
103 if (CastInst *CI = dyn_cast<CastInst>(V)) {
104 if (CI->getOperand(0)->getType() == Ty)
105 return CI->getOperand(0);
108 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
109 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
110 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
111 if (CastInst *CI = dyn_cast<CastInst>(V))
112 if ((CI->getOpcode() == Instruction::PtrToInt ||
113 CI->getOpcode() == Instruction::IntToPtr) &&
114 SE.getTypeSizeInBits(CI->getType()) ==
115 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
116 return CI->getOperand(0);
117 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
118 if ((CE->getOpcode() == Instruction::PtrToInt ||
119 CE->getOpcode() == Instruction::IntToPtr) &&
120 SE.getTypeSizeInBits(CE->getType()) ==
121 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
122 return CE->getOperand(0);
125 // Fold a cast of a constant.
126 if (Constant *C = dyn_cast<Constant>(V))
127 return ConstantExpr::getCast(Op, C, Ty);
129 // Cast the argument at the beginning of the entry block, after
130 // any bitcasts of other arguments.
131 if (Argument *A = dyn_cast<Argument>(V)) {
132 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
133 while ((isa<BitCastInst>(IP) &&
134 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
135 cast<BitCastInst>(IP)->getOperand(0) != A) ||
136 isa<DbgInfoIntrinsic>(IP) ||
137 isa<LandingPadInst>(IP))
139 return ReuseOrCreateCast(A, Ty, Op, IP);
142 // Cast the instruction immediately after the instruction.
143 Instruction *I = cast<Instruction>(V);
144 BasicBlock::iterator IP = I; ++IP;
145 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
146 IP = II->getNormalDest()->begin();
147 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP))
149 return ReuseOrCreateCast(I, Ty, Op, IP);
152 /// InsertBinop - Insert the specified binary operator, doing a small amount
153 /// of work to avoid inserting an obviously redundant operation.
154 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
155 Value *LHS, Value *RHS) {
156 // Fold a binop with constant operands.
157 if (Constant *CLHS = dyn_cast<Constant>(LHS))
158 if (Constant *CRHS = dyn_cast<Constant>(RHS))
159 return ConstantExpr::get(Opcode, CLHS, CRHS);
161 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
162 unsigned ScanLimit = 6;
163 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
164 // Scanning starts from the last instruction before the insertion point.
165 BasicBlock::iterator IP = Builder.GetInsertPoint();
166 if (IP != BlockBegin) {
168 for (; ScanLimit; --IP, --ScanLimit) {
169 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
171 if (isa<DbgInfoIntrinsic>(IP))
173 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
174 IP->getOperand(1) == RHS)
176 if (IP == BlockBegin) break;
180 // Save the original insertion point so we can restore it when we're done.
181 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
182 BuilderType::InsertPointGuard Guard(Builder);
184 // Move the insertion point out of as many loops as we can.
185 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
186 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
187 BasicBlock *Preheader = L->getLoopPreheader();
188 if (!Preheader) break;
190 // Ok, move up a level.
191 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
194 // If we haven't found this binop, insert it.
195 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
196 BO->setDebugLoc(Loc);
197 rememberInstruction(BO);
202 /// FactorOutConstant - Test if S is divisible by Factor, using signed
203 /// division. If so, update S with Factor divided out and return true.
204 /// S need not be evenly divisible if a reasonable remainder can be
206 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
207 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
208 /// check to see if the divide was folded.
209 static bool FactorOutConstant(const SCEV *&S,
210 const SCEV *&Remainder,
213 const DataLayout *DL) {
214 // Everything is divisible by one.
220 S = SE.getConstant(S->getType(), 1);
224 // For a Constant, check for a multiple of the given factor.
225 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
229 // Check for divisibility.
230 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
232 ConstantInt::get(SE.getContext(),
233 C->getValue()->getValue().sdiv(
234 FC->getValue()->getValue()));
235 // If the quotient is zero and the remainder is non-zero, reject
236 // the value at this scale. It will be considered for subsequent
239 const SCEV *Div = SE.getConstant(CI);
242 SE.getAddExpr(Remainder,
243 SE.getConstant(C->getValue()->getValue().srem(
244 FC->getValue()->getValue())));
250 // In a Mul, check if there is a constant operand which is a multiple
251 // of the given factor.
252 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
254 // With DataLayout, the size is known. Check if there is a constant
255 // operand which is a multiple of the given factor. If so, we can
257 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
258 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
259 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
260 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
262 SE.getConstant(C->getValue()->getValue().sdiv(
263 FC->getValue()->getValue()));
264 S = SE.getMulExpr(NewMulOps);
268 // Without DataLayout, check if Factor can be factored out of any of the
269 // Mul's operands. If so, we can just remove it.
270 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
271 const SCEV *SOp = M->getOperand(i);
272 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
273 if (FactorOutConstant(SOp, Remainder, Factor, SE, DL) &&
274 Remainder->isZero()) {
275 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
277 S = SE.getMulExpr(NewMulOps);
284 // In an AddRec, check if both start and step are divisible.
285 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
286 const SCEV *Step = A->getStepRecurrence(SE);
287 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
288 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
290 if (!StepRem->isZero())
292 const SCEV *Start = A->getStart();
293 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
295 S = SE.getAddRecExpr(Start, Step, A->getLoop(),
296 A->getNoWrapFlags(SCEV::FlagNW));
303 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
304 /// is the number of SCEVAddRecExprs present, which are kept at the end of
307 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
309 ScalarEvolution &SE) {
310 unsigned NumAddRecs = 0;
311 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
313 // Group Ops into non-addrecs and addrecs.
314 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
315 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
316 // Let ScalarEvolution sort and simplify the non-addrecs list.
317 const SCEV *Sum = NoAddRecs.empty() ?
318 SE.getConstant(Ty, 0) :
319 SE.getAddExpr(NoAddRecs);
320 // If it returned an add, use the operands. Otherwise it simplified
321 // the sum into a single value, so just use that.
323 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
324 Ops.append(Add->op_begin(), Add->op_end());
325 else if (!Sum->isZero())
327 // Then append the addrecs.
328 Ops.append(AddRecs.begin(), AddRecs.end());
331 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
332 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
333 /// This helps expose more opportunities for folding parts of the expressions
334 /// into GEP indices.
336 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
338 ScalarEvolution &SE) {
340 SmallVector<const SCEV *, 8> AddRecs;
341 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
342 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
343 const SCEV *Start = A->getStart();
344 if (Start->isZero()) break;
345 const SCEV *Zero = SE.getConstant(Ty, 0);
346 AddRecs.push_back(SE.getAddRecExpr(Zero,
347 A->getStepRecurrence(SE),
349 A->getNoWrapFlags(SCEV::FlagNW)));
350 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
352 Ops.append(Add->op_begin(), Add->op_end());
353 e += Add->getNumOperands();
358 if (!AddRecs.empty()) {
359 // Add the addrecs onto the end of the list.
360 Ops.append(AddRecs.begin(), AddRecs.end());
361 // Resort the operand list, moving any constants to the front.
362 SimplifyAddOperands(Ops, Ty, SE);
366 /// expandAddToGEP - Expand an addition expression with a pointer type into
367 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
368 /// BasicAliasAnalysis and other passes analyze the result. See the rules
369 /// for getelementptr vs. inttoptr in
370 /// http://llvm.org/docs/LangRef.html#pointeraliasing
373 /// Design note: The correctness of using getelementptr here depends on
374 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
375 /// they may introduce pointer arithmetic which may not be safely converted
376 /// into getelementptr.
378 /// Design note: It might seem desirable for this function to be more
379 /// loop-aware. If some of the indices are loop-invariant while others
380 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
381 /// loop-invariant portions of the overall computation outside the loop.
382 /// However, there are a few reasons this is not done here. Hoisting simple
383 /// arithmetic is a low-level optimization that often isn't very
384 /// important until late in the optimization process. In fact, passes
385 /// like InstructionCombining will combine GEPs, even if it means
386 /// pushing loop-invariant computation down into loops, so even if the
387 /// GEPs were split here, the work would quickly be undone. The
388 /// LoopStrengthReduction pass, which is usually run quite late (and
389 /// after the last InstructionCombining pass), takes care of hoisting
390 /// loop-invariant portions of expressions, after considering what
391 /// can be folded using target addressing modes.
393 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
394 const SCEV *const *op_end,
398 Type *ElTy = PTy->getElementType();
399 SmallVector<Value *, 4> GepIndices;
400 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
401 bool AnyNonZeroIndices = false;
403 // Split AddRecs up into parts as either of the parts may be usable
404 // without the other.
405 SplitAddRecs(Ops, Ty, SE);
407 Type *IntPtrTy = SE.DL
408 ? SE.DL->getIntPtrType(PTy)
409 : Type::getInt64Ty(PTy->getContext());
411 // Descend down the pointer's type and attempt to convert the other
412 // operands into GEP indices, at each level. The first index in a GEP
413 // indexes into the array implied by the pointer operand; the rest of
414 // the indices index into the element or field type selected by the
417 // If the scale size is not 0, attempt to factor out a scale for
419 SmallVector<const SCEV *, 8> ScaledOps;
420 if (ElTy->isSized()) {
421 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
422 if (!ElSize->isZero()) {
423 SmallVector<const SCEV *, 8> NewOps;
424 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
425 const SCEV *Op = Ops[i];
426 const SCEV *Remainder = SE.getConstant(Ty, 0);
427 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.DL)) {
428 // Op now has ElSize factored out.
429 ScaledOps.push_back(Op);
430 if (!Remainder->isZero())
431 NewOps.push_back(Remainder);
432 AnyNonZeroIndices = true;
434 // The operand was not divisible, so add it to the list of operands
435 // we'll scan next iteration.
436 NewOps.push_back(Ops[i]);
439 // If we made any changes, update Ops.
440 if (!ScaledOps.empty()) {
442 SimplifyAddOperands(Ops, Ty, SE);
447 // Record the scaled array index for this level of the type. If
448 // we didn't find any operands that could be factored, tentatively
449 // assume that element zero was selected (since the zero offset
450 // would obviously be folded away).
451 Value *Scaled = ScaledOps.empty() ?
452 Constant::getNullValue(Ty) :
453 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
454 GepIndices.push_back(Scaled);
456 // Collect struct field index operands.
457 while (StructType *STy = dyn_cast<StructType>(ElTy)) {
458 bool FoundFieldNo = false;
459 // An empty struct has no fields.
460 if (STy->getNumElements() == 0) break;
462 // With DataLayout, field offsets are known. See if a constant offset
463 // falls within any of the struct fields.
464 if (Ops.empty()) break;
465 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
466 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
467 const StructLayout &SL = *SE.DL->getStructLayout(STy);
468 uint64_t FullOffset = C->getValue()->getZExtValue();
469 if (FullOffset < SL.getSizeInBytes()) {
470 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
471 GepIndices.push_back(
472 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
473 ElTy = STy->getTypeAtIndex(ElIdx);
475 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
476 AnyNonZeroIndices = true;
481 // Without DataLayout, just check for an offsetof expression of the
482 // appropriate struct type.
483 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
484 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
487 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
488 GepIndices.push_back(FieldNo);
490 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
491 Ops[i] = SE.getConstant(Ty, 0);
492 AnyNonZeroIndices = true;
498 // If no struct field offsets were found, tentatively assume that
499 // field zero was selected (since the zero offset would obviously
502 ElTy = STy->getTypeAtIndex(0u);
503 GepIndices.push_back(
504 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
508 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
509 ElTy = ATy->getElementType();
514 // If none of the operands were convertible to proper GEP indices, cast
515 // the base to i8* and do an ugly getelementptr with that. It's still
516 // better than ptrtoint+arithmetic+inttoptr at least.
517 if (!AnyNonZeroIndices) {
518 // Cast the base to i8*.
519 V = InsertNoopCastOfTo(V,
520 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
522 assert(!isa<Instruction>(V) ||
523 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint()));
525 // Expand the operands for a plain byte offset.
526 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
528 // Fold a GEP with constant operands.
529 if (Constant *CLHS = dyn_cast<Constant>(V))
530 if (Constant *CRHS = dyn_cast<Constant>(Idx))
531 return ConstantExpr::getGetElementPtr(CLHS, CRHS);
533 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
534 unsigned ScanLimit = 6;
535 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
536 // Scanning starts from the last instruction before the insertion point.
537 BasicBlock::iterator IP = Builder.GetInsertPoint();
538 if (IP != BlockBegin) {
540 for (; ScanLimit; --IP, --ScanLimit) {
541 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
543 if (isa<DbgInfoIntrinsic>(IP))
545 if (IP->getOpcode() == Instruction::GetElementPtr &&
546 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
548 if (IP == BlockBegin) break;
552 // Save the original insertion point so we can restore it when we're done.
553 BuilderType::InsertPointGuard Guard(Builder);
555 // Move the insertion point out of as many loops as we can.
556 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
557 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
558 BasicBlock *Preheader = L->getLoopPreheader();
559 if (!Preheader) break;
561 // Ok, move up a level.
562 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
566 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
567 rememberInstruction(GEP);
572 // Save the original insertion point so we can restore it when we're done.
573 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP();
575 // Move the insertion point out of as many loops as we can.
576 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
577 if (!L->isLoopInvariant(V)) break;
579 bool AnyIndexNotLoopInvariant = false;
580 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
581 E = GepIndices.end(); I != E; ++I)
582 if (!L->isLoopInvariant(*I)) {
583 AnyIndexNotLoopInvariant = true;
586 if (AnyIndexNotLoopInvariant)
589 BasicBlock *Preheader = L->getLoopPreheader();
590 if (!Preheader) break;
592 // Ok, move up a level.
593 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
596 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
597 // because ScalarEvolution may have changed the address arithmetic to
598 // compute a value which is beyond the end of the allocated object.
600 if (V->getType() != PTy)
601 Casted = InsertNoopCastOfTo(Casted, PTy);
602 Value *GEP = Builder.CreateGEP(Casted,
605 Ops.push_back(SE.getUnknown(GEP));
606 rememberInstruction(GEP);
608 // Restore the original insert point.
609 Builder.restoreIP(SaveInsertPt);
611 return expand(SE.getAddExpr(Ops));
614 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
615 /// SCEV expansion. If they are nested, this is the most nested. If they are
616 /// neighboring, pick the later.
617 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
621 if (A->contains(B)) return B;
622 if (B->contains(A)) return A;
623 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
624 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
625 return A; // Arbitrarily break the tie.
628 /// getRelevantLoop - Get the most relevant loop associated with the given
629 /// expression, according to PickMostRelevantLoop.
630 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
631 // Test whether we've already computed the most relevant loop for this SCEV.
632 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair =
633 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0)));
635 return Pair.first->second;
637 if (isa<SCEVConstant>(S))
638 // A constant has no relevant loops.
640 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
641 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
642 return Pair.first->second = SE.LI->getLoopFor(I->getParent());
643 // A non-instruction has no relevant loops.
646 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
648 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
650 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
652 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT);
653 return RelevantLoops[N] = L;
655 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
656 const Loop *Result = getRelevantLoop(C->getOperand());
657 return RelevantLoops[C] = Result;
659 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
661 PickMostRelevantLoop(getRelevantLoop(D->getLHS()),
662 getRelevantLoop(D->getRHS()),
664 return RelevantLoops[D] = Result;
666 llvm_unreachable("Unexpected SCEV type!");
671 /// LoopCompare - Compare loops by PickMostRelevantLoop.
675 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
677 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
678 std::pair<const Loop *, const SCEV *> RHS) const {
679 // Keep pointer operands sorted at the end.
680 if (LHS.second->getType()->isPointerTy() !=
681 RHS.second->getType()->isPointerTy())
682 return LHS.second->getType()->isPointerTy();
684 // Compare loops with PickMostRelevantLoop.
685 if (LHS.first != RHS.first)
686 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
688 // If one operand is a non-constant negative and the other is not,
689 // put the non-constant negative on the right so that a sub can
690 // be used instead of a negate and add.
691 if (LHS.second->isNonConstantNegative()) {
692 if (!RHS.second->isNonConstantNegative())
694 } else if (RHS.second->isNonConstantNegative())
697 // Otherwise they are equivalent according to this comparison.
704 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
705 Type *Ty = SE.getEffectiveSCEVType(S->getType());
707 // Collect all the add operands in a loop, along with their associated loops.
708 // Iterate in reverse so that constants are emitted last, all else equal, and
709 // so that pointer operands are inserted first, which the code below relies on
710 // to form more involved GEPs.
711 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
712 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
713 E(S->op_begin()); I != E; ++I)
714 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
716 // Sort by loop. Use a stable sort so that constants follow non-constants and
717 // pointer operands precede non-pointer operands.
718 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
720 // Emit instructions to add all the operands. Hoist as much as possible
721 // out of loops, and form meaningful getelementptrs where possible.
723 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
724 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
725 const Loop *CurLoop = I->first;
726 const SCEV *Op = I->second;
728 // This is the first operand. Just expand it.
731 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
732 // The running sum expression is a pointer. Try to form a getelementptr
733 // at this level with that as the base.
734 SmallVector<const SCEV *, 4> NewOps;
735 for (; I != E && I->first == CurLoop; ++I) {
736 // If the operand is SCEVUnknown and not instructions, peek through
737 // it, to enable more of it to be folded into the GEP.
738 const SCEV *X = I->second;
739 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
740 if (!isa<Instruction>(U->getValue()))
741 X = SE.getSCEV(U->getValue());
744 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
745 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
746 // The running sum is an integer, and there's a pointer at this level.
747 // Try to form a getelementptr. If the running sum is instructions,
748 // use a SCEVUnknown to avoid re-analyzing them.
749 SmallVector<const SCEV *, 4> NewOps;
750 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
752 for (++I; I != E && I->first == CurLoop; ++I)
753 NewOps.push_back(I->second);
754 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
755 } else if (Op->isNonConstantNegative()) {
756 // Instead of doing a negate and add, just do a subtract.
757 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
758 Sum = InsertNoopCastOfTo(Sum, Ty);
759 Sum = InsertBinop(Instruction::Sub, Sum, W);
763 Value *W = expandCodeFor(Op, Ty);
764 Sum = InsertNoopCastOfTo(Sum, Ty);
765 // Canonicalize a constant to the RHS.
766 if (isa<Constant>(Sum)) std::swap(Sum, W);
767 Sum = InsertBinop(Instruction::Add, Sum, W);
775 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
776 Type *Ty = SE.getEffectiveSCEVType(S->getType());
778 // Collect all the mul operands in a loop, along with their associated loops.
779 // Iterate in reverse so that constants are emitted last, all else equal.
780 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
781 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
782 E(S->op_begin()); I != E; ++I)
783 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
785 // Sort by loop. Use a stable sort so that constants follow non-constants.
786 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
788 // Emit instructions to mul all the operands. Hoist as much as possible
791 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
792 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
793 const SCEV *Op = I->second;
795 // This is the first operand. Just expand it.
798 } else if (Op->isAllOnesValue()) {
799 // Instead of doing a multiply by negative one, just do a negate.
800 Prod = InsertNoopCastOfTo(Prod, Ty);
801 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
805 Value *W = expandCodeFor(Op, Ty);
806 Prod = InsertNoopCastOfTo(Prod, Ty);
807 // Canonicalize a constant to the RHS.
808 if (isa<Constant>(Prod)) std::swap(Prod, W);
809 Prod = InsertBinop(Instruction::Mul, Prod, W);
817 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
818 Type *Ty = SE.getEffectiveSCEVType(S->getType());
820 Value *LHS = expandCodeFor(S->getLHS(), Ty);
821 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
822 const APInt &RHS = SC->getValue()->getValue();
823 if (RHS.isPowerOf2())
824 return InsertBinop(Instruction::LShr, LHS,
825 ConstantInt::get(Ty, RHS.logBase2()));
828 Value *RHS = expandCodeFor(S->getRHS(), Ty);
829 return InsertBinop(Instruction::UDiv, LHS, RHS);
832 /// Move parts of Base into Rest to leave Base with the minimal
833 /// expression that provides a pointer operand suitable for a
835 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
836 ScalarEvolution &SE) {
837 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
838 Base = A->getStart();
839 Rest = SE.getAddExpr(Rest,
840 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
841 A->getStepRecurrence(SE),
843 A->getNoWrapFlags(SCEV::FlagNW)));
845 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
846 Base = A->getOperand(A->getNumOperands()-1);
847 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
848 NewAddOps.back() = Rest;
849 Rest = SE.getAddExpr(NewAddOps);
850 ExposePointerBase(Base, Rest, SE);
854 /// Determine if this is a well-behaved chain of instructions leading back to
855 /// the PHI. If so, it may be reused by expanded expressions.
856 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
858 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
859 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
861 // If any of the operands don't dominate the insert position, bail.
862 // Addrec operands are always loop-invariant, so this can only happen
863 // if there are instructions which haven't been hoisted.
864 if (L == IVIncInsertLoop) {
865 for (User::op_iterator OI = IncV->op_begin()+1,
866 OE = IncV->op_end(); OI != OE; ++OI)
867 if (Instruction *OInst = dyn_cast<Instruction>(OI))
868 if (!SE.DT->dominates(OInst, IVIncInsertPos))
871 // Advance to the next instruction.
872 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
876 if (IncV->mayHaveSideEffects())
882 return isNormalAddRecExprPHI(PN, IncV, L);
885 /// getIVIncOperand returns an induction variable increment's induction
886 /// variable operand.
888 /// If allowScale is set, any type of GEP is allowed as long as the nonIV
889 /// operands dominate InsertPos.
891 /// If allowScale is not set, ensure that a GEP increment conforms to one of the
892 /// simple patterns generated by getAddRecExprPHILiterally and
893 /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
894 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
895 Instruction *InsertPos,
897 if (IncV == InsertPos)
900 switch (IncV->getOpcode()) {
903 // Check for a simple Add/Sub or GEP of a loop invariant step.
904 case Instruction::Add:
905 case Instruction::Sub: {
906 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
907 if (!OInst || SE.DT->dominates(OInst, InsertPos))
908 return dyn_cast<Instruction>(IncV->getOperand(0));
911 case Instruction::BitCast:
912 return dyn_cast<Instruction>(IncV->getOperand(0));
913 case Instruction::GetElementPtr:
914 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end();
916 if (isa<Constant>(*I))
918 if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
919 if (!SE.DT->dominates(OInst, InsertPos))
923 // allow any kind of GEP as long as it can be hoisted.
926 // This must be a pointer addition of constants (pretty), which is already
927 // handled, or some number of address-size elements (ugly). Ugly geps
928 // have 2 operands. i1* is used by the expander to represent an
929 // address-size element.
930 if (IncV->getNumOperands() != 2)
932 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
933 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
934 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
938 return dyn_cast<Instruction>(IncV->getOperand(0));
942 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
943 /// it available to other uses in this loop. Recursively hoist any operands,
944 /// until we reach a value that dominates InsertPos.
945 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
946 if (SE.DT->dominates(IncV, InsertPos))
949 // InsertPos must itself dominate IncV so that IncV's new position satisfies
950 // its existing users.
951 if (isa<PHINode>(InsertPos)
952 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent()))
955 // Check that the chain of IV operands leading back to Phi can be hoisted.
956 SmallVector<Instruction*, 4> IVIncs;
958 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
961 // IncV is safe to hoist.
962 IVIncs.push_back(IncV);
964 if (SE.DT->dominates(IncV, InsertPos))
967 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(),
968 E = IVIncs.rend(); I != E; ++I) {
969 (*I)->moveBefore(InsertPos);
974 /// Determine if this cyclic phi is in a form that would have been generated by
975 /// LSR. We don't care if the phi was actually expanded in this pass, as long
976 /// as it is in a low-cost form, for example, no implied multiplication. This
977 /// should match any patterns generated by getAddRecExprPHILiterally and
979 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
981 for(Instruction *IVOper = IncV;
982 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
983 /*allowScale=*/false));) {
990 /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
991 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
992 /// need to materialize IV increments elsewhere to handle difficult situations.
993 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
994 Type *ExpandTy, Type *IntTy,
997 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
998 if (ExpandTy->isPointerTy()) {
999 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1000 // If the step isn't constant, don't use an implicitly scaled GEP, because
1001 // that would require a multiply inside the loop.
1002 if (!isa<ConstantInt>(StepV))
1003 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1004 GEPPtrTy->getAddressSpace());
1005 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
1006 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
1007 if (IncV->getType() != PN->getType()) {
1008 IncV = Builder.CreateBitCast(IncV, PN->getType());
1009 rememberInstruction(IncV);
1012 IncV = useSubtract ?
1013 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1014 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1015 rememberInstruction(IncV);
1020 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the
1021 /// position. This routine assumes that this is possible (has been checked).
1022 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
1023 Instruction *Pos, PHINode *LoopPhi) {
1025 if (DT->dominates(InstToHoist, Pos))
1027 // Make sure the increment is where we want it. But don't move it
1028 // down past a potential existing post-inc user.
1029 InstToHoist->moveBefore(Pos);
1031 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
1032 } while (InstToHoist != LoopPhi);
1035 /// \brief Check whether we can cheaply express the requested SCEV in terms of
1036 /// the available PHI SCEV by truncation and/or invertion of the step.
1037 static bool canBeCheaplyTransformed(ScalarEvolution &SE,
1038 const SCEVAddRecExpr *Phi,
1039 const SCEVAddRecExpr *Requested,
1041 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1042 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1044 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1047 // Try truncate it if necessary.
1048 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1052 // Check whether truncation will help.
1053 if (Phi == Requested) {
1058 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1059 if (SE.getAddExpr(Requested->getStart(),
1060 SE.getNegativeSCEV(Requested)) == Phi) {
1068 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1069 /// the base addrec, which is the addrec without any non-loop-dominating
1070 /// values, and return the PHI.
1072 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1078 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1080 // Reuse a previously-inserted PHI, if present.
1081 BasicBlock *LatchBlock = L->getLoopLatch();
1083 PHINode *AddRecPhiMatch = 0;
1084 Instruction *IncV = 0;
1088 // Only try partially matching scevs that need truncation and/or
1089 // step-inversion if we know this loop is outside the current loop.
1090 bool TryNonMatchingSCEV = IVIncInsertLoop &&
1091 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1093 for (BasicBlock::iterator I = L->getHeader()->begin();
1094 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1095 if (!SE.isSCEVable(PN->getType()))
1098 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN));
1102 bool IsMatchingSCEV = PhiSCEV == Normalized;
1103 // We only handle truncation and inversion of phi recurrences for the
1104 // expanded expression if the expanded expression's loop dominates the
1105 // loop we insert to. Check now, so we can bail out early.
1106 if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1109 Instruction *TempIncV =
1110 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
1112 // Check whether we can reuse this PHI node.
1114 if (!isExpandedAddRecExprPHI(PN, TempIncV, L))
1116 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1119 if (!isNormalAddRecExprPHI(PN, TempIncV, L))
1123 // Stop if we have found an exact match SCEV.
1124 if (IsMatchingSCEV) {
1128 AddRecPhiMatch = PN;
1132 // Try whether the phi can be translated into the requested form
1133 // (truncated and/or offset by a constant).
1134 if ((!TruncTy || InvertStep) &&
1135 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1136 // Record the phi node. But don't stop we might find an exact match
1138 AddRecPhiMatch = PN;
1140 TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1144 if (AddRecPhiMatch) {
1145 // Potentially, move the increment. We have made sure in
1146 // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1147 if (L == IVIncInsertLoop)
1148 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1150 // Ok, the add recurrence looks usable.
1151 // Remember this PHI, even in post-inc mode.
1152 InsertedValues.insert(AddRecPhiMatch);
1153 // Remember the increment.
1154 rememberInstruction(IncV);
1155 return AddRecPhiMatch;
1159 // Save the original insertion point so we can restore it when we're done.
1160 BuilderType::InsertPointGuard Guard(Builder);
1162 // Another AddRec may need to be recursively expanded below. For example, if
1163 // this AddRec is quadratic, the StepV may itself be an AddRec in this
1164 // loop. Remove this loop from the PostIncLoops set before expanding such
1165 // AddRecs. Otherwise, we cannot find a valid position for the step
1166 // (i.e. StepV can never dominate its loop header). Ideally, we could do
1167 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1168 // so it's not worth implementing SmallPtrSet::swap.
1169 PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1170 PostIncLoops.clear();
1172 // Expand code for the start value.
1173 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1174 L->getHeader()->begin());
1176 // StartV must be hoisted into L's preheader to dominate the new phi.
1177 assert(!isa<Instruction>(StartV) ||
1178 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(),
1181 // Expand code for the step value. Do this before creating the PHI so that PHI
1182 // reuse code doesn't see an incomplete PHI.
1183 const SCEV *Step = Normalized->getStepRecurrence(SE);
1184 // If the stride is negative, insert a sub instead of an add for the increment
1185 // (unless it's a constant, because subtracts of constants are canonicalized
1187 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1189 Step = SE.getNegativeSCEV(Step);
1190 // Expand the step somewhere that dominates the loop header.
1191 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1194 BasicBlock *Header = L->getHeader();
1195 Builder.SetInsertPoint(Header, Header->begin());
1196 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1197 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1198 Twine(IVName) + ".iv");
1199 rememberInstruction(PN);
1201 // Create the step instructions and populate the PHI.
1202 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1203 BasicBlock *Pred = *HPI;
1205 // Add a start value.
1206 if (!L->contains(Pred)) {
1207 PN->addIncoming(StartV, Pred);
1211 // Create a step value and add it to the PHI.
1212 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1213 // instructions at IVIncInsertPos.
1214 Instruction *InsertPos = L == IVIncInsertLoop ?
1215 IVIncInsertPos : Pred->getTerminator();
1216 Builder.SetInsertPoint(InsertPos);
1217 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1218 if (isa<OverflowingBinaryOperator>(IncV)) {
1219 if (Normalized->getNoWrapFlags(SCEV::FlagNUW))
1220 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1221 if (Normalized->getNoWrapFlags(SCEV::FlagNSW))
1222 cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1224 PN->addIncoming(IncV, Pred);
1227 // After expanding subexpressions, restore the PostIncLoops set so the caller
1228 // can ensure that IVIncrement dominates the current uses.
1229 PostIncLoops = SavedPostIncLoops;
1231 // Remember this PHI, even in post-inc mode.
1232 InsertedValues.insert(PN);
1237 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1238 Type *STy = S->getType();
1239 Type *IntTy = SE.getEffectiveSCEVType(STy);
1240 const Loop *L = S->getLoop();
1242 // Determine a normalized form of this expression, which is the expression
1243 // before any post-inc adjustment is made.
1244 const SCEVAddRecExpr *Normalized = S;
1245 if (PostIncLoops.count(L)) {
1246 PostIncLoopSet Loops;
1249 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
1250 Loops, SE, *SE.DT));
1253 // Strip off any non-loop-dominating component from the addrec start.
1254 const SCEV *Start = Normalized->getStart();
1255 const SCEV *PostLoopOffset = 0;
1256 if (!SE.properlyDominates(Start, L->getHeader())) {
1257 PostLoopOffset = Start;
1258 Start = SE.getConstant(Normalized->getType(), 0);
1259 Normalized = cast<SCEVAddRecExpr>(
1260 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1261 Normalized->getLoop(),
1262 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1265 // Strip off any non-loop-dominating component from the addrec step.
1266 const SCEV *Step = Normalized->getStepRecurrence(SE);
1267 const SCEV *PostLoopScale = 0;
1268 if (!SE.dominates(Step, L->getHeader())) {
1269 PostLoopScale = Step;
1270 Step = SE.getConstant(Normalized->getType(), 1);
1272 cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1273 Start, Step, Normalized->getLoop(),
1274 Normalized->getNoWrapFlags(SCEV::FlagNW)));
1277 // Expand the core addrec. If we need post-loop scaling, force it to
1278 // expand to an integer type to avoid the need for additional casting.
1279 Type *ExpandTy = PostLoopScale ? IntTy : STy;
1280 // In some cases, we decide to reuse an existing phi node but need to truncate
1281 // it and/or invert the step.
1283 bool InvertStep = false;
1284 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy,
1285 TruncTy, InvertStep);
1287 // Accommodate post-inc mode, if necessary.
1289 if (!PostIncLoops.count(L))
1292 // In PostInc mode, use the post-incremented value.
1293 BasicBlock *LatchBlock = L->getLoopLatch();
1294 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1295 Result = PN->getIncomingValueForBlock(LatchBlock);
1297 // For an expansion to use the postinc form, the client must call
1298 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1299 // or dominated by IVIncInsertPos.
1300 if (isa<Instruction>(Result)
1301 && !SE.DT->dominates(cast<Instruction>(Result),
1302 Builder.GetInsertPoint())) {
1303 // The induction variable's postinc expansion does not dominate this use.
1304 // IVUsers tries to prevent this case, so it is rare. However, it can
1305 // happen when an IVUser outside the loop is not dominated by the latch
1306 // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1307 // all cases. Consider a phi outide whose operand is replaced during
1308 // expansion with the value of the postinc user. Without fundamentally
1309 // changing the way postinc users are tracked, the only remedy is
1310 // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1311 // but hopefully expandCodeFor handles that.
1313 !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1315 Step = SE.getNegativeSCEV(Step);
1318 // Expand the step somewhere that dominates the loop header.
1319 BuilderType::InsertPointGuard Guard(Builder);
1320 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
1322 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1326 // We have decided to reuse an induction variable of a dominating loop. Apply
1327 // truncation and/or invertion of the step.
1329 Type *ResTy = Result->getType();
1330 // Normalize the result type.
1331 if (ResTy != SE.getEffectiveSCEVType(ResTy))
1332 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1333 // Truncate the result.
1334 if (TruncTy != Result->getType()) {
1335 Result = Builder.CreateTrunc(Result, TruncTy);
1336 rememberInstruction(Result);
1338 // Invert the result.
1340 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1342 rememberInstruction(Result);
1346 // Re-apply any non-loop-dominating scale.
1347 if (PostLoopScale) {
1348 assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1349 Result = InsertNoopCastOfTo(Result, IntTy);
1350 Result = Builder.CreateMul(Result,
1351 expandCodeFor(PostLoopScale, IntTy));
1352 rememberInstruction(Result);
1355 // Re-apply any non-loop-dominating offset.
1356 if (PostLoopOffset) {
1357 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1358 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1359 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1361 Result = InsertNoopCastOfTo(Result, IntTy);
1362 Result = Builder.CreateAdd(Result,
1363 expandCodeFor(PostLoopOffset, IntTy));
1364 rememberInstruction(Result);
1371 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1372 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1374 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1375 const Loop *L = S->getLoop();
1377 // First check for an existing canonical IV in a suitable type.
1378 PHINode *CanonicalIV = 0;
1379 if (PHINode *PN = L->getCanonicalInductionVariable())
1380 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1383 // Rewrite an AddRec in terms of the canonical induction variable, if
1384 // its type is more narrow.
1386 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1387 SE.getTypeSizeInBits(Ty)) {
1388 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1389 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1390 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1391 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1392 S->getNoWrapFlags(SCEV::FlagNW)));
1393 BasicBlock::iterator NewInsertPt =
1394 std::next(BasicBlock::iterator(cast<Instruction>(V)));
1395 BuilderType::InsertPointGuard Guard(Builder);
1396 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) ||
1397 isa<LandingPadInst>(NewInsertPt))
1399 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1404 // {X,+,F} --> X + {0,+,F}
1405 if (!S->getStart()->isZero()) {
1406 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1407 NewOps[0] = SE.getConstant(Ty, 0);
1408 const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1409 S->getNoWrapFlags(SCEV::FlagNW));
1411 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1412 // comments on expandAddToGEP for details.
1413 const SCEV *Base = S->getStart();
1414 const SCEV *RestArray[1] = { Rest };
1415 // Dig into the expression to find the pointer base for a GEP.
1416 ExposePointerBase(Base, RestArray[0], SE);
1417 // If we found a pointer, expand the AddRec with a GEP.
1418 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1419 // Make sure the Base isn't something exotic, such as a multiplied
1420 // or divided pointer value. In those cases, the result type isn't
1421 // actually a pointer type.
1422 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1423 Value *StartV = expand(Base);
1424 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1425 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1429 // Just do a normal add. Pre-expand the operands to suppress folding.
1430 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1431 SE.getUnknown(expand(Rest))));
1434 // If we don't yet have a canonical IV, create one.
1436 // Create and insert the PHI node for the induction variable in the
1438 BasicBlock *Header = L->getHeader();
1439 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1440 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1442 rememberInstruction(CanonicalIV);
1444 SmallSet<BasicBlock *, 4> PredSeen;
1445 Constant *One = ConstantInt::get(Ty, 1);
1446 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1447 BasicBlock *HP = *HPI;
1448 if (!PredSeen.insert(HP))
1451 if (L->contains(HP)) {
1452 // Insert a unit add instruction right before the terminator
1453 // corresponding to the back-edge.
1454 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1456 HP->getTerminator());
1457 Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1458 rememberInstruction(Add);
1459 CanonicalIV->addIncoming(Add, HP);
1461 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1466 // {0,+,1} --> Insert a canonical induction variable into the loop!
1467 if (S->isAffine() && S->getOperand(1)->isOne()) {
1468 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1469 "IVs with types different from the canonical IV should "
1470 "already have been handled!");
1474 // {0,+,F} --> {0,+,1} * F
1476 // If this is a simple linear addrec, emit it now as a special case.
1477 if (S->isAffine()) // {0,+,F} --> i*F
1479 expand(SE.getTruncateOrNoop(
1480 SE.getMulExpr(SE.getUnknown(CanonicalIV),
1481 SE.getNoopOrAnyExtend(S->getOperand(1),
1482 CanonicalIV->getType())),
1485 // If this is a chain of recurrences, turn it into a closed form, using the
1486 // folders, then expandCodeFor the closed form. This allows the folders to
1487 // simplify the expression without having to build a bunch of special code
1488 // into this folder.
1489 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1491 // Promote S up to the canonical IV type, if the cast is foldable.
1492 const SCEV *NewS = S;
1493 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1494 if (isa<SCEVAddRecExpr>(Ext))
1497 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1498 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1500 // Truncate the result down to the original type, if needed.
1501 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1505 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1506 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1507 Value *V = expandCodeFor(S->getOperand(),
1508 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1509 Value *I = Builder.CreateTrunc(V, Ty);
1510 rememberInstruction(I);
1514 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1515 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1516 Value *V = expandCodeFor(S->getOperand(),
1517 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1518 Value *I = Builder.CreateZExt(V, Ty);
1519 rememberInstruction(I);
1523 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1524 Type *Ty = SE.getEffectiveSCEVType(S->getType());
1525 Value *V = expandCodeFor(S->getOperand(),
1526 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1527 Value *I = Builder.CreateSExt(V, Ty);
1528 rememberInstruction(I);
1532 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1533 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1534 Type *Ty = LHS->getType();
1535 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1536 // In the case of mixed integer and pointer types, do the
1537 // rest of the comparisons as integer.
1538 if (S->getOperand(i)->getType() != Ty) {
1539 Ty = SE.getEffectiveSCEVType(Ty);
1540 LHS = InsertNoopCastOfTo(LHS, Ty);
1542 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1543 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1544 rememberInstruction(ICmp);
1545 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1546 rememberInstruction(Sel);
1549 // In the case of mixed integer and pointer types, cast the
1550 // final result back to the pointer type.
1551 if (LHS->getType() != S->getType())
1552 LHS = InsertNoopCastOfTo(LHS, S->getType());
1556 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1557 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1558 Type *Ty = LHS->getType();
1559 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1560 // In the case of mixed integer and pointer types, do the
1561 // rest of the comparisons as integer.
1562 if (S->getOperand(i)->getType() != Ty) {
1563 Ty = SE.getEffectiveSCEVType(Ty);
1564 LHS = InsertNoopCastOfTo(LHS, Ty);
1566 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1567 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1568 rememberInstruction(ICmp);
1569 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1570 rememberInstruction(Sel);
1573 // In the case of mixed integer and pointer types, cast the
1574 // final result back to the pointer type.
1575 if (LHS->getType() != S->getType())
1576 LHS = InsertNoopCastOfTo(LHS, S->getType());
1580 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1582 Builder.SetInsertPoint(IP->getParent(), IP);
1583 return expandCodeFor(SH, Ty);
1586 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1587 // Expand the code for this SCEV.
1588 Value *V = expand(SH);
1590 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1591 "non-trivial casts should be done with the SCEVs directly!");
1592 V = InsertNoopCastOfTo(V, Ty);
1597 Value *SCEVExpander::expand(const SCEV *S) {
1598 // Compute an insertion point for this SCEV object. Hoist the instructions
1599 // as far out in the loop nest as possible.
1600 Instruction *InsertPt = Builder.GetInsertPoint();
1601 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1602 L = L->getParentLoop())
1603 if (SE.isLoopInvariant(S, L)) {
1605 if (BasicBlock *Preheader = L->getLoopPreheader())
1606 InsertPt = Preheader->getTerminator();
1608 // LSR sets the insertion point for AddRec start/step values to the
1609 // block start to simplify value reuse, even though it's an invalid
1610 // position. SCEVExpander must correct for this in all cases.
1611 InsertPt = L->getHeader()->getFirstInsertionPt();
1614 // If the SCEV is computable at this level, insert it into the header
1615 // after the PHIs (and after any other instructions that we've inserted
1616 // there) so that it is guaranteed to dominate any user inside the loop.
1617 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
1618 InsertPt = L->getHeader()->getFirstInsertionPt();
1619 while (InsertPt != Builder.GetInsertPoint()
1620 && (isInsertedInstruction(InsertPt)
1621 || isa<DbgInfoIntrinsic>(InsertPt))) {
1622 InsertPt = std::next(BasicBlock::iterator(InsertPt));
1627 // Check to see if we already expanded this here.
1628 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator
1629 I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1630 if (I != InsertedExpressions.end())
1633 BuilderType::InsertPointGuard Guard(Builder);
1634 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1636 // Expand the expression into instructions.
1637 Value *V = visit(S);
1639 // Remember the expanded value for this SCEV at this location.
1641 // This is independent of PostIncLoops. The mapped value simply materializes
1642 // the expression at this insertion point. If the mapped value happened to be
1643 // a postinc expansion, it could be reused by a non-postinc user, but only if
1644 // its insertion point was already at the head of the loop.
1645 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1649 void SCEVExpander::rememberInstruction(Value *I) {
1650 if (!PostIncLoops.empty())
1651 InsertedPostIncValues.insert(I);
1653 InsertedValues.insert(I);
1656 /// getOrInsertCanonicalInductionVariable - This method returns the
1657 /// canonical induction variable of the specified type for the specified
1658 /// loop (inserting one if there is none). A canonical induction variable
1659 /// starts at zero and steps by one on each iteration.
1661 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1663 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1665 // Build a SCEV for {0,+,1}<L>.
1666 // Conservatively use FlagAnyWrap for now.
1667 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1668 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1670 // Emit code for it.
1671 BuilderType::InsertPointGuard Guard(Builder);
1672 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
1677 /// Sort values by integer width for replaceCongruentIVs.
1678 static bool width_descending(Value *lhs, Value *rhs) {
1679 // Put pointers at the back and make sure pointer < pointer = false.
1680 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy())
1681 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy();
1682 return rhs->getType()->getPrimitiveSizeInBits()
1683 < lhs->getType()->getPrimitiveSizeInBits();
1686 /// replaceCongruentIVs - Check for congruent phis in this loop header and
1687 /// replace them with their most canonical representative. Return the number of
1688 /// phis eliminated.
1690 /// This does not depend on any SCEVExpander state but should be used in
1691 /// the same context that SCEVExpander is used.
1692 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1693 SmallVectorImpl<WeakVH> &DeadInsts,
1694 const TargetTransformInfo *TTI) {
1695 // Find integer phis in order of increasing width.
1696 SmallVector<PHINode*, 8> Phis;
1697 for (BasicBlock::iterator I = L->getHeader()->begin();
1698 PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
1699 Phis.push_back(Phi);
1702 std::sort(Phis.begin(), Phis.end(), width_descending);
1704 unsigned NumElim = 0;
1705 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1706 // Process phis from wide to narrow. Mapping wide phis to the their truncation
1707 // so narrow phis can reuse them.
1708 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(),
1709 PEnd = Phis.end(); PIter != PEnd; ++PIter) {
1710 PHINode *Phi = *PIter;
1712 // Fold constant phis. They may be congruent to other constant phis and
1713 // would confuse the logic below that expects proper IVs.
1714 if (Value *V = Phi->hasConstantValue()) {
1715 Phi->replaceAllUsesWith(V);
1716 DeadInsts.push_back(Phi);
1718 DEBUG_WITH_TYPE(DebugType, dbgs()
1719 << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1723 if (!SE.isSCEVable(Phi->getType()))
1726 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1729 if (Phi->getType()->isIntegerTy() && TTI
1730 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1731 // This phi can be freely truncated to the narrowest phi type. Map the
1732 // truncated expression to it so it will be reused for narrow types.
1733 const SCEV *TruncExpr =
1734 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1735 ExprToIVMap[TruncExpr] = Phi;
1740 // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1742 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1745 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1746 Instruction *OrigInc =
1747 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1748 Instruction *IsomorphicInc =
1749 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1751 // If this phi has the same width but is more canonical, replace the
1752 // original with it. As part of the "more canonical" determination,
1753 // respect a prior decision to use an IV chain.
1754 if (OrigPhiRef->getType() == Phi->getType()
1755 && !(ChainedPhis.count(Phi)
1756 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L))
1757 && (ChainedPhis.count(Phi)
1758 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1759 std::swap(OrigPhiRef, Phi);
1760 std::swap(OrigInc, IsomorphicInc);
1762 // Replacing the congruent phi is sufficient because acyclic redundancy
1763 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1764 // that a phi is congruent, it's often the head of an IV user cycle that
1765 // is isomorphic with the original phi. It's worth eagerly cleaning up the
1766 // common case of a single IV increment so that DeleteDeadPHIs can remove
1767 // cycles that had postinc uses.
1768 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc),
1769 IsomorphicInc->getType());
1770 if (OrigInc != IsomorphicInc
1771 && TruncExpr == SE.getSCEV(IsomorphicInc)
1772 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc))
1773 || hoistIVInc(OrigInc, IsomorphicInc))) {
1774 DEBUG_WITH_TYPE(DebugType, dbgs()
1775 << "INDVARS: Eliminated congruent iv.inc: "
1776 << *IsomorphicInc << '\n');
1777 Value *NewInc = OrigInc;
1778 if (OrigInc->getType() != IsomorphicInc->getType()) {
1779 Instruction *IP = isa<PHINode>(OrigInc)
1780 ? (Instruction*)L->getHeader()->getFirstInsertionPt()
1781 : OrigInc->getNextNode();
1782 IRBuilder<> Builder(IP);
1783 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1785 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName);
1787 IsomorphicInc->replaceAllUsesWith(NewInc);
1788 DeadInsts.push_back(IsomorphicInc);
1791 DEBUG_WITH_TYPE(DebugType, dbgs()
1792 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1794 Value *NewIV = OrigPhiRef;
1795 if (OrigPhiRef->getType() != Phi->getType()) {
1796 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt());
1797 Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1798 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1800 Phi->replaceAllUsesWith(NewIV);
1801 DeadInsts.push_back(Phi);
1807 // Search for a SCEV subexpression that is not safe to expand. Any expression
1808 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
1809 // UDiv expressions. We don't know if the UDiv is derived from an IR divide
1810 // instruction, but the important thing is that we prove the denominator is
1811 // nonzero before expansion.
1813 // IVUsers already checks that IV-derived expressions are safe. So this check is
1814 // only needed when the expression includes some subexpression that is not IV
1817 // Currently, we only allow division by a nonzero constant here. If this is
1818 // inadequate, we could easily allow division by SCEVUnknown by using
1819 // ValueTracking to check isKnownNonZero().
1821 // We cannot generally expand recurrences unless the step dominates the loop
1822 // header. The expander handles the special case of affine recurrences by
1823 // scaling the recurrence outside the loop, but this technique isn't generally
1824 // applicable. Expanding a nested recurrence outside a loop requires computing
1825 // binomial coefficients. This could be done, but the recurrence has to be in a
1826 // perfectly reduced form, which can't be guaranteed.
1827 struct SCEVFindUnsafe {
1828 ScalarEvolution &SE;
1831 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
1833 bool follow(const SCEV *S) {
1834 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1835 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
1836 if (!SC || SC->getValue()->isZero()) {
1841 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1842 const SCEV *Step = AR->getStepRecurrence(SE);
1843 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
1850 bool isDone() const { return IsUnsafe; }
1855 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
1856 SCEVFindUnsafe Search(SE);
1857 visitAll(S, Search);
1858 return !Search.IsUnsafe;