ConstantFolding.cpp revision ecc0274a54658a4ee76d783b285a0690ad760928 
1//===-- ConstantFolding.cpp - Analyze constant folding possibilities ------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This family of functions determines the possibility of performing constant
11// folding.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Analysis/ConstantFolding.h"
16#include "llvm/Constants.h"
17#include "llvm/DerivedTypes.h"
18#include "llvm/Function.h"
19#include "llvm/Instructions.h"
20#include "llvm/Intrinsics.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/ADT/StringMap.h"
23#include "llvm/Target/TargetData.h"
24#include "llvm/Support/GetElementPtrTypeIterator.h"
25#include "llvm/Support/MathExtras.h"
26#include <cerrno>
27#include <cmath>
28using namespace llvm;
29
30//===----------------------------------------------------------------------===//
31// Constant Folding internal helper functions
32//===----------------------------------------------------------------------===//
33
34/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
35/// from a global, return the global and the constant.  Because of
36/// constantexprs, this function is recursive.
37static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
38                                       int64_t &Offset, const TargetData &TD) {
39  // Trivial case, constant is the global.
40  if ((GV = dyn_cast<GlobalValue>(C))) {
41    Offset = 0;
42    return true;
43  }
44
45  // Otherwise, if this isn't a constant expr, bail out.
46  ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
47  if (!CE) return false;
48
49  // Look through ptr->int and ptr->ptr casts.
50  if (CE->getOpcode() == Instruction::PtrToInt ||
51      CE->getOpcode() == Instruction::BitCast)
52    return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
53
54  // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
55  if (CE->getOpcode() == Instruction::GetElementPtr) {
56    // Cannot compute this if the element type of the pointer is missing size
57    // info.
58    if (!cast<PointerType>(CE->getOperand(0)->getType())->getElementType()->isSized())
59      return false;
60
61    // If the base isn't a global+constant, we aren't either.
62    if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
63      return false;
64
65    // Otherwise, add any offset that our operands provide.
66    gep_type_iterator GTI = gep_type_begin(CE);
67    for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i, ++GTI) {
68      ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(i));
69      if (!CI) return false;  // Index isn't a simple constant?
70      if (CI->getZExtValue() == 0) continue;  // Not adding anything.
71
72      if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
73        // N = N + Offset
74        Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
75      } else {
76        const SequentialType *SQT = cast<SequentialType>(*GTI);
77        Offset += TD.getABITypeSize(SQT->getElementType())*CI->getSExtValue();
78      }
79    }
80    return true;
81  }
82
83  return false;
84}
85
86
87/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
88/// Attempt to symbolically evaluate the result of  a binary operator merging
89/// these together.  If target data info is available, it is provided as TD,
90/// otherwise TD is null.
91static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
92                                           Constant *Op1, const TargetData *TD){
93  // SROA
94
95  // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
96  // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
97  // bits.
98
99
100  // If the constant expr is something like &A[123] - &A[4].f, fold this into a
101  // constant.  This happens frequently when iterating over a global array.
102  if (Opc == Instruction::Sub && TD) {
103    GlobalValue *GV1, *GV2;
104    int64_t Offs1, Offs2;
105
106    if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
107      if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
108          GV1 == GV2) {
109        // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
110        return ConstantInt::get(Op0->getType(), Offs1-Offs2);
111      }
112  }
113
114  // TODO: Fold icmp setne/seteq as well.
115  return 0;
116}
117
118/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
119/// constant expression, do so.
120static Constant *SymbolicallyEvaluateGEP(Constant** Ops, unsigned NumOps,
121                                         const Type *ResultTy,
122                                         const TargetData *TD) {
123  Constant *Ptr = Ops[0];
124  if (!cast<PointerType>(Ptr->getType())->getElementType()->isSized())
125    return 0;
126
127  if (TD && Ptr->isNullValue()) {
128    // If this is a constant expr gep that is effectively computing an
129    // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
130    bool isFoldableGEP = true;
131    for (unsigned i = 1; i != NumOps; ++i)
132      if (!isa<ConstantInt>(Ops[i])) {
133        isFoldableGEP = false;
134        break;
135      }
136    if (isFoldableGEP) {
137      uint64_t Offset = TD->getIndexedOffset(Ptr->getType(),
138                                             (Value**)Ops+1, NumOps-1);
139      Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
140      return ConstantExpr::getIntToPtr(C, ResultTy);
141    }
142  }
143
144  return 0;
145}
146
147
148//===----------------------------------------------------------------------===//
149// Constant Folding public APIs
150//===----------------------------------------------------------------------===//
151
152
153/// ConstantFoldInstruction - Attempt to constant fold the specified
154/// instruction.  If successful, the constant result is returned, if not, null
155/// is returned.  Note that this function can only fail when attempting to fold
156/// instructions like loads and stores, which have no constant expression form.
157///
158Constant *llvm::ConstantFoldInstruction(Instruction *I, const TargetData *TD) {
159  if (PHINode *PN = dyn_cast<PHINode>(I)) {
160    if (PN->getNumIncomingValues() == 0)
161      return Constant::getNullValue(PN->getType());
162
163    Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
164    if (Result == 0) return 0;
165
166    // Handle PHI nodes specially here...
167    for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
168      if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
169        return 0;   // Not all the same incoming constants...
170
171    // If we reach here, all incoming values are the same constant.
172    return Result;
173  }
174
175  // Scan the operand list, checking to see if they are all constants, if so,
176  // hand off to ConstantFoldInstOperands.
177  SmallVector<Constant*, 8> Ops;
178  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
179    if (Constant *Op = dyn_cast<Constant>(I->getOperand(i)))
180      Ops.push_back(Op);
181    else
182      return 0;  // All operands not constant!
183
184  return ConstantFoldInstOperands(I, &Ops[0], Ops.size(), TD);
185}
186
187/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
188/// specified opcode and operands.  If successful, the constant result is
189/// returned, if not, null is returned.  Note that this function can fail when
190/// attempting to fold instructions like loads and stores, which have no
191/// constant expression form.
192///
193Constant *llvm::ConstantFoldInstOperands(const Instruction* I,
194                                         Constant** Ops, unsigned NumOps,
195                                         const TargetData *TD) {
196  unsigned Opc = I->getOpcode();
197  const Type *DestTy = I->getType();
198
199  // Handle easy binops first.
200  if (isa<BinaryOperator>(I)) {
201    if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
202      if (Constant *C = SymbolicallyEvaluateBinop(I->getOpcode(), Ops[0],
203                                                  Ops[1], TD))
204        return C;
205
206    return ConstantExpr::get(Opc, Ops[0], Ops[1]);
207  }
208
209  switch (Opc) {
210  default: return 0;
211  case Instruction::Call:
212    if (Function *F = dyn_cast<Function>(Ops[0]))
213      if (canConstantFoldCallTo(F))
214        return ConstantFoldCall(F, Ops+1, NumOps-1);
215    return 0;
216  case Instruction::ICmp:
217  case Instruction::FCmp:
218    return ConstantExpr::getCompare(cast<CmpInst>(I)->getPredicate(), Ops[0],
219                                    Ops[1]);
220  case Instruction::PtrToInt:
221    // If the input is a inttoptr, eliminate the pair.  This requires knowing
222    // the width of a pointer, so it can't be done in ConstantExpr::getCast.
223    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
224      if (TD && CE->getOpcode() == Instruction::IntToPtr) {
225        Constant *Input = CE->getOperand(0);
226        unsigned InWidth = Input->getType()->getPrimitiveSizeInBits();
227        Constant *Mask =
228          ConstantInt::get(APInt::getLowBitsSet(InWidth,
229                                                TD->getPointerSizeInBits()));
230        Input = ConstantExpr::getAnd(Input, Mask);
231        // Do a zext or trunc to get to the dest size.
232        return ConstantExpr::getIntegerCast(Input, I->getType(), false);
233      }
234    }
235    // FALL THROUGH.
236  case Instruction::IntToPtr:
237  case Instruction::Trunc:
238  case Instruction::ZExt:
239  case Instruction::SExt:
240  case Instruction::FPTrunc:
241  case Instruction::FPExt:
242  case Instruction::UIToFP:
243  case Instruction::SIToFP:
244  case Instruction::FPToUI:
245  case Instruction::FPToSI:
246  case Instruction::BitCast:
247    return ConstantExpr::getCast(Opc, Ops[0], DestTy);
248  case Instruction::Select:
249    return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
250  case Instruction::ExtractElement:
251    return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
252  case Instruction::InsertElement:
253    return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
254  case Instruction::ShuffleVector:
255    return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
256  case Instruction::GetElementPtr:
257    if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, I->getType(), TD))
258      return C;
259
260    return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
261  }
262}
263
264/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
265/// getelementptr constantexpr, return the constant value being addressed by the
266/// constant expression, or null if something is funny and we can't decide.
267Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
268                                                       ConstantExpr *CE) {
269  if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
270    return 0;  // Do not allow stepping over the value!
271
272  // Loop over all of the operands, tracking down which value we are
273  // addressing...
274  gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
275  for (++I; I != E; ++I)
276    if (const StructType *STy = dyn_cast<StructType>(*I)) {
277      ConstantInt *CU = cast<ConstantInt>(I.getOperand());
278      assert(CU->getZExtValue() < STy->getNumElements() &&
279             "Struct index out of range!");
280      unsigned El = (unsigned)CU->getZExtValue();
281      if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
282        C = CS->getOperand(El);
283      } else if (isa<ConstantAggregateZero>(C)) {
284        C = Constant::getNullValue(STy->getElementType(El));
285      } else if (isa<UndefValue>(C)) {
286        C = UndefValue::get(STy->getElementType(El));
287      } else {
288        return 0;
289      }
290    } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
291      if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
292        if (CI->getZExtValue() >= ATy->getNumElements())
293         return 0;
294        if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
295          C = CA->getOperand(CI->getZExtValue());
296        else if (isa<ConstantAggregateZero>(C))
297          C = Constant::getNullValue(ATy->getElementType());
298        else if (isa<UndefValue>(C))
299          C = UndefValue::get(ATy->getElementType());
300        else
301          return 0;
302      } else if (const VectorType *PTy = dyn_cast<VectorType>(*I)) {
303        if (CI->getZExtValue() >= PTy->getNumElements())
304          return 0;
305        if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
306          C = CP->getOperand(CI->getZExtValue());
307        else if (isa<ConstantAggregateZero>(C))
308          C = Constant::getNullValue(PTy->getElementType());
309        else if (isa<UndefValue>(C))
310          C = UndefValue::get(PTy->getElementType());
311        else
312          return 0;
313      } else {
314        return 0;
315      }
316    } else {
317      return 0;
318    }
319  return C;
320}
321
322
323//===----------------------------------------------------------------------===//
324//  Constant Folding for Calls
325//
326
327/// canConstantFoldCallTo - Return true if its even possible to fold a call to
328/// the specified function.
329bool
330llvm::canConstantFoldCallTo(Function *F) {
331  switch (F->getIntrinsicID()) {
332  case Intrinsic::sqrt:
333  case Intrinsic::powi:
334  case Intrinsic::bswap:
335  case Intrinsic::ctpop:
336  case Intrinsic::ctlz:
337  case Intrinsic::cttz:
338    return true;
339  default: break;
340  }
341
342  const ValueName *NameVal = F->getValueName();
343  if (NameVal == 0) return false;
344  const char *Str = NameVal->getKeyData();
345  unsigned Len = NameVal->getKeyLength();
346
347  // In these cases, the check of the length is required.  We don't want to
348  // return true for a name like "cos\0blah" which strcmp would return equal to
349  // "cos", but has length 8.
350  switch (Str[0]) {
351  default: return false;
352  case 'a':
353    if (Len == 4)
354      return !strcmp(Str, "acos") || !strcmp(Str, "asin") ||
355             !strcmp(Str, "atan");
356    else if (Len == 5)
357      return !strcmp(Str, "atan2");
358    return false;
359  case 'c':
360    if (Len == 3)
361      return !strcmp(Str, "cos");
362    else if (Len == 4)
363      return !strcmp(Str, "ceil") || !strcmp(Str, "cosf") ||
364             !strcmp(Str, "cosh");
365    return false;
366  case 'e':
367    if (Len == 3)
368      return !strcmp(Str, "exp");
369    return false;
370  case 'f':
371    if (Len == 4)
372      return !strcmp(Str, "fabs") || !strcmp(Str, "fmod");
373    else if (Len == 5)
374      return !strcmp(Str, "floor");
375    return false;
376    break;
377  case 'l':
378    if (Len == 3 && !strcmp(Str, "log"))
379      return true;
380    if (Len == 5 && !strcmp(Str, "log10"))
381      return true;
382    return false;
383  case 'p':
384    if (Len == 3 && !strcmp(Str, "pow"))
385      return true;
386    return false;
387  case 's':
388    if (Len == 3)
389      return !strcmp(Str, "sin");
390    if (Len == 4)
391      return !strcmp(Str, "sinh") || !strcmp(Str, "sqrt");
392    if (Len == 5)
393      return !strcmp(Str, "sqrtf");
394    return false;
395  case 't':
396    if (Len == 3 && !strcmp(Str, "tan"))
397      return true;
398    else if (Len == 4 && !strcmp(Str, "tanh"))
399      return true;
400    return false;
401  }
402}
403
404static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
405                                const Type *Ty) {
406  errno = 0;
407  V = NativeFP(V);
408  if (errno == 0) {
409    if (Ty==Type::FloatTy)
410      return ConstantFP::get(Ty, APFloat((float)V));
411    else if (Ty==Type::DoubleTy)
412      return ConstantFP::get(Ty, APFloat(V));
413    else
414      assert(0);
415  }
416  errno = 0;
417  return 0;
418}
419
420static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
421                                      double V, double W,
422                                      const Type *Ty) {
423  errno = 0;
424  V = NativeFP(V, W);
425  if (errno == 0) {
426    if (Ty==Type::FloatTy)
427      return ConstantFP::get(Ty, APFloat((float)V));
428    else if (Ty==Type::DoubleTy)
429      return ConstantFP::get(Ty, APFloat(V));
430    else
431      assert(0);
432  }
433  errno = 0;
434  return 0;
435}
436
437/// ConstantFoldCall - Attempt to constant fold a call to the specified function
438/// with the specified arguments, returning null if unsuccessful.
439
440Constant *
441llvm::ConstantFoldCall(Function *F, Constant** Operands, unsigned NumOperands) {
442  const ValueName *NameVal = F->getValueName();
443  if (NameVal == 0) return 0;
444  const char *Str = NameVal->getKeyData();
445  unsigned Len = NameVal->getKeyLength();
446
447  const Type *Ty = F->getReturnType();
448  if (NumOperands == 1) {
449    if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
450      if (Ty!=Type::FloatTy && Ty!=Type::DoubleTy)
451        return 0;
452      /// Currently APFloat versions of these functions do not exist, so we use
453      /// the host native double versions.  Float versions are not called
454      /// directly but for all these it is true (float)(f((double)arg)) ==
455      /// f(arg).  Long double not supported yet.
456      double V = Ty==Type::FloatTy ? (double)Op->getValueAPF().convertToFloat():
457                                     Op->getValueAPF().convertToDouble();
458      switch (Str[0]) {
459      case 'a':
460        if (Len == 4 && !strcmp(Str, "acos"))
461          return ConstantFoldFP(acos, V, Ty);
462        else if (Len == 4 && !strcmp(Str, "asin"))
463          return ConstantFoldFP(asin, V, Ty);
464        else if (Len == 4 && !strcmp(Str, "atan"))
465          return ConstantFoldFP(atan, V, Ty);
466        break;
467      case 'c':
468        if (Len == 4 && !strcmp(Str, "ceil"))
469          return ConstantFoldFP(ceil, V, Ty);
470        else if (Len == 3 && !strcmp(Str, "cos"))
471          return ConstantFoldFP(cos, V, Ty);
472        else if (Len == 4 && !strcmp(Str, "cosh"))
473          return ConstantFoldFP(cosh, V, Ty);
474        break;
475      case 'e':
476        if (Len == 3 && !strcmp(Str, "exp"))
477          return ConstantFoldFP(exp, V, Ty);
478        break;
479      case 'f':
480        if (Len == 4 && !strcmp(Str, "fabs"))
481          return ConstantFoldFP(fabs, V, Ty);
482        else if (Len == 5 && !strcmp(Str, "floor"))
483          return ConstantFoldFP(floor, V, Ty);
484        break;
485      case 'l':
486        if (Len == 3 && !strcmp(Str, "log") && V > 0)
487          return ConstantFoldFP(log, V, Ty);
488        else if (Len == 5 && !strcmp(Str, "log10") && V > 0)
489          return ConstantFoldFP(log10, V, Ty);
490        else if (!strcmp(Str, "llvm.sqrt.f32") ||
491                 !strcmp(Str, "llvm.sqrt.f64")) {
492          if (V >= -0.0)
493            return ConstantFoldFP(sqrt, V, Ty);
494          else // Undefined
495            return ConstantFP::get(Ty, Ty==Type::FloatTy ? APFloat(0.0f) :
496                                       APFloat(0.0));
497        }
498        break;
499      case 's':
500        if (Len == 3 && !strcmp(Str, "sin"))
501          return ConstantFoldFP(sin, V, Ty);
502        else if (Len == 4 && !strcmp(Str, "sinh"))
503          return ConstantFoldFP(sinh, V, Ty);
504        else if (Len == 4 && !strcmp(Str, "sqrt") && V >= 0)
505          return ConstantFoldFP(sqrt, V, Ty);
506        else if (Len == 5 && !strcmp(Str, "sqrtf") && V >= 0)
507          return ConstantFoldFP(sqrt, V, Ty);
508        break;
509      case 't':
510        if (Len == 3 && !strcmp(Str, "tan"))
511          return ConstantFoldFP(tan, V, Ty);
512        else if (Len == 4 && !strcmp(Str, "tanh"))
513          return ConstantFoldFP(tanh, V, Ty);
514        break;
515      default:
516        break;
517      }
518    } else if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
519      if (Len > 11 && !memcmp(Str, "llvm.bswap", 10))
520        return ConstantInt::get(Op->getValue().byteSwap());
521      else if (Len > 11 && !memcmp(Str, "llvm.ctpop", 10))
522        return ConstantInt::get(Ty, Op->getValue().countPopulation());
523      else if (Len > 10 && !memcmp(Str, "llvm.cttz", 9))
524        return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
525      else if (Len > 10 && !memcmp(Str, "llvm.ctlz", 9))
526        return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
527    }
528  } else if (NumOperands == 2) {
529    if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
530      if (Ty!=Type::FloatTy && Ty!=Type::DoubleTy)
531        return 0;
532      double Op1V = Ty==Type::FloatTy ?
533                      (double)Op1->getValueAPF().convertToFloat():
534                      Op1->getValueAPF().convertToDouble();
535      if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
536        double Op2V = Ty==Type::FloatTy ?
537                      (double)Op2->getValueAPF().convertToFloat():
538                      Op2->getValueAPF().convertToDouble();
539
540        if (Len == 3 && !strcmp(Str, "pow")) {
541          return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty);
542        } else if (Len == 4 && !strcmp(Str, "fmod")) {
543          return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty);
544        } else if (Len == 5 && !strcmp(Str, "atan2")) {
545          return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
546        }
547      } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
548        if (!strcmp(Str, "llvm.powi.f32")) {
549          return ConstantFP::get(Ty, APFloat((float)std::pow((float)Op1V,
550                                              (int)Op2C->getZExtValue())));
551        } else if (!strcmp(Str, "llvm.powi.f64")) {
552          return ConstantFP::get(Ty, APFloat((double)std::pow((double)Op1V,
553                                              (int)Op2C->getZExtValue())));
554        }
555      }
556    }
557  }
558  return 0;
559}
560
561