ScalarEvolutionExpander.cpp revision 278b49af8a08f6ab6c486a3cfc7a9c1c1acd2b23 
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"
19using namespace llvm;
20
21/// InsertCastOfTo - Insert a cast of V to the specified type, doing what
22/// we can to share the casts.
23Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
24                                    const Type *Ty) {
25  // Short-circuit unnecessary bitcasts.
26  if (opcode == Instruction::BitCast && V->getType() == Ty)
27    return V;
28
29  // Short-circuit unnecessary inttoptr<->ptrtoint casts.
30  if ((opcode == Instruction::PtrToInt || opcode == Instruction::IntToPtr) &&
31      SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
32    if (CastInst *CI = dyn_cast<CastInst>(V))
33      if ((CI->getOpcode() == Instruction::PtrToInt ||
34           CI->getOpcode() == Instruction::IntToPtr) &&
35          SE.getTypeSizeInBits(CI->getType()) ==
36          SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
37        return CI->getOperand(0);
38    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
39      if ((CE->getOpcode() == Instruction::PtrToInt ||
40           CE->getOpcode() == Instruction::IntToPtr) &&
41          SE.getTypeSizeInBits(CE->getType()) ==
42          SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
43        return CE->getOperand(0);
44  }
45
46  // FIXME: keep track of the cast instruction.
47  if (Constant *C = dyn_cast<Constant>(V))
48    return ConstantExpr::getCast(opcode, C, Ty);
49
50  if (Argument *A = dyn_cast<Argument>(V)) {
51    // Check to see if there is already a cast!
52    for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
53         UI != E; ++UI) {
54      if ((*UI)->getType() == Ty)
55        if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
56          if (CI->getOpcode() == opcode) {
57            // If the cast isn't the first instruction of the function, move it.
58            if (BasicBlock::iterator(CI) !=
59                A->getParent()->getEntryBlock().begin()) {
60              // If the CastInst is the insert point, change the insert point.
61              if (CI == InsertPt) ++InsertPt;
62              // Splice the cast at the beginning of the entry block.
63              CI->moveBefore(A->getParent()->getEntryBlock().begin());
64            }
65            return CI;
66          }
67    }
68    Instruction *I = CastInst::Create(opcode, V, Ty, V->getName(),
69                                      A->getParent()->getEntryBlock().begin());
70    InsertedValues.insert(I);
71    return I;
72  }
73
74  Instruction *I = cast<Instruction>(V);
75
76  // Check to see if there is already a cast.  If there is, use it.
77  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
78       UI != E; ++UI) {
79    if ((*UI)->getType() == Ty)
80      if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
81        if (CI->getOpcode() == opcode) {
82          BasicBlock::iterator It = I; ++It;
83          if (isa<InvokeInst>(I))
84            It = cast<InvokeInst>(I)->getNormalDest()->begin();
85          while (isa<PHINode>(It)) ++It;
86          if (It != BasicBlock::iterator(CI)) {
87            // If the CastInst is the insert point, change the insert point.
88            if (CI == InsertPt) ++InsertPt;
89            // Splice the cast immediately after the operand in question.
90            CI->moveBefore(It);
91          }
92          return CI;
93        }
94  }
95  BasicBlock::iterator IP = I; ++IP;
96  if (InvokeInst *II = dyn_cast<InvokeInst>(I))
97    IP = II->getNormalDest()->begin();
98  while (isa<PHINode>(IP)) ++IP;
99  Instruction *CI = CastInst::Create(opcode, V, Ty, V->getName(), IP);
100  InsertedValues.insert(CI);
101  return CI;
102}
103
104/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
105/// which must be possible with a noop cast.
106Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
107  Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
108  assert((Op == Instruction::BitCast ||
109          Op == Instruction::PtrToInt ||
110          Op == Instruction::IntToPtr) &&
111         "InsertNoopCastOfTo cannot perform non-noop casts!");
112  assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
113         "InsertNoopCastOfTo cannot change sizes!");
114  return InsertCastOfTo(Op, V, Ty);
115}
116
117/// InsertBinop - Insert the specified binary operator, doing a small amount
118/// of work to avoid inserting an obviously redundant operation.
119Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS,
120                                 Value *RHS, BasicBlock::iterator InsertPt) {
121  // Fold a binop with constant operands.
122  if (Constant *CLHS = dyn_cast<Constant>(LHS))
123    if (Constant *CRHS = dyn_cast<Constant>(RHS))
124      return ConstantExpr::get(Opcode, CLHS, CRHS);
125
126  // Do a quick scan to see if we have this binop nearby.  If so, reuse it.
127  unsigned ScanLimit = 6;
128  BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
129  if (InsertPt != BlockBegin) {
130    // Scanning starts from the last instruction before InsertPt.
131    BasicBlock::iterator IP = InsertPt;
132    --IP;
133    for (; ScanLimit; --IP, --ScanLimit) {
134      if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
135          IP->getOperand(1) == RHS)
136        return IP;
137      if (IP == BlockBegin) break;
138    }
139  }
140
141  // If we haven't found this binop, insert it.
142  Instruction *BO = BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt);
143  InsertedValues.insert(BO);
144  return BO;
145}
146
147/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP
148/// instead of using ptrtoint+arithmetic+inttoptr.
149Value *SCEVExpander::expandAddToGEP(const SCEVAddExpr *S,
150                                    const PointerType *PTy,
151                                    const Type *Ty,
152                                    Value *V) {
153  const Type *ElTy = PTy->getElementType();
154  SmallVector<Value *, 4> GepIndices;
155  std::vector<SCEVHandle> Ops = S->getOperands();
156  bool AnyNonZeroIndices = false;
157  Ops.pop_back();
158
159  // Decend down the pointer's type and attempt to convert the other
160  // operands into GEP indices, at each level. The first index in a GEP
161  // indexes into the array implied by the pointer operand; the rest of
162  // the indices index into the element or field type selected by the
163  // preceding index.
164  for (;;) {
165    APInt ElSize = APInt(SE.getTypeSizeInBits(Ty),
166                         ElTy->isSized() ?  SE.TD->getTypeAllocSize(ElTy) : 0);
167    std::vector<SCEVHandle> NewOps;
168    std::vector<SCEVHandle> ScaledOps;
169    for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
170      if (ElSize != 0) {
171        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i]))
172          if (!C->getValue()->getValue().srem(ElSize)) {
173            ConstantInt *CI =
174              ConstantInt::get(C->getValue()->getValue().sdiv(ElSize));
175            SCEVHandle Div = SE.getConstant(CI);
176            ScaledOps.push_back(Div);
177            continue;
178          }
179        if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i]))
180          if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
181            if (C->getValue()->getValue() == ElSize) {
182              for (unsigned j = 1, f = M->getNumOperands(); j != f; ++j)
183                ScaledOps.push_back(M->getOperand(j));
184              continue;
185            }
186        if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i]))
187          if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getValue()))
188            if (BO->getOpcode() == Instruction::Mul)
189              if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1)))
190                if (CI->getValue() == ElSize) {
191                  ScaledOps.push_back(SE.getUnknown(BO->getOperand(0)));
192                  continue;
193                }
194        if (ElSize == 1) {
195          ScaledOps.push_back(Ops[i]);
196          continue;
197        }
198      }
199      NewOps.push_back(Ops[i]);
200    }
201    Ops = NewOps;
202    AnyNonZeroIndices |= !ScaledOps.empty();
203    Value *Scaled = ScaledOps.empty() ?
204                    Constant::getNullValue(Ty) :
205                    expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
206    GepIndices.push_back(Scaled);
207
208    // Collect struct field index operands.
209    if (!Ops.empty())
210      while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
211        if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
212          if (SE.getTypeSizeInBits(C->getType()) <= 64) {
213            const StructLayout &SL = *SE.TD->getStructLayout(STy);
214            uint64_t FullOffset = C->getValue()->getZExtValue();
215            if (FullOffset < SL.getSizeInBytes()) {
216              unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
217              GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx));
218              ElTy = STy->getTypeAtIndex(ElIdx);
219              Ops[0] =
220                SE.getConstant(ConstantInt::get(Ty,
221                                                FullOffset -
222                                                  SL.getElementOffset(ElIdx)));
223              AnyNonZeroIndices = true;
224              continue;
225            }
226          }
227        break;
228      }
229
230    if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) {
231      ElTy = ATy->getElementType();
232      continue;
233    }
234    break;
235  }
236
237  // If none of the operands were convertable to proper GEP indices, cast
238  // the base to i8* and do an ugly getelementptr with that. It's still
239  // better than ptrtoint+arithmetic+inttoptr at least.
240  if (!AnyNonZeroIndices) {
241    V = InsertNoopCastOfTo(V,
242                           Type::Int8Ty->getPointerTo(PTy->getAddressSpace()));
243    Value *Idx = expand(SE.getAddExpr(Ops));
244    Idx = InsertNoopCastOfTo(Idx, Ty);
245
246    // Fold a GEP with constant operands.
247    if (Constant *CLHS = dyn_cast<Constant>(V))
248      if (Constant *CRHS = dyn_cast<Constant>(Idx))
249        return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
250
251    // Do a quick scan to see if we have this GEP nearby.  If so, reuse it.
252    unsigned ScanLimit = 6;
253    BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin();
254    if (InsertPt != BlockBegin) {
255      // Scanning starts from the last instruction before InsertPt.
256      BasicBlock::iterator IP = InsertPt;
257      --IP;
258      for (; ScanLimit; --IP, --ScanLimit) {
259        if (IP->getOpcode() == Instruction::GetElementPtr &&
260            IP->getOperand(0) == V && IP->getOperand(1) == Idx)
261          return IP;
262        if (IP == BlockBegin) break;
263      }
264    }
265
266    Value *GEP = GetElementPtrInst::Create(V, Idx, "scevgep", InsertPt);
267    InsertedValues.insert(GEP);
268    return GEP;
269  }
270
271  // Insert a pretty getelementptr.
272  Value *GEP = GetElementPtrInst::Create(V,
273                                         GepIndices.begin(),
274                                         GepIndices.end(),
275                                         "scevgep", InsertPt);
276  Ops.push_back(SE.getUnknown(GEP));
277  InsertedValues.insert(GEP);
278  return expand(SE.getAddExpr(Ops));
279}
280
281Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
282  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
283  Value *V = expand(S->getOperand(S->getNumOperands()-1));
284
285  // Turn things like ptrtoint+arithmetic+inttoptr into GEP. This helps
286  // BasicAliasAnalysis analyze the result. However, it suffers from the
287  // underlying bug described in PR2831. Addition in LLVM currently always
288  // has two's complement wrapping guaranteed. However, the semantics for
289  // getelementptr overflow are ambiguous. In the common case though, this
290  // expansion gets used when a GEP in the original code has been converted
291  // into integer arithmetic, in which case the resulting code will be no
292  // more undefined than it was originally.
293  if (SE.TD)
294    if (const PointerType *PTy = dyn_cast<PointerType>(V->getType()))
295      return expandAddToGEP(S, PTy, Ty, V);
296
297  V = InsertNoopCastOfTo(V, Ty);
298
299  // Emit a bunch of add instructions
300  for (int i = S->getNumOperands()-2; i >= 0; --i) {
301    Value *W = expand(S->getOperand(i));
302    W = InsertNoopCastOfTo(W, Ty);
303    V = InsertBinop(Instruction::Add, V, W, InsertPt);
304  }
305  return V;
306}
307
308Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
309  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
310  int FirstOp = 0;  // Set if we should emit a subtract.
311  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
312    if (SC->getValue()->isAllOnesValue())
313      FirstOp = 1;
314
315  int i = S->getNumOperands()-2;
316  Value *V = expand(S->getOperand(i+1));
317  V = InsertNoopCastOfTo(V, Ty);
318
319  // Emit a bunch of multiply instructions
320  for (; i >= FirstOp; --i) {
321    Value *W = expand(S->getOperand(i));
322    W = InsertNoopCastOfTo(W, Ty);
323    V = InsertBinop(Instruction::Mul, V, W, InsertPt);
324  }
325
326  // -1 * ...  --->  0 - ...
327  if (FirstOp == 1)
328    V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V, InsertPt);
329  return V;
330}
331
332Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
333  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
334
335  Value *LHS = expand(S->getLHS());
336  LHS = InsertNoopCastOfTo(LHS, Ty);
337  if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
338    const APInt &RHS = SC->getValue()->getValue();
339    if (RHS.isPowerOf2())
340      return InsertBinop(Instruction::LShr, LHS,
341                         ConstantInt::get(Ty, RHS.logBase2()),
342                         InsertPt);
343  }
344
345  Value *RHS = expand(S->getRHS());
346  RHS = InsertNoopCastOfTo(RHS, Ty);
347  return InsertBinop(Instruction::UDiv, LHS, RHS, InsertPt);
348}
349
350Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
351  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
352  const Loop *L = S->getLoop();
353
354  // {X,+,F} --> X + {0,+,F}
355  if (!S->getStart()->isZero()) {
356    std::vector<SCEVHandle> NewOps(S->getOperands());
357    NewOps[0] = SE.getIntegerSCEV(0, Ty);
358    Value *Rest = expand(SE.getAddRecExpr(NewOps, L));
359    return expand(SE.getAddExpr(S->getStart(), SE.getUnknown(Rest)));
360  }
361
362  // {0,+,1} --> Insert a canonical induction variable into the loop!
363  if (S->isAffine() &&
364      S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
365    // Create and insert the PHI node for the induction variable in the
366    // specified loop.
367    BasicBlock *Header = L->getHeader();
368    PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
369    InsertedValues.insert(PN);
370    PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
371
372    pred_iterator HPI = pred_begin(Header);
373    assert(HPI != pred_end(Header) && "Loop with zero preds???");
374    if (!L->contains(*HPI)) ++HPI;
375    assert(HPI != pred_end(Header) && L->contains(*HPI) &&
376           "No backedge in loop?");
377
378    // Insert a unit add instruction right before the terminator corresponding
379    // to the back-edge.
380    Constant *One = ConstantInt::get(Ty, 1);
381    Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
382                                                 (*HPI)->getTerminator());
383    InsertedValues.insert(Add);
384
385    pred_iterator PI = pred_begin(Header);
386    if (*PI == L->getLoopPreheader())
387      ++PI;
388    PN->addIncoming(Add, *PI);
389    return PN;
390  }
391
392  // Get the canonical induction variable I for this loop.
393  Value *I = getOrInsertCanonicalInductionVariable(L, Ty);
394
395  // If this is a simple linear addrec, emit it now as a special case.
396  if (S->isAffine()) {   // {0,+,F} --> i*F
397    Value *F = expand(S->getOperand(1));
398    F = InsertNoopCastOfTo(F, Ty);
399
400    // IF the step is by one, just return the inserted IV.
401    if (ConstantInt *CI = dyn_cast<ConstantInt>(F))
402      if (CI->getValue() == 1)
403        return I;
404
405    // If the insert point is directly inside of the loop, emit the multiply at
406    // the insert point.  Otherwise, L is a loop that is a parent of the insert
407    // point loop.  If we can, move the multiply to the outer most loop that it
408    // is safe to be in.
409    BasicBlock::iterator MulInsertPt = getInsertionPoint();
410    Loop *InsertPtLoop = SE.LI->getLoopFor(MulInsertPt->getParent());
411    if (InsertPtLoop != L && InsertPtLoop &&
412        L->contains(InsertPtLoop->getHeader())) {
413      do {
414        // If we cannot hoist the multiply out of this loop, don't.
415        if (!InsertPtLoop->isLoopInvariant(F)) break;
416
417        BasicBlock *InsertPtLoopPH = InsertPtLoop->getLoopPreheader();
418
419        // If this loop hasn't got a preheader, we aren't able to hoist the
420        // multiply.
421        if (!InsertPtLoopPH)
422          break;
423
424        // Otherwise, move the insert point to the preheader.
425        MulInsertPt = InsertPtLoopPH->getTerminator();
426        InsertPtLoop = InsertPtLoop->getParentLoop();
427      } while (InsertPtLoop != L);
428    }
429
430    return InsertBinop(Instruction::Mul, I, F, MulInsertPt);
431  }
432
433  // If this is a chain of recurrences, turn it into a closed form, using the
434  // folders, then expandCodeFor the closed form.  This allows the folders to
435  // simplify the expression without having to build a bunch of special code
436  // into this folder.
437  SCEVHandle IH = SE.getUnknown(I);   // Get I as a "symbolic" SCEV.
438
439  SCEVHandle V = S->evaluateAtIteration(IH, SE);
440  //cerr << "Evaluated: " << *this << "\n     to: " << *V << "\n";
441
442  return expand(V);
443}
444
445Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
446  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
447  Value *V = expand(S->getOperand());
448  V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType()));
449  Instruction *I = new TruncInst(V, Ty, "tmp.", InsertPt);
450  InsertedValues.insert(I);
451  return I;
452}
453
454Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
455  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
456  Value *V = expand(S->getOperand());
457  V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType()));
458  Instruction *I = new ZExtInst(V, Ty, "tmp.", InsertPt);
459  InsertedValues.insert(I);
460  return I;
461}
462
463Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
464  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
465  Value *V = expand(S->getOperand());
466  V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType()));
467  Instruction *I = new SExtInst(V, Ty, "tmp.", InsertPt);
468  InsertedValues.insert(I);
469  return I;
470}
471
472Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
473  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
474  Value *LHS = expand(S->getOperand(0));
475  LHS = InsertNoopCastOfTo(LHS, Ty);
476  for (unsigned i = 1; i < S->getNumOperands(); ++i) {
477    Value *RHS = expand(S->getOperand(i));
478    RHS = InsertNoopCastOfTo(RHS, Ty);
479    Instruction *ICmp =
480      new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt);
481    InsertedValues.insert(ICmp);
482    Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt);
483    InsertedValues.insert(Sel);
484    LHS = Sel;
485  }
486  return LHS;
487}
488
489Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
490  const Type *Ty = SE.getEffectiveSCEVType(S->getType());
491  Value *LHS = expand(S->getOperand(0));
492  LHS = InsertNoopCastOfTo(LHS, Ty);
493  for (unsigned i = 1; i < S->getNumOperands(); ++i) {
494    Value *RHS = expand(S->getOperand(i));
495    RHS = InsertNoopCastOfTo(RHS, Ty);
496    Instruction *ICmp =
497      new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt);
498    InsertedValues.insert(ICmp);
499    Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt);
500    InsertedValues.insert(Sel);
501    LHS = Sel;
502  }
503  return LHS;
504}
505
506Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty) {
507  // Expand the code for this SCEV.
508  Value *V = expand(SH);
509  if (Ty) {
510    assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
511           "non-trivial casts should be done with the SCEVs directly!");
512    V = InsertNoopCastOfTo(V, Ty);
513  }
514  return V;
515}
516
517Value *SCEVExpander::expand(const SCEV *S) {
518  // Check to see if we already expanded this.
519  std::map<SCEVHandle, Value*>::iterator I = InsertedExpressions.find(S);
520  if (I != InsertedExpressions.end())
521    return I->second;
522
523  Value *V = visit(S);
524  InsertedExpressions[S] = V;
525  return V;
526}
527