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/Analysis/LoopInfo.h"
18 #include "llvm/IntrinsicInst.h"
19 #include "llvm/LLVMContext.h"
20 #include "llvm/Target/TargetData.h"
21 #include "llvm/ADT/STLExtras.h"
24 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
25 /// reusing an existing cast if a suitable one exists, moving an existing
26 /// cast if a suitable one exists but isn't in the right place, or
27 /// creating a new one.
28 Value *SCEVExpander::ReuseOrCreateCast(Value *V, const Type *Ty,
29 Instruction::CastOps Op,
30 BasicBlock::iterator IP) {
31 // Check to see if there is already a cast!
32 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
35 if (U->getType() == Ty)
36 if (CastInst *CI = dyn_cast<CastInst>(U))
37 if (CI->getOpcode() == Op) {
38 // If the cast isn't where we want it, fix it.
39 if (BasicBlock::iterator(CI) != IP) {
40 // Create a new cast, and leave the old cast in place in case
41 // it is being used as an insert point. Clear its operand
42 // so that it doesn't hold anything live.
43 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
45 CI->replaceAllUsesWith(NewCI);
46 CI->setOperand(0, UndefValue::get(V->getType()));
47 rememberInstruction(NewCI);
50 rememberInstruction(CI);
56 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
57 rememberInstruction(I);
61 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
62 /// which must be possible with a noop cast, doing what we can to share
64 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
65 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
66 assert((Op == Instruction::BitCast ||
67 Op == Instruction::PtrToInt ||
68 Op == Instruction::IntToPtr) &&
69 "InsertNoopCastOfTo cannot perform non-noop casts!");
70 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
71 "InsertNoopCastOfTo cannot change sizes!");
73 // Short-circuit unnecessary bitcasts.
74 if (Op == Instruction::BitCast && V->getType() == Ty)
77 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
78 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
79 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
80 if (CastInst *CI = dyn_cast<CastInst>(V))
81 if ((CI->getOpcode() == Instruction::PtrToInt ||
82 CI->getOpcode() == Instruction::IntToPtr) &&
83 SE.getTypeSizeInBits(CI->getType()) ==
84 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
85 return CI->getOperand(0);
86 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
87 if ((CE->getOpcode() == Instruction::PtrToInt ||
88 CE->getOpcode() == Instruction::IntToPtr) &&
89 SE.getTypeSizeInBits(CE->getType()) ==
90 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
91 return CE->getOperand(0);
94 // Fold a cast of a constant.
95 if (Constant *C = dyn_cast<Constant>(V))
96 return ConstantExpr::getCast(Op, C, Ty);
98 // Cast the argument at the beginning of the entry block, after
99 // any bitcasts of other arguments.
100 if (Argument *A = dyn_cast<Argument>(V)) {
101 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
102 while ((isa<BitCastInst>(IP) &&
103 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
104 cast<BitCastInst>(IP)->getOperand(0) != A) ||
105 isa<DbgInfoIntrinsic>(IP))
107 return ReuseOrCreateCast(A, Ty, Op, IP);
110 // Cast the instruction immediately after the instruction.
111 Instruction *I = cast<Instruction>(V);
112 BasicBlock::iterator IP = I; ++IP;
113 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
114 IP = II->getNormalDest()->begin();
115 while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP)) ++IP;
116 return ReuseOrCreateCast(I, Ty, Op, IP);
119 /// InsertBinop - Insert the specified binary operator, doing a small amount
120 /// of work to avoid inserting an obviously redundant operation.
121 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
122 Value *LHS, Value *RHS) {
123 // Fold a binop with constant operands.
124 if (Constant *CLHS = dyn_cast<Constant>(LHS))
125 if (Constant *CRHS = dyn_cast<Constant>(RHS))
126 return ConstantExpr::get(Opcode, CLHS, CRHS);
128 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
129 unsigned ScanLimit = 6;
130 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
131 // Scanning starts from the last instruction before the insertion point.
132 BasicBlock::iterator IP = Builder.GetInsertPoint();
133 if (IP != BlockBegin) {
135 for (; ScanLimit; --IP, --ScanLimit) {
136 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
138 if (isa<DbgInfoIntrinsic>(IP))
140 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
141 IP->getOperand(1) == RHS)
143 if (IP == BlockBegin) break;
147 // Save the original insertion point so we can restore it when we're done.
148 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
149 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
151 // Move the insertion point out of as many loops as we can.
152 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
153 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
154 BasicBlock *Preheader = L->getLoopPreheader();
155 if (!Preheader) break;
157 // Ok, move up a level.
158 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
161 // If we haven't found this binop, insert it.
162 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
163 rememberInstruction(BO);
165 // Restore the original insert point.
167 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
172 /// FactorOutConstant - Test if S is divisible by Factor, using signed
173 /// division. If so, update S with Factor divided out and return true.
174 /// S need not be evenly divisible if a reasonable remainder can be
176 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
177 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
178 /// check to see if the divide was folded.
179 static bool FactorOutConstant(const SCEV *&S,
180 const SCEV *&Remainder,
183 const TargetData *TD) {
184 // Everything is divisible by one.
190 S = SE.getConstant(S->getType(), 1);
194 // For a Constant, check for a multiple of the given factor.
195 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
199 // Check for divisibility.
200 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
202 ConstantInt::get(SE.getContext(),
203 C->getValue()->getValue().sdiv(
204 FC->getValue()->getValue()));
205 // If the quotient is zero and the remainder is non-zero, reject
206 // the value at this scale. It will be considered for subsequent
209 const SCEV *Div = SE.getConstant(CI);
212 SE.getAddExpr(Remainder,
213 SE.getConstant(C->getValue()->getValue().srem(
214 FC->getValue()->getValue())));
220 // In a Mul, check if there is a constant operand which is a multiple
221 // of the given factor.
222 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
224 // With TargetData, the size is known. Check if there is a constant
225 // operand which is a multiple of the given factor. If so, we can
227 const SCEVConstant *FC = cast<SCEVConstant>(Factor);
228 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
229 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
230 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
232 SE.getConstant(C->getValue()->getValue().sdiv(
233 FC->getValue()->getValue()));
234 S = SE.getMulExpr(NewMulOps);
238 // Without TargetData, check if Factor can be factored out of any of the
239 // Mul's operands. If so, we can just remove it.
240 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
241 const SCEV *SOp = M->getOperand(i);
242 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
243 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
244 Remainder->isZero()) {
245 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
247 S = SE.getMulExpr(NewMulOps);
254 // In an AddRec, check if both start and step are divisible.
255 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
256 const SCEV *Step = A->getStepRecurrence(SE);
257 const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
258 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
260 if (!StepRem->isZero())
262 const SCEV *Start = A->getStart();
263 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
265 S = SE.getAddRecExpr(Start, Step, A->getLoop());
272 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
273 /// is the number of SCEVAddRecExprs present, which are kept at the end of
276 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
278 ScalarEvolution &SE) {
279 unsigned NumAddRecs = 0;
280 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
282 // Group Ops into non-addrecs and addrecs.
283 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
284 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
285 // Let ScalarEvolution sort and simplify the non-addrecs list.
286 const SCEV *Sum = NoAddRecs.empty() ?
287 SE.getConstant(Ty, 0) :
288 SE.getAddExpr(NoAddRecs);
289 // If it returned an add, use the operands. Otherwise it simplified
290 // the sum into a single value, so just use that.
292 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
293 Ops.append(Add->op_begin(), Add->op_end());
294 else if (!Sum->isZero())
296 // Then append the addrecs.
297 Ops.append(AddRecs.begin(), AddRecs.end());
300 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
301 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
302 /// This helps expose more opportunities for folding parts of the expressions
303 /// into GEP indices.
305 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
307 ScalarEvolution &SE) {
309 SmallVector<const SCEV *, 8> AddRecs;
310 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
311 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
312 const SCEV *Start = A->getStart();
313 if (Start->isZero()) break;
314 const SCEV *Zero = SE.getConstant(Ty, 0);
315 AddRecs.push_back(SE.getAddRecExpr(Zero,
316 A->getStepRecurrence(SE),
318 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
320 Ops.append(Add->op_begin(), Add->op_end());
321 e += Add->getNumOperands();
326 if (!AddRecs.empty()) {
327 // Add the addrecs onto the end of the list.
328 Ops.append(AddRecs.begin(), AddRecs.end());
329 // Resort the operand list, moving any constants to the front.
330 SimplifyAddOperands(Ops, Ty, SE);
334 /// expandAddToGEP - Expand an addition expression with a pointer type into
335 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
336 /// BasicAliasAnalysis and other passes analyze the result. See the rules
337 /// for getelementptr vs. inttoptr in
338 /// http://llvm.org/docs/LangRef.html#pointeraliasing
341 /// Design note: The correctness of using getelementptr here depends on
342 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
343 /// they may introduce pointer arithmetic which may not be safely converted
344 /// into getelementptr.
346 /// Design note: It might seem desirable for this function to be more
347 /// loop-aware. If some of the indices are loop-invariant while others
348 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
349 /// loop-invariant portions of the overall computation outside the loop.
350 /// However, there are a few reasons this is not done here. Hoisting simple
351 /// arithmetic is a low-level optimization that often isn't very
352 /// important until late in the optimization process. In fact, passes
353 /// like InstructionCombining will combine GEPs, even if it means
354 /// pushing loop-invariant computation down into loops, so even if the
355 /// GEPs were split here, the work would quickly be undone. The
356 /// LoopStrengthReduction pass, which is usually run quite late (and
357 /// after the last InstructionCombining pass), takes care of hoisting
358 /// loop-invariant portions of expressions, after considering what
359 /// can be folded using target addressing modes.
361 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
362 const SCEV *const *op_end,
363 const PointerType *PTy,
366 const Type *ElTy = PTy->getElementType();
367 SmallVector<Value *, 4> GepIndices;
368 SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
369 bool AnyNonZeroIndices = false;
371 // Split AddRecs up into parts as either of the parts may be usable
372 // without the other.
373 SplitAddRecs(Ops, Ty, SE);
375 // Descend down the pointer's type and attempt to convert the other
376 // operands into GEP indices, at each level. The first index in a GEP
377 // indexes into the array implied by the pointer operand; the rest of
378 // the indices index into the element or field type selected by the
381 // If the scale size is not 0, attempt to factor out a scale for
383 SmallVector<const SCEV *, 8> ScaledOps;
384 if (ElTy->isSized()) {
385 const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
386 if (!ElSize->isZero()) {
387 SmallVector<const SCEV *, 8> NewOps;
388 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
389 const SCEV *Op = Ops[i];
390 const SCEV *Remainder = SE.getConstant(Ty, 0);
391 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
392 // Op now has ElSize factored out.
393 ScaledOps.push_back(Op);
394 if (!Remainder->isZero())
395 NewOps.push_back(Remainder);
396 AnyNonZeroIndices = true;
398 // The operand was not divisible, so add it to the list of operands
399 // we'll scan next iteration.
400 NewOps.push_back(Ops[i]);
403 // If we made any changes, update Ops.
404 if (!ScaledOps.empty()) {
406 SimplifyAddOperands(Ops, Ty, SE);
411 // Record the scaled array index for this level of the type. If
412 // we didn't find any operands that could be factored, tentatively
413 // assume that element zero was selected (since the zero offset
414 // would obviously be folded away).
415 Value *Scaled = ScaledOps.empty() ?
416 Constant::getNullValue(Ty) :
417 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
418 GepIndices.push_back(Scaled);
420 // Collect struct field index operands.
421 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
422 bool FoundFieldNo = false;
423 // An empty struct has no fields.
424 if (STy->getNumElements() == 0) break;
426 // With TargetData, field offsets are known. See if a constant offset
427 // falls within any of the struct fields.
428 if (Ops.empty()) break;
429 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
430 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
431 const StructLayout &SL = *SE.TD->getStructLayout(STy);
432 uint64_t FullOffset = C->getValue()->getZExtValue();
433 if (FullOffset < SL.getSizeInBytes()) {
434 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
435 GepIndices.push_back(
436 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
437 ElTy = STy->getTypeAtIndex(ElIdx);
439 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
440 AnyNonZeroIndices = true;
445 // Without TargetData, just check for an offsetof expression of the
446 // appropriate struct type.
447 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
448 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
451 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
452 GepIndices.push_back(FieldNo);
454 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
455 Ops[i] = SE.getConstant(Ty, 0);
456 AnyNonZeroIndices = true;
462 // If no struct field offsets were found, tentatively assume that
463 // field zero was selected (since the zero offset would obviously
466 ElTy = STy->getTypeAtIndex(0u);
467 GepIndices.push_back(
468 Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
472 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
473 ElTy = ATy->getElementType();
478 // If none of the operands were convertible to proper GEP indices, cast
479 // the base to i8* and do an ugly getelementptr with that. It's still
480 // better than ptrtoint+arithmetic+inttoptr at least.
481 if (!AnyNonZeroIndices) {
482 // Cast the base to i8*.
483 V = InsertNoopCastOfTo(V,
484 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
486 // Expand the operands for a plain byte offset.
487 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
489 // Fold a GEP with constant operands.
490 if (Constant *CLHS = dyn_cast<Constant>(V))
491 if (Constant *CRHS = dyn_cast<Constant>(Idx))
492 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
494 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
495 unsigned ScanLimit = 6;
496 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
497 // Scanning starts from the last instruction before the insertion point.
498 BasicBlock::iterator IP = Builder.GetInsertPoint();
499 if (IP != BlockBegin) {
501 for (; ScanLimit; --IP, --ScanLimit) {
502 // Don't count dbg.value against the ScanLimit, to avoid perturbing the
504 if (isa<DbgInfoIntrinsic>(IP))
506 if (IP->getOpcode() == Instruction::GetElementPtr &&
507 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
509 if (IP == BlockBegin) break;
513 // Save the original insertion point so we can restore it when we're done.
514 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
515 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
517 // Move the insertion point out of as many loops as we can.
518 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
519 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
520 BasicBlock *Preheader = L->getLoopPreheader();
521 if (!Preheader) break;
523 // Ok, move up a level.
524 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
528 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
529 rememberInstruction(GEP);
531 // Restore the original insert point.
533 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
538 // Save the original insertion point so we can restore it when we're done.
539 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
540 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
542 // Move the insertion point out of as many loops as we can.
543 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
544 if (!L->isLoopInvariant(V)) break;
546 bool AnyIndexNotLoopInvariant = false;
547 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
548 E = GepIndices.end(); I != E; ++I)
549 if (!L->isLoopInvariant(*I)) {
550 AnyIndexNotLoopInvariant = true;
553 if (AnyIndexNotLoopInvariant)
556 BasicBlock *Preheader = L->getLoopPreheader();
557 if (!Preheader) break;
559 // Ok, move up a level.
560 Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
563 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
564 // because ScalarEvolution may have changed the address arithmetic to
565 // compute a value which is beyond the end of the allocated object.
567 if (V->getType() != PTy)
568 Casted = InsertNoopCastOfTo(Casted, PTy);
569 Value *GEP = Builder.CreateGEP(Casted,
573 Ops.push_back(SE.getUnknown(GEP));
574 rememberInstruction(GEP);
576 // Restore the original insert point.
578 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
580 return expand(SE.getAddExpr(Ops));
583 /// isNonConstantNegative - Return true if the specified scev is negated, but
585 static bool isNonConstantNegative(const SCEV *F) {
586 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F);
587 if (!Mul) return false;
589 // If there is a constant factor, it will be first.
590 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
591 if (!SC) return false;
593 // Return true if the value is negative, this matches things like (-42 * V).
594 return SC->getValue()->getValue().isNegative();
597 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
598 /// SCEV expansion. If they are nested, this is the most nested. If they are
599 /// neighboring, pick the later.
600 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
604 if (A->contains(B)) return B;
605 if (B->contains(A)) return A;
606 if (DT.dominates(A->getHeader(), B->getHeader())) return B;
607 if (DT.dominates(B->getHeader(), A->getHeader())) return A;
608 return A; // Arbitrarily break the tie.
611 /// GetRelevantLoop - Get the most relevant loop associated with the given
612 /// expression, according to PickMostRelevantLoop.
613 static const Loop *GetRelevantLoop(const SCEV *S, LoopInfo &LI,
615 if (isa<SCEVConstant>(S))
617 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
618 if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
619 return LI.getLoopFor(I->getParent());
622 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
624 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
626 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
628 L = PickMostRelevantLoop(L, GetRelevantLoop(*I, LI, DT), DT);
631 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
632 return GetRelevantLoop(C->getOperand(), LI, DT);
633 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S))
634 return PickMostRelevantLoop(GetRelevantLoop(D->getLHS(), LI, DT),
635 GetRelevantLoop(D->getRHS(), LI, DT),
637 llvm_unreachable("Unexpected SCEV type!");
642 /// LoopCompare - Compare loops by PickMostRelevantLoop.
646 explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
648 bool operator()(std::pair<const Loop *, const SCEV *> LHS,
649 std::pair<const Loop *, const SCEV *> RHS) const {
650 // Compare loops with PickMostRelevantLoop.
651 if (LHS.first != RHS.first)
652 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
654 // If one operand is a non-constant negative and the other is not,
655 // put the non-constant negative on the right so that a sub can
656 // be used instead of a negate and add.
657 if (isNonConstantNegative(LHS.second)) {
658 if (!isNonConstantNegative(RHS.second))
660 } else if (isNonConstantNegative(RHS.second))
663 // Otherwise they are equivalent according to this comparison.
670 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
671 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
673 // Collect all the add operands in a loop, along with their associated loops.
674 // Iterate in reverse so that constants are emitted last, all else equal, and
675 // so that pointer operands are inserted first, which the code below relies on
676 // to form more involved GEPs.
677 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
678 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
679 E(S->op_begin()); I != E; ++I)
680 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
683 // Sort by loop. Use a stable sort so that constants follow non-constants and
684 // pointer operands precede non-pointer operands.
685 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
687 // Emit instructions to add all the operands. Hoist as much as possible
688 // out of loops, and form meaningful getelementptrs where possible.
690 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
691 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
692 const Loop *CurLoop = I->first;
693 const SCEV *Op = I->second;
695 // This is the first operand. Just expand it.
698 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
699 // The running sum expression is a pointer. Try to form a getelementptr
700 // at this level with that as the base.
701 SmallVector<const SCEV *, 4> NewOps;
702 for (; I != E && I->first == CurLoop; ++I)
703 NewOps.push_back(I->second);
704 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
705 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
706 // The running sum is an integer, and there's a pointer at this level.
707 // Try to form a getelementptr. If the running sum is instructions,
708 // use a SCEVUnknown to avoid re-analyzing them.
709 SmallVector<const SCEV *, 4> NewOps;
710 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
712 for (++I; I != E && I->first == CurLoop; ++I)
713 NewOps.push_back(I->second);
714 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
715 } else if (isNonConstantNegative(Op)) {
716 // Instead of doing a negate and add, just do a subtract.
717 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
718 Sum = InsertNoopCastOfTo(Sum, Ty);
719 Sum = InsertBinop(Instruction::Sub, Sum, W);
723 Value *W = expandCodeFor(Op, Ty);
724 Sum = InsertNoopCastOfTo(Sum, Ty);
725 // Canonicalize a constant to the RHS.
726 if (isa<Constant>(Sum)) std::swap(Sum, W);
727 Sum = InsertBinop(Instruction::Add, Sum, W);
735 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
736 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
738 // Collect all the mul operands in a loop, along with their associated loops.
739 // Iterate in reverse so that constants are emitted last, all else equal.
740 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
741 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
742 E(S->op_begin()); I != E; ++I)
743 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
746 // Sort by loop. Use a stable sort so that constants follow non-constants.
747 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
749 // Emit instructions to mul all the operands. Hoist as much as possible
752 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
753 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
754 const SCEV *Op = I->second;
756 // This is the first operand. Just expand it.
759 } else if (Op->isAllOnesValue()) {
760 // Instead of doing a multiply by negative one, just do a negate.
761 Prod = InsertNoopCastOfTo(Prod, Ty);
762 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
766 Value *W = expandCodeFor(Op, Ty);
767 Prod = InsertNoopCastOfTo(Prod, Ty);
768 // Canonicalize a constant to the RHS.
769 if (isa<Constant>(Prod)) std::swap(Prod, W);
770 Prod = InsertBinop(Instruction::Mul, Prod, W);
778 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
779 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
781 Value *LHS = expandCodeFor(S->getLHS(), Ty);
782 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
783 const APInt &RHS = SC->getValue()->getValue();
784 if (RHS.isPowerOf2())
785 return InsertBinop(Instruction::LShr, LHS,
786 ConstantInt::get(Ty, RHS.logBase2()));
789 Value *RHS = expandCodeFor(S->getRHS(), Ty);
790 return InsertBinop(Instruction::UDiv, LHS, RHS);
793 /// Move parts of Base into Rest to leave Base with the minimal
794 /// expression that provides a pointer operand suitable for a
796 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
797 ScalarEvolution &SE) {
798 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
799 Base = A->getStart();
800 Rest = SE.getAddExpr(Rest,
801 SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
802 A->getStepRecurrence(SE),
805 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
806 Base = A->getOperand(A->getNumOperands()-1);
807 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
808 NewAddOps.back() = Rest;
809 Rest = SE.getAddExpr(NewAddOps);
810 ExposePointerBase(Base, Rest, SE);
814 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
815 /// the base addrec, which is the addrec without any non-loop-dominating
816 /// values, and return the PHI.
818 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
820 const Type *ExpandTy,
822 // Reuse a previously-inserted PHI, if present.
823 for (BasicBlock::iterator I = L->getHeader()->begin();
824 PHINode *PN = dyn_cast<PHINode>(I); ++I)
825 if (SE.isSCEVable(PN->getType()) &&
826 (SE.getEffectiveSCEVType(PN->getType()) ==
827 SE.getEffectiveSCEVType(Normalized->getType())) &&
828 SE.getSCEV(PN) == Normalized)
829 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
831 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
833 // Determine if this is a well-behaved chain of instructions leading
834 // back to the PHI. It probably will be, if we're scanning an inner
835 // loop already visited by LSR for example, but it wouldn't have
838 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) {
842 // If any of the operands don't dominate the insert position, bail.
843 // Addrec operands are always loop-invariant, so this can only happen
844 // if there are instructions which haven't been hoisted.
845 for (User::op_iterator OI = IncV->op_begin()+1,
846 OE = IncV->op_end(); OI != OE; ++OI)
847 if (Instruction *OInst = dyn_cast<Instruction>(OI))
848 if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
854 // Advance to the next instruction.
855 IncV = dyn_cast<Instruction>(IncV->getOperand(0));
858 if (IncV->mayHaveSideEffects()) {
862 } while (IncV != PN);
865 // Ok, the add recurrence looks usable.
866 // Remember this PHI, even in post-inc mode.
867 InsertedValues.insert(PN);
868 // Remember the increment.
869 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
870 rememberInstruction(IncV);
871 if (L == IVIncInsertLoop)
873 if (SE.DT->dominates(IncV, IVIncInsertPos))
875 // Make sure the increment is where we want it. But don't move it
876 // down past a potential existing post-inc user.
877 IncV->moveBefore(IVIncInsertPos);
878 IVIncInsertPos = IncV;
879 IncV = cast<Instruction>(IncV->getOperand(0));
880 } while (IncV != PN);
885 // Save the original insertion point so we can restore it when we're done.
886 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
887 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
889 // Expand code for the start value.
890 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
891 L->getHeader()->begin());
893 // Expand code for the step value. Insert instructions right before the
894 // terminator corresponding to the back-edge. Do this before creating the PHI
895 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is
896 // negative, insert a sub instead of an add for the increment (unless it's a
897 // constant, because subtracts of constants are canonicalized to adds).
898 const SCEV *Step = Normalized->getStepRecurrence(SE);
899 bool isPointer = ExpandTy->isPointerTy();
900 bool isNegative = !isPointer && isNonConstantNegative(Step);
902 Step = SE.getNegativeSCEV(Step);
903 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin());
906 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin());
907 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv");
908 rememberInstruction(PN);
910 // Create the step instructions and populate the PHI.
911 BasicBlock *Header = L->getHeader();
912 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
914 BasicBlock *Pred = *HPI;
916 // Add a start value.
917 if (!L->contains(Pred)) {
918 PN->addIncoming(StartV, Pred);
922 // Create a step value and add it to the PHI. If IVIncInsertLoop is
923 // non-null and equal to the addrec's loop, insert the instructions
924 // at IVIncInsertPos.
925 Instruction *InsertPos = L == IVIncInsertLoop ?
926 IVIncInsertPos : Pred->getTerminator();
927 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos);
929 // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
931 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
932 // If the step isn't constant, don't use an implicitly scaled GEP, because
933 // that would require a multiply inside the loop.
934 if (!isa<ConstantInt>(StepV))
935 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
936 GEPPtrTy->getAddressSpace());
937 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) };
938 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN);
939 if (IncV->getType() != PN->getType()) {
940 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp");
941 rememberInstruction(IncV);
945 Builder.CreateSub(PN, StepV, "lsr.iv.next") :
946 Builder.CreateAdd(PN, StepV, "lsr.iv.next");
947 rememberInstruction(IncV);
949 PN->addIncoming(IncV, Pred);
952 // Restore the original insert point.
954 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
956 // Remember this PHI, even in post-inc mode.
957 InsertedValues.insert(PN);
962 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
963 const Type *STy = S->getType();
964 const Type *IntTy = SE.getEffectiveSCEVType(STy);
965 const Loop *L = S->getLoop();
967 // Determine a normalized form of this expression, which is the expression
968 // before any post-inc adjustment is made.
969 const SCEVAddRecExpr *Normalized = S;
970 if (PostIncLoops.count(L)) {
971 PostIncLoopSet Loops;
974 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
978 // Strip off any non-loop-dominating component from the addrec start.
979 const SCEV *Start = Normalized->getStart();
980 const SCEV *PostLoopOffset = 0;
981 if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
982 PostLoopOffset = Start;
983 Start = SE.getConstant(Normalized->getType(), 0);
985 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
986 Normalized->getStepRecurrence(SE),
987 Normalized->getLoop()));
990 // Strip off any non-loop-dominating component from the addrec step.
991 const SCEV *Step = Normalized->getStepRecurrence(SE);
992 const SCEV *PostLoopScale = 0;
993 if (!Step->dominates(L->getHeader(), SE.DT)) {
994 PostLoopScale = Step;
995 Step = SE.getConstant(Normalized->getType(), 1);
997 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
998 Normalized->getLoop()));
1001 // Expand the core addrec. If we need post-loop scaling, force it to
1002 // expand to an integer type to avoid the need for additional casting.
1003 const Type *ExpandTy = PostLoopScale ? IntTy : STy;
1004 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
1006 // Accommodate post-inc mode, if necessary.
1008 if (!PostIncLoops.count(L))
1011 // In PostInc mode, use the post-incremented value.
1012 BasicBlock *LatchBlock = L->getLoopLatch();
1013 assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1014 Result = PN->getIncomingValueForBlock(LatchBlock);
1017 // Re-apply any non-loop-dominating scale.
1018 if (PostLoopScale) {
1019 Result = InsertNoopCastOfTo(Result, IntTy);
1020 Result = Builder.CreateMul(Result,
1021 expandCodeFor(PostLoopScale, IntTy));
1022 rememberInstruction(Result);
1025 // Re-apply any non-loop-dominating offset.
1026 if (PostLoopOffset) {
1027 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1028 const SCEV *const OffsetArray[1] = { PostLoopOffset };
1029 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
1031 Result = InsertNoopCastOfTo(Result, IntTy);
1032 Result = Builder.CreateAdd(Result,
1033 expandCodeFor(PostLoopOffset, IntTy));
1034 rememberInstruction(Result);
1041 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1042 if (!CanonicalMode) return expandAddRecExprLiterally(S);
1044 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1045 const Loop *L = S->getLoop();
1047 // First check for an existing canonical IV in a suitable type.
1048 PHINode *CanonicalIV = 0;
1049 if (PHINode *PN = L->getCanonicalInductionVariable())
1050 if (SE.isSCEVable(PN->getType()) &&
1051 SE.getEffectiveSCEVType(PN->getType())->isIntegerTy() &&
1052 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1055 // Rewrite an AddRec in terms of the canonical induction variable, if
1056 // its type is more narrow.
1058 SE.getTypeSizeInBits(CanonicalIV->getType()) >
1059 SE.getTypeSizeInBits(Ty)) {
1060 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1061 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1062 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1063 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
1064 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1065 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1066 BasicBlock::iterator NewInsertPt =
1067 llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
1068 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt))
1070 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
1072 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1076 // {X,+,F} --> X + {0,+,F}
1077 if (!S->getStart()->isZero()) {
1078 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1079 NewOps[0] = SE.getConstant(Ty, 0);
1080 const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
1082 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1083 // comments on expandAddToGEP for details.
1084 const SCEV *Base = S->getStart();
1085 const SCEV *RestArray[1] = { Rest };
1086 // Dig into the expression to find the pointer base for a GEP.
1087 ExposePointerBase(Base, RestArray[0], SE);
1088 // If we found a pointer, expand the AddRec with a GEP.
1089 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1090 // Make sure the Base isn't something exotic, such as a multiplied
1091 // or divided pointer value. In those cases, the result type isn't
1092 // actually a pointer type.
1093 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1094 Value *StartV = expand(Base);
1095 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1096 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
1100 // Just do a normal add. Pre-expand the operands to suppress folding.
1101 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
1102 SE.getUnknown(expand(Rest))));
1105 // {0,+,1} --> Insert a canonical induction variable into the loop!
1106 if (S->isAffine() && S->getOperand(1)->isOne()) {
1107 // If there's a canonical IV, just use it.
1109 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1110 "IVs with types different from the canonical IV should "
1111 "already have been handled!");
1115 // Create and insert the PHI node for the induction variable in the
1117 BasicBlock *Header = L->getHeader();
1118 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
1119 rememberInstruction(PN);
1121 Constant *One = ConstantInt::get(Ty, 1);
1122 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
1123 HPI != HPE; ++HPI) {
1124 BasicBlock *HP = *HPI;
1125 if (L->contains(HP)) {
1126 // Insert a unit add instruction right before the terminator
1127 // corresponding to the back-edge.
1128 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
1129 HP->getTerminator());
1130 rememberInstruction(Add);
1131 PN->addIncoming(Add, HP);
1133 PN->addIncoming(Constant::getNullValue(Ty), HP);
1138 // {0,+,F} --> {0,+,1} * F
1139 // Get the canonical induction variable I for this loop.
1140 Value *I = CanonicalIV ?
1142 getOrInsertCanonicalInductionVariable(L, Ty);
1144 // If this is a simple linear addrec, emit it now as a special case.
1145 if (S->isAffine()) // {0,+,F} --> i*F
1147 expand(SE.getTruncateOrNoop(
1148 SE.getMulExpr(SE.getUnknown(I),
1149 SE.getNoopOrAnyExtend(S->getOperand(1),
1153 // If this is a chain of recurrences, turn it into a closed form, using the
1154 // folders, then expandCodeFor the closed form. This allows the folders to
1155 // simplify the expression without having to build a bunch of special code
1156 // into this folder.
1157 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
1159 // Promote S up to the canonical IV type, if the cast is foldable.
1160 const SCEV *NewS = S;
1161 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
1162 if (isa<SCEVAddRecExpr>(Ext))
1165 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1166 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1168 // Truncate the result down to the original type, if needed.
1169 const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1173 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1174 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1175 Value *V = expandCodeFor(S->getOperand(),
1176 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1177 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
1178 rememberInstruction(I);
1182 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1183 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1184 Value *V = expandCodeFor(S->getOperand(),
1185 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1186 Value *I = Builder.CreateZExt(V, Ty, "tmp");
1187 rememberInstruction(I);
1191 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1192 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
1193 Value *V = expandCodeFor(S->getOperand(),
1194 SE.getEffectiveSCEVType(S->getOperand()->getType()));
1195 Value *I = Builder.CreateSExt(V, Ty, "tmp");
1196 rememberInstruction(I);
1200 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1201 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1202 const Type *Ty = LHS->getType();
1203 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1204 // In the case of mixed integer and pointer types, do the
1205 // rest of the comparisons as integer.
1206 if (S->getOperand(i)->getType() != Ty) {
1207 Ty = SE.getEffectiveSCEVType(Ty);
1208 LHS = InsertNoopCastOfTo(LHS, Ty);
1210 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1211 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
1212 rememberInstruction(ICmp);
1213 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1214 rememberInstruction(Sel);
1217 // In the case of mixed integer and pointer types, cast the
1218 // final result back to the pointer type.
1219 if (LHS->getType() != S->getType())
1220 LHS = InsertNoopCastOfTo(LHS, S->getType());
1224 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1225 Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1226 const Type *Ty = LHS->getType();
1227 for (int i = S->getNumOperands()-2; i >= 0; --i) {
1228 // In the case of mixed integer and pointer types, do the
1229 // rest of the comparisons as integer.
1230 if (S->getOperand(i)->getType() != Ty) {
1231 Ty = SE.getEffectiveSCEVType(Ty);
1232 LHS = InsertNoopCastOfTo(LHS, Ty);
1234 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1235 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
1236 rememberInstruction(ICmp);
1237 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1238 rememberInstruction(Sel);
1241 // In the case of mixed integer and pointer types, cast the
1242 // final result back to the pointer type.
1243 if (LHS->getType() != S->getType())
1244 LHS = InsertNoopCastOfTo(LHS, S->getType());
1248 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty,
1250 BasicBlock::iterator IP = I;
1251 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
1253 Builder.SetInsertPoint(IP->getParent(), IP);
1254 return expandCodeFor(SH, Ty);
1257 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
1258 // Expand the code for this SCEV.
1259 Value *V = expand(SH);
1261 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1262 "non-trivial casts should be done with the SCEVs directly!");
1263 V = InsertNoopCastOfTo(V, Ty);
1268 Value *SCEVExpander::expand(const SCEV *S) {
1269 // Compute an insertion point for this SCEV object. Hoist the instructions
1270 // as far out in the loop nest as possible.
1271 Instruction *InsertPt = Builder.GetInsertPoint();
1272 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
1273 L = L->getParentLoop())
1274 if (S->isLoopInvariant(L)) {
1276 if (BasicBlock *Preheader = L->getLoopPreheader())
1277 InsertPt = Preheader->getTerminator();
1279 // If the SCEV is computable at this level, insert it into the header
1280 // after the PHIs (and after any other instructions that we've inserted
1281 // there) so that it is guaranteed to dominate any user inside the loop.
1282 if (L && S->hasComputableLoopEvolution(L) && !PostIncLoops.count(L))
1283 InsertPt = L->getHeader()->getFirstNonPHI();
1284 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
1285 InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
1289 // Check to see if we already expanded this here.
1290 std::map<std::pair<const SCEV *, Instruction *>,
1291 AssertingVH<Value> >::iterator I =
1292 InsertedExpressions.find(std::make_pair(S, InsertPt));
1293 if (I != InsertedExpressions.end())
1296 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1297 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1298 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
1300 // Expand the expression into instructions.
1301 Value *V = visit(S);
1303 // Remember the expanded value for this SCEV at this location.
1304 if (PostIncLoops.empty())
1305 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1307 restoreInsertPoint(SaveInsertBB, SaveInsertPt);
1311 void SCEVExpander::rememberInstruction(Value *I) {
1312 if (!PostIncLoops.empty())
1313 InsertedPostIncValues.insert(I);
1315 InsertedValues.insert(I);
1317 // If we just claimed an existing instruction and that instruction had
1318 // been the insert point, adjust the insert point forward so that
1319 // subsequently inserted code will be dominated.
1320 if (Builder.GetInsertPoint() == I) {
1321 BasicBlock::iterator It = cast<Instruction>(I);
1322 do { ++It; } while (isInsertedInstruction(It) ||
1323 isa<DbgInfoIntrinsic>(It));
1324 Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
1328 void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
1329 // If we acquired more instructions since the old insert point was saved,
1330 // advance past them.
1331 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
1333 Builder.SetInsertPoint(BB, I);
1336 /// getOrInsertCanonicalInductionVariable - This method returns the
1337 /// canonical induction variable of the specified type for the specified
1338 /// loop (inserting one if there is none). A canonical induction variable
1339 /// starts at zero and steps by one on each iteration.
1341 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1343 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1344 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1345 SE.getConstant(Ty, 1), L);
1346 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
1347 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
1348 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
1350 restoreInsertPoint(SaveInsertBB, SaveInsertPt);