ScalarEvolutionExpander.cpp revision 13c5e35222afe0895f0c5e68aa9f22f134ea437a 
1//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
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
12// expression.
13//
14//===----------------------------------------------------------------------===//
15
16#include "llvm/Analysis/ScalarEvolutionExpander.h"
17#include "llvm/Analysis/LoopInfo.h"
18#include "llvm/LLVMContext.h"
19#include "llvm/Target/TargetData.h"
20#include "llvm/ADT/STLExtras.h"
21using namespace llvm;
22
23/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
24/// which must be possible with a noop cast, doing what we can to share
25/// the casts.
26Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
27  Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
28  assert((Op == Instruction::BitCast ||
29          Op == Instruction::PtrToInt ||
30          Op == Instruction::IntToPtr) &&
31         "InsertNoopCastOfTo cannot perform non-noop casts!");
32  assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
33         "InsertNoopCastOfTo cannot change sizes!");
34
35  // Short-circuit unnecessary bitcasts.
36  if (Op == Instruction::BitCast && V->getType() == Ty)
37    return V;
38
39  // Short-circuit unnecessary inttoptr<->ptrtoint casts.
40  if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
41      SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
42    if (CastInst *CI = dyn_cast<CastInst>(V))
43      if ((CI->getOpcode() == Instruction::PtrToInt ||
44           CI->getOpcode() == Instruction::IntToPtr) &&
45          SE.getTypeSizeInBits(CI->getType()) ==
46          SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
47        return CI->getOperand(0);
48    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
49      if ((CE->getOpcode() == Instruction::PtrToInt ||
50           CE->getOpcode() == Instruction::IntToPtr) &&
51          SE.getTypeSizeInBits(CE->getType()) ==
52          SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
53        return CE->getOperand(0);
54  }
55
56  // FIXME: keep track of the cast instruction.
57  if (Constant *C = dyn_cast<Constant>(V))
58    return getContext()->getConstantExprCast(Op, C, Ty);
59
60  if (Argument *A = dyn_cast<Argument>(V)) {
61    // Check to see if there is already a cast!
62    for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
63         UI != E; ++UI)
64      if ((*UI)->getType() == Ty)
65        if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
66          if (CI->getOpcode() == Op) {
67            // If the cast isn't the first instruction of the function, move it.
68            if (BasicBlock::iterator(CI) !=
69                A->getParent()->getEntryBlock().begin()) {
70              // Recreate the cast at the beginning of the entry block.
71              // The old cast is left in place in case it is being used
72              // as an insert point.
73              Instruction *NewCI =
74                CastInst::Create(Op, V, Ty, "",
75                                 A->getParent()->getEntryBlock().begin());
76              NewCI->takeName(CI);
77              CI->replaceAllUsesWith(NewCI);
78              return NewCI;
79            }
80            return CI;
81          }
82
83    Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
84                                      A->getParent()->getEntryBlock().begin());
85    InsertedValues.insert(I);
86    return I;
87  }
88
89  Instruction *I = cast<Instruction>(V);
90
91  // Check to see if there is already a cast.  If there is, use it.
92  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
93       UI != E; ++UI) {
94    if ((*UI)->getType() == Ty)
95      if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
96        if (CI->getOpcode() == Op) {
97          BasicBlock::iterator It = I; ++It;
98          if (isa<InvokeInst>(I))
99            It = cast<InvokeInst>(I)->getNormalDest()->begin();
100          while (isa<PHINode>(It)) ++It;
101          if (It != BasicBlock::iterator(CI)) {
102            // Recreate the cast at the beginning of the entry block.
103            // The old cast is left in place in case it is being used
104            // as an insert point.
105            Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
106            NewCI->takeName(CI);
107            CI->replaceAllUsesWith(NewCI);
108            return NewCI;
109          }
110          return CI;
111        }
112  }
113  BasicBlock::iterator IP = I; ++IP;
114  if (InvokeInst *II = dyn_cast<InvokeInst>(I))
115    IP = II->getNormalDest()->begin();
116  while (isa<PHINode>(IP)) ++IP;
117  Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
118  InsertedValues.insert(CI);
119  return CI;
120}
121
122/// InsertBinop - Insert the specified binary operator, doing a small amount
123/// of work to avoid inserting an obviously redundant operation.
124Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
125                                 Value *LHS, Value *RHS) {
126  // Fold a binop with constant operands.
127  if (Constant *CLHS = dyn_cast<Constant>(LHS))
128    if (Constant *CRHS = dyn_cast<Constant>(RHS))
129      return getContext()->getConstantExpr(Opcode, CLHS, CRHS);
130
131  // Do a quick scan to see if we have this binop nearby.  If so, reuse it.
132  unsigned ScanLimit = 6;
133  BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
134  // Scanning starts from the last instruction before the insertion point.
135  BasicBlock::iterator IP = Builder.GetInsertPoint();
136  if (IP != BlockBegin) {
137    --IP;
138    for (; ScanLimit; --IP, --ScanLimit) {
139      if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
140          IP->getOperand(1) == RHS)
141        return IP;
142      if (IP == BlockBegin) break;
143    }
144  }
145
146  // If we haven't found this binop, insert it.
147  Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
148  InsertedValues.insert(BO);
149  return BO;
150}
151
152/// FactorOutConstant - Test if S is divisible by Factor, using signed
153/// division. If so, update S with Factor divided out and return true.
154/// S need not be evenly divisble if a reasonable remainder can be
155/// computed.
156/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
157/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
158/// check to see if the divide was folded.
159static bool FactorOutConstant(const SCEV *&S,
160                              const SCEV *&Remainder,
161                              const APInt &Factor,
162                              ScalarEvolution &SE) {
163  // Everything is divisible by one.
164  if (Factor == 1)
165    return true;
166
167  // For a Constant, check for a multiple of the given factor.
168  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
169    ConstantInt *CI =
170      SE.getContext()->getConstantInt(C->getValue()->getValue().sdiv(Factor));
171    // If the quotient is zero and the remainder is non-zero, reject
172    // the value at this scale. It will be considered for subsequent
173    // smaller scales.
174    if (C->isZero() || !CI->isZero()) {
175      const SCEV *Div = SE.getConstant(CI);
176      S = Div;
177      Remainder =
178        SE.getAddExpr(Remainder,
179                      SE.getConstant(C->getValue()->getValue().srem(Factor)));
180      return true;
181    }
182  }
183
184  // In a Mul, check if there is a constant operand which is a multiple
185  // of the given factor.
186  if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S))
187    if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
188      if (!C->getValue()->getValue().srem(Factor)) {
189        const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
190        SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
191                                               MOperands.end());
192        NewMulOps[0] =
193          SE.getConstant(C->getValue()->getValue().sdiv(Factor));
194        S = SE.getMulExpr(NewMulOps);
195        return true;
196      }
197
198  // In an AddRec, check if both start and step are divisible.
199  if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
200    const SCEV *Step = A->getStepRecurrence(SE);
201    const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
202    if (!FactorOutConstant(Step, StepRem, Factor, SE))
203      return false;
204    if (!StepRem->isZero())
205      return false;
206    const SCEV *Start = A->getStart();
207    if (!FactorOutConstant(Start, Remainder, Factor, SE))
208      return false;
209    S = SE.getAddRecExpr(Start, Step, A->getLoop());
210    return true;
211  }
212
213  return false;
214}
215
216/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
217/// instead of using ptrtoint+arithmetic+inttoptr. This helps
218/// BasicAliasAnalysis analyze the result.
219///
220/// Design note: This depends on ScalarEvolution not recognizing inttoptr
221/// and ptrtoint operators, as they may introduce pointer arithmetic
222/// which may not be safely converted into getelementptr.
223///
224/// Design note: It might seem desirable for this function to be more
225/// loop-aware. If some of the indices are loop-invariant while others
226/// aren't, it might seem desirable to emit multiple GEPs, keeping the
227/// loop-invariant portions of the overall computation outside the loop.
228/// However, there are a few reasons this is not done here. Hoisting simple
229/// arithmetic is a low-level optimization that often isn't very
230/// important until late in the optimization process. In fact, passes
231/// like InstructionCombining will combine GEPs, even if it means
232/// pushing loop-invariant computation down into loops, so even if the
233/// GEPs were split here, the work would quickly be undone. The
234/// LoopStrengthReduction pass, which is usually run quite late (and
235/// after the last InstructionCombining pass), takes care of hoisting
236/// loop-invariant portions of expressions, after considering what
237/// can be folded using target addressing modes.
238///
239Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
240                                    const SCEV *const *op_end,
241                                    const PointerType *PTy,
242                                    const Type *Ty,
243                                    Value *V) {
244  const Type *ElTy = PTy->getElementType();
245  SmallVector<Value *, 4> GepIndices;
246  SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
247  bool AnyNonZeroIndices = false;
248
249  // Decend down the pointer's type and attempt to convert the other
250  // operands into GEP indices, at each level. The first index in a GEP
251  // indexes into the array implied by the pointer operand; the rest of
252  // the indices index into the element or field type selected by the
253  // preceding index.
254  for (;;) {
255    APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
256                         ElTy->isSized() ?  SE.TD->getTypeAllocSize(ElTy) : 0);
257    SmallVector<const SCEV *, 8> NewOps;
258    SmallVector<const SCEV *, 8> ScaledOps;
259    for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
260      // Split AddRecs up into parts as either of the parts may be usable
261      // without the other.
262      if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i]))
263        if (!A->getStart()->isZero()) {
264          const SCEV *Start = A->getStart();
265          Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
266                                         A->getStepRecurrence(SE),
267                                         A->getLoop()));
268          Ops[i] = Start;
269          ++e;
270        }
271      // If the scale size is not 0, attempt to factor out a scale.
272      if (ElSize != 0) {
273        const SCEV *Op = Ops[i];
274        const SCEV *Remainder = SE.getIntegerSCEV(0, Op->getType());
275        if (FactorOutConstant(Op, Remainder, ElSize, SE)) {
276          ScaledOps.push_back(Op); // Op now has ElSize factored out.
277          NewOps.push_back(Remainder);
278          continue;
279        }
280      }
281      // If the operand was not divisible, add it to the list of operands
282      // we'll scan next iteration.
283      NewOps.push_back(Ops[i]);
284    }
285    Ops = NewOps;
286    AnyNonZeroIndices |= !ScaledOps.empty();
287    Value *Scaled = ScaledOps.empty() ?
288                    getContext()->getNullValue(Ty) :
289                    expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
290    GepIndices.push_back(Scaled);
291
292    // Collect struct field index operands.
293    if (!Ops.empty())
294      while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
295        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
296          if (SE.getTypeSizeInBits(C->getType()) <= 64) {
297            const StructLayout &SL = *SE.TD->getStructLayout(STy);
298            uint64_t FullOffset = C->getValue()->getZExtValue();
299            if (FullOffset < SL.getSizeInBytes()) {
300              unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
301              GepIndices.push_back(
302                            getContext()->getConstantInt(Type::Int32Ty, ElIdx));
303              ElTy = STy->getTypeAtIndex(ElIdx);
304              Ops[0] =
305                SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
306              AnyNonZeroIndices = true;
307              continue;
308            }
309          }
310        break;
311      }
312
313    if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
314      ElTy = ATy->getElementType();
315      continue;
316    }
317    break;
318  }
319
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);
327
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 getContext()->getConstantExprGetElementPtr(CLHS, &CRHS, 1);
332
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) {
339      --IP;
340      for (; ScanLimit; --IP, --ScanLimit) {
341        if (IP->getOpcode() == Instruction::GetElementPtr &&
342            IP->getOperand(0) == V && IP->getOperand(1) == Idx)
343          return IP;
344        if (IP == BlockBegin) break;
345      }
346    }
347
348    Value *GEP = Builder.CreateGEP(V, Idx, "scevgep");
349    InsertedValues.insert(GEP);
350    return GEP;
351  }
352
353  // Insert a pretty getelementptr.
354  Value *GEP = Builder.CreateGEP(V,
355                                 GepIndices.begin(),
356                                 GepIndices.end(),
357                                 "scevgep");
358  Ops.push_back(SE.getUnknown(GEP));
359  InsertedValues.insert(GEP);
360  return expand(SE.getAddExpr(Ops));
361}
362
363Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
364  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
365  Value *V = expand(S->getOperand(S->getNumOperands()-1));
366
367  // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
368  // comments on expandAddToGEP for details.
369  if (SE.TD)
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);
373    }
374
375  V = InsertNoopCastOfTo(V, Ty);
376
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);
381  }
382  return V;
383}
384
385Value *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())
390      FirstOp = 1;
391
392  int i = S->getNumOperands()-2;
393  Value *V = expandCodeFor(S->getOperand(i+1), Ty);
394
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);
399  }
400
401  // -1 * ...  --->  0 - ...
402  if (FirstOp == 1)
403    V = InsertBinop(Instruction::Sub, getContext()->getNullValue(Ty), V);
404  return V;
405}
406
407Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
408  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
409
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                         getContext()->getConstantInt(Ty, RHS.logBase2()));
416  }
417
418  Value *RHS = expandCodeFor(S->getRHS(), Ty);
419  return InsertBinop(Instruction::UDiv, LHS, RHS);
420}
421
422/// Move parts of Base into Rest to leave Base with the minimal
423/// expression that provides a pointer operand suitable for a
424/// GEP expansion.
425static 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),
432                                          A->getLoop()));
433  }
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);
440  }
441}
442
443Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
444  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
445  const Loop *L = S->getLoop();
446
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))
453      CanonicalIV = PN;
454
455  // Rewrite an AddRec in terms of the canonical induction variable, if
456  // its type is more narrow.
457  if (CanonicalIV &&
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,
471                      NewInsertPt);
472    Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
473    return V;
474  }
475
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);
482
483    // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
484    // comments on expandAddToGEP for details.
485    if (SE.TD) {
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);
499        }
500      }
501    }
502
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))));
506  }
507
508  // {0,+,1} --> Insert a canonical induction variable into the loop!
509  if (S->isAffine() &&
510      S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
511    // If there's a canonical IV, just use it.
512    if (CanonicalIV) {
513      assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
514             "IVs with types different from the canonical IV should "
515             "already have been handled!");
516      return CanonicalIV;
517    }
518
519    // Create and insert the PHI node for the induction variable in the
520    // specified loop.
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(getContext()->getNullValue(Ty), Preheader);
526
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?");
532
533    // Insert a unit add instruction right before the terminator corresponding
534    // to the back-edge.
535    Constant *One = getContext()->getConstantInt(Ty, 1);
536    Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
537                                                 (*HPI)->getTerminator());
538    InsertedValues.insert(Add);
539
540    pred_iterator PI = pred_begin(Header);
541    if (*PI == Preheader)
542      ++PI;
543    PN->addIncoming(Add, *PI);
544    return PN;
545  }
546
547  // {0,+,F} --> {0,+,1} * F
548  // Get the canonical induction variable I for this loop.
549  Value *I = CanonicalIV ?
550             CanonicalIV :
551             getOrInsertCanonicalInductionVariable(L, Ty);
552
553  // If this is a simple linear addrec, emit it now as a special case.
554  if (S->isAffine())    // {0,+,F} --> i*F
555    return
556      expand(SE.getTruncateOrNoop(
557        SE.getMulExpr(SE.getUnknown(I),
558                      SE.getNoopOrAnyExtend(S->getOperand(1),
559                                            I->getType())),
560        Ty));
561
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
565  // into this folder.
566  const SCEV *IH = SE.getUnknown(I);   // Get I as a "symbolic" SCEV.
567
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))
572    NewS = Ext;
573
574  const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
575  //cerr << "Evaluated: " << *this << "\n     to: " << *V << "\n";
576
577  // Truncate the result down to the original type, if needed.
578  const SCEV *T = SE.getTruncateOrNoop(V, Ty);
579  return expand(T);
580}
581
582Value *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);
588  return I;
589}
590
591Value *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);
597  return I;
598}
599
600Value *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);
606  return I;
607}
608
609Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
610  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
611  const Type *Ty = LHS->getType();
612  for (int i = S->getNumOperands()-2; i >= 0; --i) {
613    // In the case of mixed integer and pointer types, do the
614    // rest of the comparisons as integer.
615    if (S->getOperand(i)->getType() != Ty) {
616      Ty = SE.getEffectiveSCEVType(Ty);
617      LHS = InsertNoopCastOfTo(LHS, Ty);
618    }
619    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
620    Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
621    InsertedValues.insert(ICmp);
622    Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
623    InsertedValues.insert(Sel);
624    LHS = Sel;
625  }
626  // In the case of mixed integer and pointer types, cast the
627  // final result back to the pointer type.
628  if (LHS->getType() != S->getType())
629    LHS = InsertNoopCastOfTo(LHS, S->getType());
630  return LHS;
631}
632
633Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
634  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
635  const Type *Ty = LHS->getType();
636  for (int i = S->getNumOperands()-2; i >= 0; --i) {
637    // In the case of mixed integer and pointer types, do the
638    // rest of the comparisons as integer.
639    if (S->getOperand(i)->getType() != Ty) {
640      Ty = SE.getEffectiveSCEVType(Ty);
641      LHS = InsertNoopCastOfTo(LHS, Ty);
642    }
643    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
644    Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
645    InsertedValues.insert(ICmp);
646    Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
647    InsertedValues.insert(Sel);
648    LHS = Sel;
649  }
650  // In the case of mixed integer and pointer types, cast the
651  // final result back to the pointer type.
652  if (LHS->getType() != S->getType())
653    LHS = InsertNoopCastOfTo(LHS, S->getType());
654  return LHS;
655}
656
657Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
658  // Expand the code for this SCEV.
659  Value *V = expand(SH);
660  if (Ty) {
661    assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
662           "non-trivial casts should be done with the SCEVs directly!");
663    V = InsertNoopCastOfTo(V, Ty);
664  }
665  return V;
666}
667
668Value *SCEVExpander::expand(const SCEV *S) {
669  // Compute an insertion point for this SCEV object. Hoist the instructions
670  // as far out in the loop nest as possible.
671  Instruction *InsertPt = Builder.GetInsertPoint();
672  for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
673       L = L->getParentLoop())
674    if (S->isLoopInvariant(L)) {
675      if (!L) break;
676      if (BasicBlock *Preheader = L->getLoopPreheader())
677        InsertPt = Preheader->getTerminator();
678    } else {
679      // If the SCEV is computable at this level, insert it into the header
680      // after the PHIs (and after any other instructions that we've inserted
681      // there) so that it is guaranteed to dominate any user inside the loop.
682      if (L && S->hasComputableLoopEvolution(L))
683        InsertPt = L->getHeader()->getFirstNonPHI();
684      while (isInsertedInstruction(InsertPt))
685        InsertPt = next(BasicBlock::iterator(InsertPt));
686      break;
687    }
688
689  // Check to see if we already expanded this here.
690  std::map<std::pair<const SCEV *, Instruction *>,
691           AssertingVH<Value> >::iterator I =
692    InsertedExpressions.find(std::make_pair(S, InsertPt));
693  if (I != InsertedExpressions.end())
694    return I->second;
695
696  BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
697  BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
698  Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
699
700  // Expand the expression into instructions.
701  Value *V = visit(S);
702
703  // Remember the expanded value for this SCEV at this location.
704  InsertedExpressions[std::make_pair(S, InsertPt)] = V;
705
706  Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
707  return V;
708}
709
710/// getOrInsertCanonicalInductionVariable - This method returns the
711/// canonical induction variable of the specified type for the specified
712/// loop (inserting one if there is none).  A canonical induction variable
713/// starts at zero and steps by one on each iteration.
714Value *
715SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
716                                                    const Type *Ty) {
717  assert(Ty->isInteger() && "Can only insert integer induction variables!");
718  const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
719                                   SE.getIntegerSCEV(1, Ty), L);
720  BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
721  BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
722  Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
723  if (SaveInsertBB)
724    Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
725  return V;
726}
727