ScalarEvolutionExpander.cpp revision c40f17b08774c2dcc5787fd83241e3c64ba82974 
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 ConstantExpr::getCast(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 ConstantExpr::get(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 SCEV *Factor,
162                              ScalarEvolution &SE,
163                              const TargetData *TD) {
164  // Everything is divisible by one.
165  if (Factor->isOne())
166    return true;
167
168  // x/x == 1.
169  if (S == Factor) {
170    S = SE.getIntegerSCEV(1, S->getType());
171    return true;
172  }
173
174  // For a Constant, check for a multiple of the given factor.
175  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
176    // 0/x == 0.
177    if (C->isZero())
178      return true;
179    // Check for divisibility.
180    if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
181      ConstantInt *CI =
182        ConstantInt::get(SE.getContext(),
183                         C->getValue()->getValue().sdiv(
184                                                   FC->getValue()->getValue()));
185      // If the quotient is zero and the remainder is non-zero, reject
186      // the value at this scale. It will be considered for subsequent
187      // smaller scales.
188      if (!CI->isZero()) {
189        const SCEV *Div = SE.getConstant(CI);
190        S = Div;
191        Remainder =
192          SE.getAddExpr(Remainder,
193                        SE.getConstant(C->getValue()->getValue().srem(
194                                                  FC->getValue()->getValue())));
195        return true;
196      }
197    }
198  }
199
200  // In a Mul, check if there is a constant operand which is a multiple
201  // of the given factor.
202  if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
203    if (TD) {
204      // With TargetData, the size is known. Check if there is a constant
205      // operand which is a multiple of the given factor. If so, we can
206      // factor it.
207      const SCEVConstant *FC = cast<SCEVConstant>(Factor);
208      if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
209        if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
210          const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
211          SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
212                                                 MOperands.end());
213          NewMulOps[0] =
214            SE.getConstant(C->getValue()->getValue().sdiv(
215                                                   FC->getValue()->getValue()));
216          S = SE.getMulExpr(NewMulOps);
217          return true;
218        }
219    } else {
220      // Without TargetData, check if Factor can be factored out of any of the
221      // Mul's operands. If so, we can just remove it.
222      for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
223        const SCEV *SOp = M->getOperand(i);
224        const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
225        if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
226            Remainder->isZero()) {
227          const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
228          SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
229                                                 MOperands.end());
230          NewMulOps[i] = SOp;
231          S = SE.getMulExpr(NewMulOps);
232          return true;
233        }
234      }
235    }
236  }
237
238  // In an AddRec, check if both start and step are divisible.
239  if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
240    const SCEV *Step = A->getStepRecurrence(SE);
241    const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
242    if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
243      return false;
244    if (!StepRem->isZero())
245      return false;
246    const SCEV *Start = A->getStart();
247    if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
248      return false;
249    S = SE.getAddRecExpr(Start, Step, A->getLoop());
250    return true;
251  }
252
253  return false;
254}
255
256/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
257/// is the number of SCEVAddRecExprs present, which are kept at the end of
258/// the list.
259///
260static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
261                                const Type *Ty,
262                                ScalarEvolution &SE) {
263  unsigned NumAddRecs = 0;
264  for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
265    ++NumAddRecs;
266  // Group Ops into non-addrecs and addrecs.
267  SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
268  SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
269  // Let ScalarEvolution sort and simplify the non-addrecs list.
270  const SCEV *Sum = NoAddRecs.empty() ?
271                    SE.getIntegerSCEV(0, Ty) :
272                    SE.getAddExpr(NoAddRecs);
273  // If it returned an add, use the operands. Otherwise it simplified
274  // the sum into a single value, so just use that.
275  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
276    Ops = Add->getOperands();
277  else {
278    Ops.clear();
279    if (!Sum->isZero())
280      Ops.push_back(Sum);
281  }
282  // Then append the addrecs.
283  Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
284}
285
286/// SplitAddRecs - Flatten a list of add operands, moving addrec start values
287/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
288/// This helps expose more opportunities for folding parts of the expressions
289/// into GEP indices.
290///
291static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
292                         const Type *Ty,
293                         ScalarEvolution &SE) {
294  // Find the addrecs.
295  SmallVector<const SCEV *, 8> AddRecs;
296  for (unsigned i = 0, e = Ops.size(); i != e; ++i)
297    while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
298      const SCEV *Start = A->getStart();
299      if (Start->isZero()) break;
300      const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
301      AddRecs.push_back(SE.getAddRecExpr(Zero,
302                                         A->getStepRecurrence(SE),
303                                         A->getLoop()));
304      if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
305        Ops[i] = Zero;
306        Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
307        e += Add->getNumOperands();
308      } else {
309        Ops[i] = Start;
310      }
311    }
312  if (!AddRecs.empty()) {
313    // Add the addrecs onto the end of the list.
314    Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
315    // Resort the operand list, moving any constants to the front.
316    SimplifyAddOperands(Ops, Ty, SE);
317  }
318}
319
320/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
321/// instead of using ptrtoint+arithmetic+inttoptr. This helps
322/// BasicAliasAnalysis and other passes analyze the result.
323///
324/// Design note: This depends on ScalarEvolution not recognizing inttoptr
325/// and ptrtoint operators, as they may introduce pointer arithmetic
326/// which may not be safely converted into getelementptr.
327///
328/// Design note: It might seem desirable for this function to be more
329/// loop-aware. If some of the indices are loop-invariant while others
330/// aren't, it might seem desirable to emit multiple GEPs, keeping the
331/// loop-invariant portions of the overall computation outside the loop.
332/// However, there are a few reasons this is not done here. Hoisting simple
333/// arithmetic is a low-level optimization that often isn't very
334/// important until late in the optimization process. In fact, passes
335/// like InstructionCombining will combine GEPs, even if it means
336/// pushing loop-invariant computation down into loops, so even if the
337/// GEPs were split here, the work would quickly be undone. The
338/// LoopStrengthReduction pass, which is usually run quite late (and
339/// after the last InstructionCombining pass), takes care of hoisting
340/// loop-invariant portions of expressions, after considering what
341/// can be folded using target addressing modes.
342///
343Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
344                                    const SCEV *const *op_end,
345                                    const PointerType *PTy,
346                                    const Type *Ty,
347                                    Value *V) {
348  const Type *ElTy = PTy->getElementType();
349  SmallVector<Value *, 4> GepIndices;
350  SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
351  bool AnyNonZeroIndices = false;
352
353  // Split AddRecs up into parts as either of the parts may be usable
354  // without the other.
355  SplitAddRecs(Ops, Ty, SE);
356
357  // Decend down the pointer's type and attempt to convert the other
358  // operands into GEP indices, at each level. The first index in a GEP
359  // indexes into the array implied by the pointer operand; the rest of
360  // the indices index into the element or field type selected by the
361  // preceding index.
362  for (;;) {
363    const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
364    // If the scale size is not 0, attempt to factor out a scale for
365    // array indexing.
366    SmallVector<const SCEV *, 8> ScaledOps;
367    if (ElTy->isSized() && !ElSize->isZero()) {
368      SmallVector<const SCEV *, 8> NewOps;
369      for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
370        const SCEV *Op = Ops[i];
371        const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
372        if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
373          // Op now has ElSize factored out.
374          ScaledOps.push_back(Op);
375          if (!Remainder->isZero())
376            NewOps.push_back(Remainder);
377          AnyNonZeroIndices = true;
378        } else {
379          // The operand was not divisible, so add it to the list of operands
380          // we'll scan next iteration.
381          NewOps.push_back(Ops[i]);
382        }
383      }
384      // If we made any changes, update Ops.
385      if (!ScaledOps.empty()) {
386        Ops = NewOps;
387        SimplifyAddOperands(Ops, Ty, SE);
388      }
389    }
390
391    // Record the scaled array index for this level of the type. If
392    // we didn't find any operands that could be factored, tentatively
393    // assume that element zero was selected (since the zero offset
394    // would obviously be folded away).
395    Value *Scaled = ScaledOps.empty() ?
396                    Constant::getNullValue(Ty) :
397                    expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
398    GepIndices.push_back(Scaled);
399
400    // Collect struct field index operands.
401    while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
402      bool FoundFieldNo = false;
403      // An empty struct has no fields.
404      if (STy->getNumElements() == 0) break;
405      if (SE.TD) {
406        // With TargetData, field offsets are known. See if a constant offset
407        // falls within any of the struct fields.
408        if (Ops.empty()) break;
409        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
410          if (SE.getTypeSizeInBits(C->getType()) <= 64) {
411            const StructLayout &SL = *SE.TD->getStructLayout(STy);
412            uint64_t FullOffset = C->getValue()->getZExtValue();
413            if (FullOffset < SL.getSizeInBytes()) {
414              unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
415              GepIndices.push_back(
416                  ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
417              ElTy = STy->getTypeAtIndex(ElIdx);
418              Ops[0] =
419                SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
420              AnyNonZeroIndices = true;
421              FoundFieldNo = true;
422            }
423          }
424      } else {
425        // Without TargetData, just check for a SCEVFieldOffsetExpr of the
426        // appropriate struct type.
427        for (unsigned i = 0, e = Ops.size(); i != e; ++i)
428          if (const SCEVFieldOffsetExpr *FO =
429                dyn_cast<SCEVFieldOffsetExpr>(Ops[i]))
430            if (FO->getStructType() == STy) {
431              unsigned FieldNo = FO->getFieldNo();
432              GepIndices.push_back(
433                  ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
434                                   FieldNo));
435              ElTy = STy->getTypeAtIndex(FieldNo);
436              Ops[i] = SE.getConstant(Ty, 0);
437              AnyNonZeroIndices = true;
438              FoundFieldNo = true;
439              break;
440            }
441      }
442      // If no struct field offsets were found, tentatively assume that
443      // field zero was selected (since the zero offset would obviously
444      // be folded away).
445      if (!FoundFieldNo) {
446        ElTy = STy->getTypeAtIndex(0u);
447        GepIndices.push_back(
448          Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
449      }
450    }
451
452    if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
453      ElTy = ATy->getElementType();
454    else
455      break;
456  }
457
458  // If none of the operands were convertable to proper GEP indices, cast
459  // the base to i8* and do an ugly getelementptr with that. It's still
460  // better than ptrtoint+arithmetic+inttoptr at least.
461  if (!AnyNonZeroIndices) {
462    // Cast the base to i8*.
463    V = InsertNoopCastOfTo(V,
464       Type::getInt8Ty(Ty->getContext())->getPointerTo(PTy->getAddressSpace()));
465
466    // Expand the operands for a plain byte offset.
467    Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
468
469    // Fold a GEP with constant operands.
470    if (Constant *CLHS = dyn_cast<Constant>(V))
471      if (Constant *CRHS = dyn_cast<Constant>(Idx))
472        return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
473
474    // Do a quick scan to see if we have this GEP nearby.  If so, reuse it.
475    unsigned ScanLimit = 6;
476    BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
477    // Scanning starts from the last instruction before the insertion point.
478    BasicBlock::iterator IP = Builder.GetInsertPoint();
479    if (IP != BlockBegin) {
480      --IP;
481      for (; ScanLimit; --IP, --ScanLimit) {
482        if (IP->getOpcode() == Instruction::GetElementPtr &&
483            IP->getOperand(0) == V && IP->getOperand(1) == Idx)
484          return IP;
485        if (IP == BlockBegin) break;
486      }
487    }
488
489    // Emit a GEP.
490    Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
491    InsertedValues.insert(GEP);
492    return GEP;
493  }
494
495  // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
496  // because ScalarEvolution may have changed the address arithmetic to
497  // compute a value which is beyond the end of the allocated object.
498  Value *GEP = Builder.CreateGEP(V,
499                                 GepIndices.begin(),
500                                 GepIndices.end(),
501                                 "scevgep");
502  Ops.push_back(SE.getUnknown(GEP));
503  InsertedValues.insert(GEP);
504  return expand(SE.getAddExpr(Ops));
505}
506
507Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
508  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
509  Value *V = expand(S->getOperand(S->getNumOperands()-1));
510
511  // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
512  // comments on expandAddToGEP for details.
513  if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
514    const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
515    return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1], PTy, Ty, V);
516  }
517
518  V = InsertNoopCastOfTo(V, Ty);
519
520  // Emit a bunch of add instructions
521  for (int i = S->getNumOperands()-2; i >= 0; --i) {
522    Value *W = expandCodeFor(S->getOperand(i), Ty);
523    V = InsertBinop(Instruction::Add, V, W);
524  }
525  return V;
526}
527
528Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
529  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
530  int FirstOp = 0;  // Set if we should emit a subtract.
531  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
532    if (SC->getValue()->isAllOnesValue())
533      FirstOp = 1;
534
535  int i = S->getNumOperands()-2;
536  Value *V = expandCodeFor(S->getOperand(i+1), Ty);
537
538  // Emit a bunch of multiply instructions
539  for (; i >= FirstOp; --i) {
540    Value *W = expandCodeFor(S->getOperand(i), Ty);
541    V = InsertBinop(Instruction::Mul, V, W);
542  }
543
544  // -1 * ...  --->  0 - ...
545  if (FirstOp == 1)
546    V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
547  return V;
548}
549
550Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
551  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
552
553  Value *LHS = expandCodeFor(S->getLHS(), Ty);
554  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
555    const APInt &RHS = SC->getValue()->getValue();
556    if (RHS.isPowerOf2())
557      return InsertBinop(Instruction::LShr, LHS,
558                         ConstantInt::get(Ty, RHS.logBase2()));
559  }
560
561  Value *RHS = expandCodeFor(S->getRHS(), Ty);
562  return InsertBinop(Instruction::UDiv, LHS, RHS);
563}
564
565/// Move parts of Base into Rest to leave Base with the minimal
566/// expression that provides a pointer operand suitable for a
567/// GEP expansion.
568static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
569                              ScalarEvolution &SE) {
570  while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
571    Base = A->getStart();
572    Rest = SE.getAddExpr(Rest,
573                         SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
574                                          A->getStepRecurrence(SE),
575                                          A->getLoop()));
576  }
577  if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
578    Base = A->getOperand(A->getNumOperands()-1);
579    SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
580    NewAddOps.back() = Rest;
581    Rest = SE.getAddExpr(NewAddOps);
582    ExposePointerBase(Base, Rest, SE);
583  }
584}
585
586Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
587  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
588  const Loop *L = S->getLoop();
589
590  // First check for an existing canonical IV in a suitable type.
591  PHINode *CanonicalIV = 0;
592  if (PHINode *PN = L->getCanonicalInductionVariable())
593    if (SE.isSCEVable(PN->getType()) &&
594        isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
595        SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
596      CanonicalIV = PN;
597
598  // Rewrite an AddRec in terms of the canonical induction variable, if
599  // its type is more narrow.
600  if (CanonicalIV &&
601      SE.getTypeSizeInBits(CanonicalIV->getType()) >
602      SE.getTypeSizeInBits(Ty)) {
603    const SCEV *Start = SE.getAnyExtendExpr(S->getStart(),
604                                            CanonicalIV->getType());
605    const SCEV *Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
606                                           CanonicalIV->getType());
607    Value *V = expand(SE.getAddRecExpr(Start, Step, S->getLoop()));
608    BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
609    BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
610    BasicBlock::iterator NewInsertPt =
611      next(BasicBlock::iterator(cast<Instruction>(V)));
612    while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
613    V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
614                      NewInsertPt);
615    Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
616    return V;
617  }
618
619  // {X,+,F} --> X + {0,+,F}
620  if (!S->getStart()->isZero()) {
621    const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
622    SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
623    NewOps[0] = SE.getIntegerSCEV(0, Ty);
624    const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
625
626    // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
627    // comments on expandAddToGEP for details.
628    const SCEV *Base = S->getStart();
629    const SCEV *RestArray[1] = { Rest };
630    // Dig into the expression to find the pointer base for a GEP.
631    ExposePointerBase(Base, RestArray[0], SE);
632    // If we found a pointer, expand the AddRec with a GEP.
633    if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
634      // Make sure the Base isn't something exotic, such as a multiplied
635      // or divided pointer value. In those cases, the result type isn't
636      // actually a pointer type.
637      if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
638        Value *StartV = expand(Base);
639        assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
640        return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
641      }
642    }
643
644    // Just do a normal add. Pre-expand the operands to suppress folding.
645    return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
646                                SE.getUnknown(expand(Rest))));
647  }
648
649  // {0,+,1} --> Insert a canonical induction variable into the loop!
650  if (S->isAffine() &&
651      S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
652    // If there's a canonical IV, just use it.
653    if (CanonicalIV) {
654      assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
655             "IVs with types different from the canonical IV should "
656             "already have been handled!");
657      return CanonicalIV;
658    }
659
660    // Create and insert the PHI node for the induction variable in the
661    // specified loop.
662    BasicBlock *Header = L->getHeader();
663    BasicBlock *Preheader = L->getLoopPreheader();
664    PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
665    InsertedValues.insert(PN);
666    PN->addIncoming(Constant::getNullValue(Ty), Preheader);
667
668    pred_iterator HPI = pred_begin(Header);
669    assert(HPI != pred_end(Header) && "Loop with zero preds???");
670    if (!L->contains(*HPI)) ++HPI;
671    assert(HPI != pred_end(Header) && L->contains(*HPI) &&
672           "No backedge in loop?");
673
674    // Insert a unit add instruction right before the terminator corresponding
675    // to the back-edge.
676    Constant *One = ConstantInt::get(Ty, 1);
677    Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
678                                                 (*HPI)->getTerminator());
679    InsertedValues.insert(Add);
680
681    pred_iterator PI = pred_begin(Header);
682    if (*PI == Preheader)
683      ++PI;
684    PN->addIncoming(Add, *PI);
685    return PN;
686  }
687
688  // {0,+,F} --> {0,+,1} * F
689  // Get the canonical induction variable I for this loop.
690  Value *I = CanonicalIV ?
691             CanonicalIV :
692             getOrInsertCanonicalInductionVariable(L, Ty);
693
694  // If this is a simple linear addrec, emit it now as a special case.
695  if (S->isAffine())    // {0,+,F} --> i*F
696    return
697      expand(SE.getTruncateOrNoop(
698        SE.getMulExpr(SE.getUnknown(I),
699                      SE.getNoopOrAnyExtend(S->getOperand(1),
700                                            I->getType())),
701        Ty));
702
703  // If this is a chain of recurrences, turn it into a closed form, using the
704  // folders, then expandCodeFor the closed form.  This allows the folders to
705  // simplify the expression without having to build a bunch of special code
706  // into this folder.
707  const SCEV *IH = SE.getUnknown(I);   // Get I as a "symbolic" SCEV.
708
709  // Promote S up to the canonical IV type, if the cast is foldable.
710  const SCEV *NewS = S;
711  const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
712  if (isa<SCEVAddRecExpr>(Ext))
713    NewS = Ext;
714
715  const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
716  //cerr << "Evaluated: " << *this << "\n     to: " << *V << "\n";
717
718  // Truncate the result down to the original type, if needed.
719  const SCEV *T = SE.getTruncateOrNoop(V, Ty);
720  return expand(T);
721}
722
723Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
724  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
725  Value *V = expandCodeFor(S->getOperand(),
726                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
727  Value *I = Builder.CreateTrunc(V, Ty, "tmp");
728  InsertedValues.insert(I);
729  return I;
730}
731
732Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
733  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
734  Value *V = expandCodeFor(S->getOperand(),
735                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
736  Value *I = Builder.CreateZExt(V, Ty, "tmp");
737  InsertedValues.insert(I);
738  return I;
739}
740
741Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
742  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
743  Value *V = expandCodeFor(S->getOperand(),
744                           SE.getEffectiveSCEVType(S->getOperand()->getType()));
745  Value *I = Builder.CreateSExt(V, Ty, "tmp");
746  InsertedValues.insert(I);
747  return I;
748}
749
750Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
751  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
752  const Type *Ty = LHS->getType();
753  for (int i = S->getNumOperands()-2; i >= 0; --i) {
754    // In the case of mixed integer and pointer types, do the
755    // rest of the comparisons as integer.
756    if (S->getOperand(i)->getType() != Ty) {
757      Ty = SE.getEffectiveSCEVType(Ty);
758      LHS = InsertNoopCastOfTo(LHS, Ty);
759    }
760    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
761    Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
762    InsertedValues.insert(ICmp);
763    Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
764    InsertedValues.insert(Sel);
765    LHS = Sel;
766  }
767  // In the case of mixed integer and pointer types, cast the
768  // final result back to the pointer type.
769  if (LHS->getType() != S->getType())
770    LHS = InsertNoopCastOfTo(LHS, S->getType());
771  return LHS;
772}
773
774Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
775  Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
776  const Type *Ty = LHS->getType();
777  for (int i = S->getNumOperands()-2; i >= 0; --i) {
778    // In the case of mixed integer and pointer types, do the
779    // rest of the comparisons as integer.
780    if (S->getOperand(i)->getType() != Ty) {
781      Ty = SE.getEffectiveSCEVType(Ty);
782      LHS = InsertNoopCastOfTo(LHS, Ty);
783    }
784    Value *RHS = expandCodeFor(S->getOperand(i), Ty);
785    Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
786    InsertedValues.insert(ICmp);
787    Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
788    InsertedValues.insert(Sel);
789    LHS = Sel;
790  }
791  // In the case of mixed integer and pointer types, cast the
792  // final result back to the pointer type.
793  if (LHS->getType() != S->getType())
794    LHS = InsertNoopCastOfTo(LHS, S->getType());
795  return LHS;
796}
797
798Value *SCEVExpander::visitFieldOffsetExpr(const SCEVFieldOffsetExpr *S) {
799  return ConstantExpr::getOffsetOf(S->getStructType(), S->getFieldNo());
800}
801
802Value *SCEVExpander::visitAllocSizeExpr(const SCEVAllocSizeExpr *S) {
803  return ConstantExpr::getSizeOf(S->getAllocType());
804}
805
806Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
807  // Expand the code for this SCEV.
808  Value *V = expand(SH);
809  if (Ty) {
810    assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
811           "non-trivial casts should be done with the SCEVs directly!");
812    V = InsertNoopCastOfTo(V, Ty);
813  }
814  return V;
815}
816
817Value *SCEVExpander::expand(const SCEV *S) {
818  // Compute an insertion point for this SCEV object. Hoist the instructions
819  // as far out in the loop nest as possible.
820  Instruction *InsertPt = Builder.GetInsertPoint();
821  for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
822       L = L->getParentLoop())
823    if (S->isLoopInvariant(L)) {
824      if (!L) break;
825      if (BasicBlock *Preheader = L->getLoopPreheader())
826        InsertPt = Preheader->getTerminator();
827    } else {
828      // If the SCEV is computable at this level, insert it into the header
829      // after the PHIs (and after any other instructions that we've inserted
830      // there) so that it is guaranteed to dominate any user inside the loop.
831      if (L && S->hasComputableLoopEvolution(L))
832        InsertPt = L->getHeader()->getFirstNonPHI();
833      while (isInsertedInstruction(InsertPt))
834        InsertPt = next(BasicBlock::iterator(InsertPt));
835      break;
836    }
837
838  // Check to see if we already expanded this here.
839  std::map<std::pair<const SCEV *, Instruction *>,
840           AssertingVH<Value> >::iterator I =
841    InsertedExpressions.find(std::make_pair(S, InsertPt));
842  if (I != InsertedExpressions.end())
843    return I->second;
844
845  BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
846  BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
847  Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
848
849  // Expand the expression into instructions.
850  Value *V = visit(S);
851
852  // Remember the expanded value for this SCEV at this location.
853  InsertedExpressions[std::make_pair(S, InsertPt)] = V;
854
855  Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
856  return V;
857}
858
859/// getOrInsertCanonicalInductionVariable - This method returns the
860/// canonical induction variable of the specified type for the specified
861/// loop (inserting one if there is none).  A canonical induction variable
862/// starts at zero and steps by one on each iteration.
863Value *
864SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
865                                                    const Type *Ty) {
866  assert(Ty->isInteger() && "Can only insert integer induction variables!");
867  const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
868                                   SE.getIntegerSCEV(1, Ty), L);
869  BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
870  BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
871  Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
872  if (SaveInsertBB)
873    Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
874  return V;
875}
876