ScalarEvolutionExpander.cpp revision 6de29f8d960505421d61c80cdb738e16720b6c0e 
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/Target/TargetData.h"
19#include "llvm/ADT/STLExtras.h"
20using namespace llvm;
21
22/// InsertCastOfTo - Insert a cast of V to the specified type, doing what
23/// we can to share the casts.
24Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
25                                    const Type *Ty) {
26  // Short-circuit unnecessary bitcasts.
27  if (opcode == Instruction::BitCast && V->getType() == Ty)
28    return V;
29
30  // Short-circuit unnecessary inttoptr<->ptrtoint casts.
31  if ((opcode == Instruction::PtrToInt || opcode == Instruction::IntToPtr) &&
32      SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
33    if (CastInst *CI = dyn_cast<CastInst>(V))
34      if ((CI->getOpcode() == Instruction::PtrToInt ||
35           CI->getOpcode() == Instruction::IntToPtr) &&
36          SE.getTypeSizeInBits(CI->getType()) ==
37          SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
38        return CI->getOperand(0);
39    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
40      if ((CE->getOpcode() == Instruction::PtrToInt ||
41           CE->getOpcode() == Instruction::IntToPtr) &&
42          SE.getTypeSizeInBits(CE->getType()) ==
43          SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
44        return CE->getOperand(0);
45  }
46
47  // FIXME: keep track of the cast instruction.
48  if (Constant *C = dyn_cast<Constant>(V))
49    return ConstantExpr::getCast(opcode, C, Ty);
50
51  if (Argument *A = dyn_cast<Argument>(V)) {
52    // Check to see if there is already a cast!
53    for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
54         UI != E; ++UI) {
55      if ((*UI)->getType() == Ty)
56        if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
57          if (CI->getOpcode() == opcode) {
58            // If the cast isn't the first instruction of the function, move it.
59            if (BasicBlock::iterator(CI) !=
60                A->getParent()->getEntryBlock().begin()) {
61              // If the CastInst is the insert point, change the insert point.
62              if (CI == InsertPt) ++InsertPt;
63              // Splice the cast at the beginning of the entry block.
64              CI->moveBefore(A->getParent()->getEntryBlock().begin());
65            }
66            return CI;
67          }
68    }
69    Instruction *I = CastInst::Create(opcode, V, Ty, V->getName(),
70                                      A->getParent()->getEntryBlock().begin());
71    InsertedValues.insert(I);
72    return I;
73  }
74
75  Instruction *I = cast<Instruction>(V);
76
77  // Check to see if there is already a cast.  If there is, use it.
78  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
79       UI != E; ++UI) {
80    if ((*UI)->getType() == Ty)
81      if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
82        if (CI->getOpcode() == opcode) {
83          BasicBlock::iterator It = I; ++It;
84          if (isa<InvokeInst>(I))
85            It = cast<InvokeInst>(I)->getNormalDest()->begin();
86          while (isa<PHINode>(It)) ++It;
87          if (It != BasicBlock::iterator(CI)) {
88            // If the CastInst is the insert point, change the insert point.
89            if (CI == InsertPt) ++InsertPt;
90            // Splice the cast immediately after the operand in question.
91            CI->moveBefore(It);
92          }
93          return CI;
94        }
95  }
96  BasicBlock::iterator IP = I; ++IP;
97  if (InvokeInst *II = dyn_cast<InvokeInst>(I))
98    IP = II->getNormalDest()->begin();
99  while (isa<PHINode>(IP)) ++IP;
100  Instruction *CI = CastInst::Create(opcode, V, Ty, V->getName(), IP);
101  InsertedValues.insert(CI);
102  return CI;
103}
104
105/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
106/// which must be possible with a noop cast.
107Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
108  Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
109  assert((Op == Instruction::BitCast ||
110          Op == Instruction::PtrToInt ||
111          Op == Instruction::IntToPtr) &&
112         "InsertNoopCastOfTo cannot perform non-noop casts!");
113  assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
114         "InsertNoopCastOfTo cannot change sizes!");
115  return InsertCastOfTo(Op, V, Ty);
116}
117
118/// InsertBinop - Insert the specified binary operator, doing a small amount
119/// of work to avoid inserting an obviously redundant operation.
120Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS,
121                                 Value *RHS, BasicBlock::iterator InsertPt) {
122  // Fold a binop with constant operands.
123  if (Constant *CLHS = dyn_cast<Constant>(LHS))
124    if (Constant *CRHS = dyn_cast<Constant>(RHS))
125      return ConstantExpr::get(Opcode, CLHS, CRHS);
126
127  // Do a quick scan to see if we have this binop nearby.  If so, reuse it.
128  unsigned ScanLimit = 6;
129  BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
130  if (InsertPt != BlockBegin) {
131    // Scanning starts from the last instruction before InsertPt.
132    BasicBlock::iterator IP = InsertPt;
133    --IP;
134    for (; ScanLimit; --IP, --ScanLimit) {
135      if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
136          IP->getOperand(1) == RHS)
137        return IP;
138      if (IP == BlockBegin) break;
139    }
140  }
141
142  // If we haven't found this binop, insert it.
143  Instruction *BO = BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt);
144  InsertedValues.insert(BO);
145  return BO;
146}
147
148/// FactorOutConstant - Test if S is divisible by Factor, using signed
149/// division. If so, update S with Factor divided out and return true.
150/// S need not be evenly divisble if a reasonable remainder can be
151/// computed.
152/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
153/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
154/// check to see if the divide was folded.
155static bool FactorOutConstant(SCEVHandle &S,
156                              SCEVHandle &Remainder,
157                              const APInt &Factor,
158                              ScalarEvolution &SE) {
159  // Everything is divisible by one.
160  if (Factor == 1)
161    return true;
162
163  // For a Constant, check for a multiple of the given factor.
164  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
165    ConstantInt *CI =
166      ConstantInt::get(C->getValue()->getValue().sdiv(Factor));
167    // If the quotient is zero and the remainder is non-zero, reject
168    // the value at this scale. It will be considered for subsequent
169    // smaller scales.
170    if (C->isZero() || !CI->isZero()) {
171      SCEVHandle Div = SE.getConstant(CI);
172      S = Div;
173      Remainder =
174        SE.getAddExpr(Remainder,
175                      SE.getConstant(C->getValue()->getValue().srem(Factor)));
176      return true;
177    }
178  }
179
180  // In a Mul, check if there is a constant operand which is a multiple
181  // of the given factor.
182  if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S))
183    if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
184      if (!C->getValue()->getValue().srem(Factor)) {
185        const SmallVectorImpl<SCEVHandle> &MOperands = M->getOperands();
186        SmallVector<SCEVHandle, 4> NewMulOps(MOperands.begin(), MOperands.end());
187        NewMulOps[0] =
188          SE.getConstant(C->getValue()->getValue().sdiv(Factor));
189        S = SE.getMulExpr(NewMulOps);
190        return true;
191      }
192
193  // In an AddRec, check if both start and step are divisible.
194  if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
195    SCEVHandle Step = A->getStepRecurrence(SE);
196    SCEVHandle StepRem = SE.getIntegerSCEV(0, Step->getType());
197    if (!FactorOutConstant(Step, StepRem, Factor, SE))
198      return false;
199    if (!StepRem->isZero())
200      return false;
201    SCEVHandle Start = A->getStart();
202    if (!FactorOutConstant(Start, Remainder, Factor, SE))
203      return false;
204    S = SE.getAddRecExpr(Start, Step, A->getLoop());
205    return true;
206  }
207
208  return false;
209}
210
211/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
212/// instead of using ptrtoint+arithmetic+inttoptr. This helps
213/// BasicAliasAnalysis analyze the result. However, it suffers from the
214/// underlying bug described in PR2831. Addition in LLVM currently always
215/// has two's complement wrapping guaranteed. However, the semantics for
216/// getelementptr overflow are ambiguous. In the common case though, this
217/// expansion gets used when a GEP in the original code has been converted
218/// into integer arithmetic, in which case the resulting code will be no
219/// more undefined than it was originally.
220///
221/// Design note: It might seem desirable for this function to be more
222/// loop-aware. If some of the indices are loop-invariant while others
223/// aren't, it might seem desirable to emit multiple GEPs, keeping the
224/// loop-invariant portions of the overall computation outside the loop.
225/// However, there are a few reasons this is not done here. Hoisting simple
226/// arithmetic is a low-level optimization that often isn't very
227/// important until late in the optimization process. In fact, passes
228/// like InstructionCombining will combine GEPs, even if it means
229/// pushing loop-invariant computation down into loops, so even if the
230/// GEPs were split here, the work would quickly be undone. The
231/// LoopStrengthReduction pass, which is usually run quite late (and
232/// after the last InstructionCombining pass), takes care of hoisting
233/// loop-invariant portions of expressions, after considering what
234/// can be folded using target addressing modes.
235///
236Value *SCEVExpander::expandAddToGEP(const SCEVHandle *op_begin,
237                                    const SCEVHandle *op_end,
238                                    const PointerType *PTy,
239                                    const Type *Ty,
240                                    Value *V) {
241  const Type *ElTy = PTy->getElementType();
242  SmallVector<Value *, 4> GepIndices;
243  SmallVector<SCEVHandle, 8> Ops(op_begin, op_end);
244  bool AnyNonZeroIndices = false;
245
246  // Decend down the pointer's type and attempt to convert the other
247  // operands into GEP indices, at each level. The first index in a GEP
248  // indexes into the array implied by the pointer operand; the rest of
249  // the indices index into the element or field type selected by the
250  // preceding index.
251  for (;;) {
252    APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
253                         ElTy->isSized() ?  SE.TD->getTypeAllocSize(ElTy) : 0);
254    SmallVector<SCEVHandle, 8> NewOps;
255    SmallVector<SCEVHandle, 8> ScaledOps;
256    for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
257      // Split AddRecs up into parts as either of the parts may be usable
258      // without the other.
259      if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i]))
260        if (!A->getStart()->isZero()) {
261          SCEVHandle Start = A->getStart();
262          Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
263                                         A->getStepRecurrence(SE),
264                                         A->getLoop()));
265          Ops[i] = Start;
266          ++e;
267        }
268      // If the scale size is not 0, attempt to factor out a scale.
269      if (ElSize != 0) {
270        SCEVHandle Op = Ops[i];
271        SCEVHandle Remainder = SE.getIntegerSCEV(0, Op->getType());
272        if (FactorOutConstant(Op, Remainder, ElSize, SE)) {
273          ScaledOps.push_back(Op); // Op now has ElSize factored out.
274          NewOps.push_back(Remainder);
275          continue;
276        }
277      }
278      // If the operand was not divisible, add it to the list of operands
279      // we'll scan next iteration.
280      NewOps.push_back(Ops[i]);
281    }
282    Ops = NewOps;
283    AnyNonZeroIndices |= !ScaledOps.empty();
284    Value *Scaled = ScaledOps.empty() ?
285                    Constant::getNullValue(Ty) :
286                    expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
287    GepIndices.push_back(Scaled);
288
289    // Collect struct field index operands.
290    if (!Ops.empty())
291      while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
292        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
293          if (SE.getTypeSizeInBits(C->getType()) <= 64) {
294            const StructLayout &SL = *SE.TD->getStructLayout(STy);
295            uint64_t FullOffset = C->getValue()->getZExtValue();
296            if (FullOffset < SL.getSizeInBytes()) {
297              unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
298              GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx));
299              ElTy = STy->getTypeAtIndex(ElIdx);
300              Ops[0] =
301                SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
302              AnyNonZeroIndices = true;
303              continue;
304            }
305          }
306        break;
307      }
308
309    if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
310      ElTy = ATy->getElementType();
311      continue;
312    }
313    break;
314  }
315
316  // If none of the operands were convertable to proper GEP indices, cast
317  // the base to i8* and do an ugly getelementptr with that. It's still
318  // better than ptrtoint+arithmetic+inttoptr at least.
319  if (!AnyNonZeroIndices) {
320    V = InsertNoopCastOfTo(V,
321                           Type::Int8Ty->getPointerTo(PTy->getAddressSpace()));
322    Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
323
324    // Fold a GEP with constant operands.
325    if (Constant *CLHS = dyn_cast<Constant>(V))
326      if (Constant *CRHS = dyn_cast<Constant>(Idx))
327        return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
328
329    // Do a quick scan to see if we have this GEP nearby.  If so, reuse it.
330    unsigned ScanLimit = 6;
331    BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
332    if (InsertPt != BlockBegin) {
333      // Scanning starts from the last instruction before InsertPt.
334      BasicBlock::iterator IP = InsertPt;
335      --IP;
336      for (; ScanLimit; --IP, --ScanLimit) {
337        if (IP->getOpcode() == Instruction::GetElementPtr &&
338            IP->getOperand(0) == V && IP->getOperand(1) == Idx)
339          return IP;
340        if (IP == BlockBegin) break;
341      }
342    }
343
344    Value *GEP = GetElementPtrInst::Create(V, Idx, "scevgep", InsertPt);
345    InsertedValues.insert(GEP);
346    return GEP;
347  }
348
349  // Insert a pretty getelementptr.
350  Value *GEP = GetElementPtrInst::Create(V,
351                                         GepIndices.begin(),
352                                         GepIndices.end(),
353                                         "scevgep", InsertPt);
354  Ops.push_back(SE.getUnknown(GEP));
355  InsertedValues.insert(GEP);
356  return expand(SE.getAddExpr(Ops));
357}
358
359Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
360  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
361  Value *V = expand(S->getOperand(S->getNumOperands()-1));
362
363  // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
364  // comments on expandAddToGEP for details.
365  if (SE.TD)
366    if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
367      const SmallVectorImpl<SCEVHandle> &Ops = S->getOperands();
368      return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1],
369                            PTy, Ty, V);
370    }
371
372  V = InsertNoopCastOfTo(V, Ty);
373
374  // Emit a bunch of add instructions
375  for (int i = S->getNumOperands()-2; i >= 0; --i) {
376    Value *W = expandCodeFor(S->getOperand(i), Ty);
377    V = InsertBinop(Instruction::Add, V, W, InsertPt);
378  }
379  return V;
380}
381
382Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
383  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
384  int FirstOp = 0;  // Set if we should emit a subtract.
385  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
386    if (SC->getValue()->isAllOnesValue())
387      FirstOp = 1;
388
389  int i = S->getNumOperands()-2;
390  Value *V = expandCodeFor(S->getOperand(i+1), Ty);
391
392  // Emit a bunch of multiply instructions
393  for (; i >= FirstOp; --i) {
394    Value *W = expandCodeFor(S->getOperand(i), Ty);
395    V = InsertBinop(Instruction::Mul, V, W, InsertPt);
396  }
397
398  // -1 * ...  --->  0 - ...
399  if (FirstOp == 1)
400    V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V, InsertPt);
401  return V;
402}
403
404Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
405  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
406
407  Value *LHS = expandCodeFor(S->getLHS(), Ty);
408  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
409    const APInt &RHS = SC->getValue()->getValue();
410    if (RHS.isPowerOf2())
411      return InsertBinop(Instruction::LShr, LHS,
412                         ConstantInt::get(Ty, RHS.logBase2()),
413                         InsertPt);
414  }
415
416  Value *RHS = expandCodeFor(S->getRHS(), Ty);
417  return InsertBinop(Instruction::UDiv, LHS, RHS, InsertPt);
418}
419
420/// Move parts of Base into Rest to leave Base with the minimal
421/// expression that provides a pointer operand suitable for a
422/// GEP expansion.
423static void ExposePointerBase(SCEVHandle &Base, SCEVHandle &Rest,
424                              ScalarEvolution &SE) {
425  while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
426    Base = A->getStart();
427    Rest = SE.getAddExpr(Rest,
428                         SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
429                                          A->getStepRecurrence(SE),
430                                          A->getLoop()));
431  }
432  if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
433    Base = A->getOperand(A->getNumOperands()-1);
434    SmallVector<SCEVHandle, 8> NewAddOps(A->op_begin(), A->op_end());
435    NewAddOps.back() = Rest;
436    Rest = SE.getAddExpr(NewAddOps);
437    ExposePointerBase(Base, Rest, SE);
438  }
439}
440
441Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
442  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
443  const Loop *L = S->getLoop();
444
445  // First check for an existing canonical IV in a suitable type.
446  PHINode *CanonicalIV = 0;
447  if (PHINode *PN = L->getCanonicalInductionVariable())
448    if (SE.isSCEVable(PN->getType()) &&
449        isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
450        SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
451      CanonicalIV = PN;
452
453  // Rewrite an AddRec in terms of the canonical induction variable, if
454  // its type is more narrow.
455  if (CanonicalIV &&
456      SE.getTypeSizeInBits(CanonicalIV->getType()) >
457      SE.getTypeSizeInBits(Ty)) {
458    SCEVHandle Start = SE.getAnyExtendExpr(S->getStart(),
459                                           CanonicalIV->getType());
460    SCEVHandle Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
461                                          CanonicalIV->getType());
462    Value *V = expand(SE.getAddRecExpr(Start, Step, S->getLoop()));
463    BasicBlock::iterator SaveInsertPt = getInsertionPoint();
464    BasicBlock::iterator NewInsertPt =
465      next(BasicBlock::iterator(cast<Instruction>(V)));
466    while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
467    V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
468                      NewInsertPt);
469    setInsertionPoint(SaveInsertPt);
470    return V;
471  }
472
473  // {X,+,F} --> X + {0,+,F}
474  if (!S->getStart()->isZero()) {
475    const SmallVectorImpl<SCEVHandle> &SOperands = S->getOperands();
476    SmallVector<SCEVHandle, 4> NewOps(SOperands.begin(), SOperands.end());
477    NewOps[0] = SE.getIntegerSCEV(0, Ty);
478    SCEVHandle Rest = SE.getAddRecExpr(NewOps, L);
479
480    // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
481    // comments on expandAddToGEP for details.
482    if (SE.TD) {
483      SCEVHandle Base = S->getStart();
484      SCEVHandle RestArray[1] = { Rest };
485      // Dig into the expression to find the pointer base for a GEP.
486      ExposePointerBase(Base, RestArray[0], SE);
487      // If we found a pointer, expand the AddRec with a GEP.
488      if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
489        // Make sure the Base isn't something exotic, such as a multiplied
490        // or divided pointer value. In those cases, the result type isn't
491        // actually a pointer type.
492        if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
493          Value *StartV = expand(Base);
494          assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
495          return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
496        }
497      }
498    }
499
500    Value *RestV = expand(Rest);
501    return expand(SE.getAddExpr(S->getStart(), SE.getUnknown(RestV)));
502  }
503
504  // {0,+,1} --> Insert a canonical induction variable into the loop!
505  if (S->isAffine() &&
506      S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
507    // If there's a canonical IV, just use it.
508    if (CanonicalIV) {
509      assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
510             "IVs with types different from the canonical IV should "
511             "already have been handled!");
512      return CanonicalIV;
513    }
514
515    // Create and insert the PHI node for the induction variable in the
516    // specified loop.
517    BasicBlock *Header = L->getHeader();
518    PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
519    InsertedValues.insert(PN);
520    PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
521
522    pred_iterator HPI = pred_begin(Header);
523    assert(HPI != pred_end(Header) && "Loop with zero preds???");
524    if (!L->contains(*HPI)) ++HPI;
525    assert(HPI != pred_end(Header) && L->contains(*HPI) &&
526           "No backedge in loop?");
527
528    // Insert a unit add instruction right before the terminator corresponding
529    // to the back-edge.
530    Constant *One = ConstantInt::get(Ty, 1);
531    Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
532                                                 (*HPI)->getTerminator());
533    InsertedValues.insert(Add);
534
535    pred_iterator PI = pred_begin(Header);
536    if (*PI == L->getLoopPreheader())
537      ++PI;
538    PN->addIncoming(Add, *PI);
539    return PN;
540  }
541
542  // {0,+,F} --> {0,+,1} * F
543  // Get the canonical induction variable I for this loop.
544  Value *I = CanonicalIV ?
545             CanonicalIV :
546             getOrInsertCanonicalInductionVariable(L, Ty);
547
548  // If this is a simple linear addrec, emit it now as a special case.
549  if (S->isAffine()) {   // {0,+,F} --> i*F
550    Value *F = expandCodeFor(S->getOperand(1), Ty);
551
552    // If the insert point is directly inside of the loop, emit the multiply at
553    // the insert point.  Otherwise, L is a loop that is a parent of the insert
554    // point loop.  If we can, move the multiply to the outer most loop that it
555    // is safe to be in.
556    BasicBlock::iterator MulInsertPt = getInsertionPoint();
557    Loop *InsertPtLoop = SE.LI->getLoopFor(MulInsertPt->getParent());
558    if (InsertPtLoop != L && InsertPtLoop &&
559        L->contains(InsertPtLoop->getHeader())) {
560      do {
561        // If we cannot hoist the multiply out of this loop, don't.
562        if (!InsertPtLoop->isLoopInvariant(F)) break;
563
564        BasicBlock *InsertPtLoopPH = InsertPtLoop->getLoopPreheader();
565
566        // If this loop hasn't got a preheader, we aren't able to hoist the
567        // multiply.
568        if (!InsertPtLoopPH)
569          break;
570
571        // Otherwise, move the insert point to the preheader.
572        MulInsertPt = InsertPtLoopPH->getTerminator();
573        InsertPtLoop = InsertPtLoop->getParentLoop();
574      } while (InsertPtLoop != L);
575    }
576
577    return InsertBinop(Instruction::Mul, I, F, MulInsertPt);
578  }
579
580  // If this is a chain of recurrences, turn it into a closed form, using the
581  // folders, then expandCodeFor the closed form.  This allows the folders to
582  // simplify the expression without having to build a bunch of special code
583  // into this folder.
584  SCEVHandle IH = SE.getUnknown(I);   // Get I as a "symbolic" SCEV.
585
586  // Promote S up to the canonical IV type, if the cast is foldable.
587  SCEVHandle NewS = S;
588  SCEVHandle Ext = SE.getNoopOrAnyExtend(S, I->getType());
589  if (isa<SCEVAddRecExpr>(Ext))
590    NewS = Ext;
591
592  SCEVHandle V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
593  //cerr << "Evaluated: " << *this << "\n     to: " << *V << "\n";
594
595  // Truncate the result down to the original type, if needed.
596  SCEVHandle T = SE.getTruncateOrNoop(V, Ty);
597  return expand(V);
598}
599
600Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
601  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
602  Value *V = expandCodeFor(S->getOperand(),
603                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
604  Instruction *I = new TruncInst(V, Ty, "tmp.", InsertPt);
605  InsertedValues.insert(I);
606  return I;
607}
608
609Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
610  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
611  Value *V = expandCodeFor(S->getOperand(),
612                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
613  Instruction *I = new ZExtInst(V, Ty, "tmp.", InsertPt);
614  InsertedValues.insert(I);
615  return I;
616}
617
618Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
619  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
620  Value *V = expandCodeFor(S->getOperand(),
621                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
622  Instruction *I = new SExtInst(V, Ty, "tmp.", InsertPt);
623  InsertedValues.insert(I);
624  return I;
625}
626
627Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
628  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
629  Value *LHS = expandCodeFor(S->getOperand(0), Ty);
630  for (unsigned i = 1; i < S->getNumOperands(); ++i) {
631    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
632    Instruction *ICmp =
633      new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt);
634    InsertedValues.insert(ICmp);
635    Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt);
636    InsertedValues.insert(Sel);
637    LHS = Sel;
638  }
639  return LHS;
640}
641
642Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
643  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
644  Value *LHS = expandCodeFor(S->getOperand(0), Ty);
645  for (unsigned i = 1; i < S->getNumOperands(); ++i) {
646    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
647    Instruction *ICmp =
648      new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt);
649    InsertedValues.insert(ICmp);
650    Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt);
651    InsertedValues.insert(Sel);
652    LHS = Sel;
653  }
654  return LHS;
655}
656
657Value *SCEVExpander::expandCodeFor(SCEVHandle 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  // Check to see if we already expanded this.
670  std::map<SCEVHandle, AssertingVH<Value> >::iterator I =
671    InsertedExpressions.find(S);
672  if (I != InsertedExpressions.end())
673    return I->second;
674
675  Value *V = visit(S);
676  InsertedExpressions[S] = V;
677  return V;
678}
679
680/// getOrInsertCanonicalInductionVariable - This method returns the
681/// canonical induction variable of the specified type for the specified
682/// loop (inserting one if there is none).  A canonical induction variable
683/// starts at zero and steps by one on each iteration.
684Value *
685SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
686                                                    const Type *Ty) {
687  assert(Ty->isInteger() && "Can only insert integer induction variables!");
688  SCEVHandle H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
689                                  SE.getIntegerSCEV(1, Ty), L);
690  return expand(H);
691}
692