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/Target/TargetData.h"
19 #include "llvm/ADT/STLExtras.h"
22 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
23 /// which must be possible with a noop cast, doing what we can to share
25 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
26 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
27 assert((Op == Instruction::BitCast ||
28 Op == Instruction::PtrToInt ||
29 Op == Instruction::IntToPtr) &&
30 "InsertNoopCastOfTo cannot perform non-noop casts!");
31 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
32 "InsertNoopCastOfTo cannot change sizes!");
34 // Short-circuit unnecessary bitcasts.
35 if (Op == Instruction::BitCast && V->getType() == Ty)
38 // Short-circuit unnecessary inttoptr<->ptrtoint casts.
39 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
40 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
41 if (CastInst *CI = dyn_cast<CastInst>(V))
42 if ((CI->getOpcode() == Instruction::PtrToInt ||
43 CI->getOpcode() == Instruction::IntToPtr) &&
44 SE.getTypeSizeInBits(CI->getType()) ==
45 SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
46 return CI->getOperand(0);
47 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
48 if ((CE->getOpcode() == Instruction::PtrToInt ||
49 CE->getOpcode() == Instruction::IntToPtr) &&
50 SE.getTypeSizeInBits(CE->getType()) ==
51 SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
52 return CE->getOperand(0);
55 // FIXME: keep track of the cast instruction.
56 if (Constant *C = dyn_cast<Constant>(V))
57 return ConstantExpr::getCast(Op, C, Ty);
59 if (Argument *A = dyn_cast<Argument>(V)) {
60 // Check to see if there is already a cast!
61 for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
63 if ((*UI)->getType() == Ty)
64 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
65 if (CI->getOpcode() == Op) {
66 // If the cast isn't the first instruction of the function, move it.
67 if (BasicBlock::iterator(CI) !=
68 A->getParent()->getEntryBlock().begin()) {
69 // Recreate the cast at the beginning of the entry block.
70 // The old cast is left in place in case it is being used
71 // as an insert point.
73 CastInst::Create(Op, V, Ty, "",
74 A->getParent()->getEntryBlock().begin());
76 CI->replaceAllUsesWith(NewCI);
82 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
83 A->getParent()->getEntryBlock().begin());
84 InsertedValues.insert(I);
88 Instruction *I = cast<Instruction>(V);
90 // Check to see if there is already a cast. If there is, use it.
91 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
93 if ((*UI)->getType() == Ty)
94 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
95 if (CI->getOpcode() == Op) {
96 BasicBlock::iterator It = I; ++It;
97 if (isa<InvokeInst>(I))
98 It = cast<InvokeInst>(I)->getNormalDest()->begin();
99 while (isa<PHINode>(It)) ++It;
100 if (It != BasicBlock::iterator(CI)) {
101 // Recreate the cast at the beginning of the entry block.
102 // The old cast is left in place in case it is being used
103 // as an insert point.
104 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
106 CI->replaceAllUsesWith(NewCI);
112 BasicBlock::iterator IP = I; ++IP;
113 if (InvokeInst *II = dyn_cast<InvokeInst>(I))
114 IP = II->getNormalDest()->begin();
115 while (isa<PHINode>(IP)) ++IP;
116 Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
117 InsertedValues.insert(CI);
121 /// InsertBinop - Insert the specified binary operator, doing a small amount
122 /// of work to avoid inserting an obviously redundant operation.
123 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
124 Value *LHS, Value *RHS) {
125 // Fold a binop with constant operands.
126 if (Constant *CLHS = dyn_cast<Constant>(LHS))
127 if (Constant *CRHS = dyn_cast<Constant>(RHS))
128 return ConstantExpr::get(Opcode, CLHS, CRHS);
130 // Do a quick scan to see if we have this binop nearby. If so, reuse it.
131 unsigned ScanLimit = 6;
132 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
133 // Scanning starts from the last instruction before the insertion point.
134 BasicBlock::iterator IP = Builder.GetInsertPoint();
135 if (IP != BlockBegin) {
137 for (; ScanLimit; --IP, --ScanLimit) {
138 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
139 IP->getOperand(1) == RHS)
141 if (IP == BlockBegin) break;
145 // If we haven't found this binop, insert it.
146 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
147 InsertedValues.insert(BO);
151 /// FactorOutConstant - Test if S is divisible by Factor, using signed
152 /// division. If so, update S with Factor divided out and return true.
153 /// S need not be evenly divisble if a reasonable remainder can be
155 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
156 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
157 /// check to see if the divide was folded.
158 static bool FactorOutConstant(const SCEV* &S,
159 const SCEV* &Remainder,
161 ScalarEvolution &SE) {
162 // Everything is divisible by one.
166 // For a Constant, check for a multiple of the given factor.
167 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
169 ConstantInt::get(C->getValue()->getValue().sdiv(Factor));
170 // If the quotient is zero and the remainder is non-zero, reject
171 // the value at this scale. It will be considered for subsequent
173 if (C->isZero() || !CI->isZero()) {
174 const SCEV* Div = SE.getConstant(CI);
177 SE.getAddExpr(Remainder,
178 SE.getConstant(C->getValue()->getValue().srem(Factor)));
183 // In a Mul, check if there is a constant operand which is a multiple
184 // of the given factor.
185 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S))
186 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
187 if (!C->getValue()->getValue().srem(Factor)) {
188 const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
189 SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
192 SE.getConstant(C->getValue()->getValue().sdiv(Factor));
193 S = SE.getMulExpr(NewMulOps);
197 // In an AddRec, check if both start and step are divisible.
198 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
199 const SCEV* Step = A->getStepRecurrence(SE);
200 const SCEV* StepRem = SE.getIntegerSCEV(0, Step->getType());
201 if (!FactorOutConstant(Step, StepRem, Factor, SE))
203 if (!StepRem->isZero())
205 const SCEV* Start = A->getStart();
206 if (!FactorOutConstant(Start, Remainder, Factor, SE))
208 S = SE.getAddRecExpr(Start, Step, A->getLoop());
215 /// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
216 /// instead of using ptrtoint+arithmetic+inttoptr. This helps
217 /// BasicAliasAnalysis analyze the result. However, it suffers from the
218 /// underlying bug described in PR2831. Addition in LLVM currently always
219 /// has two's complement wrapping guaranteed. However, the semantics for
220 /// getelementptr overflow are ambiguous. In the common case though, this
221 /// expansion gets used when a GEP in the original code has been converted
222 /// into integer arithmetic, in which case the resulting code will be no
223 /// more undefined than it was originally.
225 /// Design note: It might seem desirable for this function to be more
226 /// loop-aware. If some of the indices are loop-invariant while others
227 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
228 /// loop-invariant portions of the overall computation outside the loop.
229 /// However, there are a few reasons this is not done here. Hoisting simple
230 /// arithmetic is a low-level optimization that often isn't very
231 /// important until late in the optimization process. In fact, passes
232 /// like InstructionCombining will combine GEPs, even if it means
233 /// pushing loop-invariant computation down into loops, so even if the
234 /// GEPs were split here, the work would quickly be undone. The
235 /// LoopStrengthReduction pass, which is usually run quite late (and
236 /// after the last InstructionCombining pass), takes care of hoisting
237 /// loop-invariant portions of expressions, after considering what
238 /// can be folded using target addressing modes.
240 Value *SCEVExpander::expandAddToGEP(const SCEV* const *op_begin,
241 const SCEV* const *op_end,
242 const PointerType *PTy,
245 const Type *ElTy = PTy->getElementType();
246 SmallVector<Value *, 4> GepIndices;
247 SmallVector<const SCEV*, 8> Ops(op_begin, op_end);
248 bool AnyNonZeroIndices = false;
250 // Decend down the pointer's type and attempt to convert the other
251 // operands into GEP indices, at each level. The first index in a GEP
252 // indexes into the array implied by the pointer operand; the rest of
253 // the indices index into the element or field type selected by the
256 APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
257 ElTy->isSized() ? SE.TD->getTypeAllocSize(ElTy) : 0);
258 SmallVector<const SCEV*, 8> NewOps;
259 SmallVector<const SCEV*, 8> ScaledOps;
260 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
261 // Split AddRecs up into parts as either of the parts may be usable
262 // without the other.
263 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i]))
264 if (!A->getStart()->isZero()) {
265 const SCEV* Start = A->getStart();
266 Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
267 A->getStepRecurrence(SE),
272 // If the scale size is not 0, attempt to factor out a scale.
274 const SCEV* Op = Ops[i];
275 const SCEV* Remainder = SE.getIntegerSCEV(0, Op->getType());
276 if (FactorOutConstant(Op, Remainder, ElSize, SE)) {
277 ScaledOps.push_back(Op); // Op now has ElSize factored out.
278 NewOps.push_back(Remainder);
282 // If the operand was not divisible, add it to the list of operands
283 // we'll scan next iteration.
284 NewOps.push_back(Ops[i]);
287 AnyNonZeroIndices |= !ScaledOps.empty();
288 Value *Scaled = ScaledOps.empty() ?
289 Constant::getNullValue(Ty) :
290 expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
291 GepIndices.push_back(Scaled);
293 // Collect struct field index operands.
295 while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
296 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
297 if (SE.getTypeSizeInBits(C->getType()) <= 64) {
298 const StructLayout &SL = *SE.TD->getStructLayout(STy);
299 uint64_t FullOffset = C->getValue()->getZExtValue();
300 if (FullOffset < SL.getSizeInBytes()) {
301 unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
302 GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx));
303 ElTy = STy->getTypeAtIndex(ElIdx);
305 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
306 AnyNonZeroIndices = true;
313 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
314 ElTy = ATy->getElementType();
320 // If none of the operands were convertable to proper GEP indices, cast
321 // the base to i8* and do an ugly getelementptr with that. It's still
322 // better than ptrtoint+arithmetic+inttoptr at least.
323 if (!AnyNonZeroIndices) {
324 V = InsertNoopCastOfTo(V,
325 Type::Int8Ty->getPointerTo(PTy->getAddressSpace()));
326 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
328 // Fold a GEP with constant operands.
329 if (Constant *CLHS = dyn_cast<Constant>(V))
330 if (Constant *CRHS = dyn_cast<Constant>(Idx))
331 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
333 // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
334 unsigned ScanLimit = 6;
335 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
336 // Scanning starts from the last instruction before the insertion point.
337 BasicBlock::iterator IP = Builder.GetInsertPoint();
338 if (IP != BlockBegin) {
340 for (; ScanLimit; --IP, --ScanLimit) {
341 if (IP->getOpcode() == Instruction::GetElementPtr &&
342 IP->getOperand(0) == V && IP->getOperand(1) == Idx)
344 if (IP == BlockBegin) break;
348 Value *GEP = Builder.CreateGEP(V, Idx, "scevgep");
349 InsertedValues.insert(GEP);
353 // Insert a pretty getelementptr.
354 Value *GEP = Builder.CreateGEP(V,
358 Ops.push_back(SE.getUnknown(GEP));
359 InsertedValues.insert(GEP);
360 return expand(SE.getAddExpr(Ops));
363 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
364 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
365 Value *V = expand(S->getOperand(S->getNumOperands()-1));
367 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
368 // comments on expandAddToGEP for details.
370 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
371 const SmallVectorImpl<const SCEV*> &Ops = S->getOperands();
372 return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1], PTy, Ty, V);
375 V = InsertNoopCastOfTo(V, Ty);
377 // Emit a bunch of add instructions
378 for (int i = S->getNumOperands()-2; i >= 0; --i) {
379 Value *W = expandCodeFor(S->getOperand(i), Ty);
380 V = InsertBinop(Instruction::Add, V, W);
385 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
386 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
387 int FirstOp = 0; // Set if we should emit a subtract.
388 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
389 if (SC->getValue()->isAllOnesValue())
392 int i = S->getNumOperands()-2;
393 Value *V = expandCodeFor(S->getOperand(i+1), Ty);
395 // Emit a bunch of multiply instructions
396 for (; i >= FirstOp; --i) {
397 Value *W = expandCodeFor(S->getOperand(i), Ty);
398 V = InsertBinop(Instruction::Mul, V, W);
401 // -1 * ... ---> 0 - ...
403 V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
407 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
408 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
410 Value *LHS = expandCodeFor(S->getLHS(), Ty);
411 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
412 const APInt &RHS = SC->getValue()->getValue();
413 if (RHS.isPowerOf2())
414 return InsertBinop(Instruction::LShr, LHS,
415 ConstantInt::get(Ty, RHS.logBase2()));
418 Value *RHS = expandCodeFor(S->getRHS(), Ty);
419 return InsertBinop(Instruction::UDiv, LHS, RHS);
422 /// Move parts of Base into Rest to leave Base with the minimal
423 /// expression that provides a pointer operand suitable for a
425 static void ExposePointerBase(const SCEV* &Base, const SCEV* &Rest,
426 ScalarEvolution &SE) {
427 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
428 Base = A->getStart();
429 Rest = SE.getAddExpr(Rest,
430 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
431 A->getStepRecurrence(SE),
434 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
435 Base = A->getOperand(A->getNumOperands()-1);
436 SmallVector<const SCEV*, 8> NewAddOps(A->op_begin(), A->op_end());
437 NewAddOps.back() = Rest;
438 Rest = SE.getAddExpr(NewAddOps);
439 ExposePointerBase(Base, Rest, SE);
443 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
444 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
445 const Loop *L = S->getLoop();
447 // First check for an existing canonical IV in a suitable type.
448 PHINode *CanonicalIV = 0;
449 if (PHINode *PN = L->getCanonicalInductionVariable())
450 if (SE.isSCEVable(PN->getType()) &&
451 isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
452 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
455 // Rewrite an AddRec in terms of the canonical induction variable, if
456 // its type is more narrow.
458 SE.getTypeSizeInBits(CanonicalIV->getType()) >
459 SE.getTypeSizeInBits(Ty)) {
460 const SCEV *Start = SE.getAnyExtendExpr(S->getStart(),
461 CanonicalIV->getType());
462 const SCEV *Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
463 CanonicalIV->getType());
464 Value *V = expand(SE.getAddRecExpr(Start, Step, S->getLoop()));
465 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
466 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
467 BasicBlock::iterator NewInsertPt =
468 next(BasicBlock::iterator(cast<Instruction>(V)));
469 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
470 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
472 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
476 // {X,+,F} --> X + {0,+,F}
477 if (!S->getStart()->isZero()) {
478 const SmallVectorImpl<const SCEV*> &SOperands = S->getOperands();
479 SmallVector<const SCEV*, 4> NewOps(SOperands.begin(), SOperands.end());
480 NewOps[0] = SE.getIntegerSCEV(0, Ty);
481 const SCEV* Rest = SE.getAddRecExpr(NewOps, L);
483 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
484 // comments on expandAddToGEP for details.
486 const SCEV* Base = S->getStart();
487 const SCEV* RestArray[1] = { Rest };
488 // Dig into the expression to find the pointer base for a GEP.
489 ExposePointerBase(Base, RestArray[0], SE);
490 // If we found a pointer, expand the AddRec with a GEP.
491 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
492 // Make sure the Base isn't something exotic, such as a multiplied
493 // or divided pointer value. In those cases, the result type isn't
494 // actually a pointer type.
495 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
496 Value *StartV = expand(Base);
497 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
498 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
503 // Just do a normal add. Pre-expand the operands to suppress folding.
504 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
505 SE.getUnknown(expand(Rest))));
508 // {0,+,1} --> Insert a canonical induction variable into the loop!
510 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
511 // If there's a canonical IV, just use it.
513 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
514 "IVs with types different from the canonical IV should "
515 "already have been handled!");
519 // Create and insert the PHI node for the induction variable in the
521 BasicBlock *Header = L->getHeader();
522 BasicBlock *Preheader = L->getLoopPreheader();
523 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
524 InsertedValues.insert(PN);
525 PN->addIncoming(Constant::getNullValue(Ty), Preheader);
527 pred_iterator HPI = pred_begin(Header);
528 assert(HPI != pred_end(Header) && "Loop with zero preds???");
529 if (!L->contains(*HPI)) ++HPI;
530 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
531 "No backedge in loop?");
533 // Insert a unit add instruction right before the terminator corresponding
535 Constant *One = ConstantInt::get(Ty, 1);
536 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
537 (*HPI)->getTerminator());
538 InsertedValues.insert(Add);
540 pred_iterator PI = pred_begin(Header);
541 if (*PI == Preheader)
543 PN->addIncoming(Add, *PI);
547 // {0,+,F} --> {0,+,1} * F
548 // Get the canonical induction variable I for this loop.
549 Value *I = CanonicalIV ?
551 getOrInsertCanonicalInductionVariable(L, Ty);
553 // If this is a simple linear addrec, emit it now as a special case.
554 if (S->isAffine()) // {0,+,F} --> i*F
556 expand(SE.getTruncateOrNoop(
557 SE.getMulExpr(SE.getUnknown(I),
558 SE.getNoopOrAnyExtend(S->getOperand(1),
562 // If this is a chain of recurrences, turn it into a closed form, using the
563 // folders, then expandCodeFor the closed form. This allows the folders to
564 // simplify the expression without having to build a bunch of special code
566 const SCEV* IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
568 // Promote S up to the canonical IV type, if the cast is foldable.
569 const SCEV* NewS = S;
570 const SCEV* Ext = SE.getNoopOrAnyExtend(S, I->getType());
571 if (isa<SCEVAddRecExpr>(Ext))
574 const SCEV* V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
575 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
577 // Truncate the result down to the original type, if needed.
578 const SCEV* T = SE.getTruncateOrNoop(V, Ty);
582 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
583 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
584 Value *V = expandCodeFor(S->getOperand(),
585 SE.getEffectiveSCEVType(S->getOperand()->getType()));
586 Value *I = Builder.CreateTrunc(V, Ty, "tmp");
587 InsertedValues.insert(I);
591 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
592 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
593 Value *V = expandCodeFor(S->getOperand(),
594 SE.getEffectiveSCEVType(S->getOperand()->getType()));
595 Value *I = Builder.CreateZExt(V, Ty, "tmp");
596 InsertedValues.insert(I);
600 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
601 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
602 Value *V = expandCodeFor(S->getOperand(),
603 SE.getEffectiveSCEVType(S->getOperand()->getType()));
604 Value *I = Builder.CreateSExt(V, Ty, "tmp");
605 InsertedValues.insert(I);
609 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
610 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
611 Value *LHS = expandCodeFor(S->getOperand(0), Ty);
612 for (unsigned i = 1; i < S->getNumOperands(); ++i) {
613 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
614 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
615 InsertedValues.insert(ICmp);
616 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
617 InsertedValues.insert(Sel);
623 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
624 const Type *Ty = SE.getEffectiveSCEVType(S->getType());
625 Value *LHS = expandCodeFor(S->getOperand(0), Ty);
626 for (unsigned i = 1; i < S->getNumOperands(); ++i) {
627 Value *RHS = expandCodeFor(S->getOperand(i), Ty);
628 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
629 InsertedValues.insert(ICmp);
630 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
631 InsertedValues.insert(Sel);
637 Value *SCEVExpander::expandCodeFor(const SCEV* SH, const Type *Ty) {
638 // Expand the code for this SCEV.
639 Value *V = expand(SH);
641 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
642 "non-trivial casts should be done with the SCEVs directly!");
643 V = InsertNoopCastOfTo(V, Ty);
648 Value *SCEVExpander::expand(const SCEV *S) {
649 // Compute an insertion point for this SCEV object. Hoist the instructions
650 // as far out in the loop nest as possible.
651 Instruction *InsertPt = Builder.GetInsertPoint();
652 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
653 L = L->getParentLoop())
654 if (S->isLoopInvariant(L)) {
656 if (BasicBlock *Preheader = L->getLoopPreheader())
657 InsertPt = Preheader->getTerminator();
659 // If the SCEV is computable at this level, insert it into the header
660 // after the PHIs (and after any other instructions that we've inserted
661 // there) so that it is guaranteed to dominate any user inside the loop.
662 if (L && S->hasComputableLoopEvolution(L))
663 InsertPt = L->getHeader()->getFirstNonPHI();
664 while (isInsertedInstruction(InsertPt))
665 InsertPt = next(BasicBlock::iterator(InsertPt));
669 // Check to see if we already expanded this here.
670 std::map<std::pair<const SCEV *, Instruction *>,
671 AssertingVH<Value> >::iterator I =
672 InsertedExpressions.find(std::make_pair(S, InsertPt));
673 if (I != InsertedExpressions.end())
676 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
677 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
678 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
680 // Expand the expression into instructions.
683 // Remember the expanded value for this SCEV at this location.
684 InsertedExpressions[std::make_pair(S, InsertPt)] = V;
686 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
690 /// getOrInsertCanonicalInductionVariable - This method returns the
691 /// canonical induction variable of the specified type for the specified
692 /// loop (inserting one if there is none). A canonical induction variable
693 /// starts at zero and steps by one on each iteration.
695 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
697 assert(Ty->isInteger() && "Can only insert integer induction variables!");
698 const SCEV* H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
699 SE.getIntegerSCEV(1, Ty), L);
700 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
701 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
702 Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
704 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);