ConstantFolding.cpp revision 58665d47aea69875c433fb039c8f10106abc1cee
1//===-- ConstantFolding.cpp - Fold instructions into constants ------------===//
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 defines routines for folding instructions into constants.
11//
12// Also, to supplement the basic VMCore ConstantExpr simplifications,
13// this file defines some additional folding routines that can make use of
14// TargetData information. These functions cannot go in VMCore due to library
15// dependency issues.
16//
17//===----------------------------------------------------------------------===//
18
19#include "llvm/Analysis/ConstantFolding.h"
20#include "llvm/Constants.h"
21#include "llvm/DerivedTypes.h"
22#include "llvm/Function.h"
23#include "llvm/GlobalVariable.h"
24#include "llvm/Instructions.h"
25#include "llvm/Intrinsics.h"
26#include "llvm/LLVMContext.h"
27#include "llvm/ADT/SmallVector.h"
28#include "llvm/ADT/StringMap.h"
29#include "llvm/Target/TargetData.h"
30#include "llvm/Support/ErrorHandling.h"
31#include "llvm/Support/GetElementPtrTypeIterator.h"
32#include "llvm/Support/MathExtras.h"
33#include <cerrno>
34#include <cmath>
35using namespace llvm;
36
37//===----------------------------------------------------------------------===//
38// Constant Folding internal helper functions
39//===----------------------------------------------------------------------===//
40
41/// IsConstantOffsetFromGlobal - If this constant is actually a constant offset
42/// from a global, return the global and the constant.  Because of
43/// constantexprs, this function is recursive.
44static bool IsConstantOffsetFromGlobal(Constant *C, GlobalValue *&GV,
45                                       int64_t &Offset, const TargetData &TD) {
46  // Trivial case, constant is the global.
47  if ((GV = dyn_cast<GlobalValue>(C))) {
48    Offset = 0;
49    return true;
50  }
51
52  // Otherwise, if this isn't a constant expr, bail out.
53  ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
54  if (!CE) return false;
55
56  // Look through ptr->int and ptr->ptr casts.
57  if (CE->getOpcode() == Instruction::PtrToInt ||
58      CE->getOpcode() == Instruction::BitCast)
59    return IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD);
60
61  // i32* getelementptr ([5 x i32]* @a, i32 0, i32 5)
62  if (CE->getOpcode() == Instruction::GetElementPtr) {
63    // Cannot compute this if the element type of the pointer is missing size
64    // info.
65    if (!cast<PointerType>(CE->getOperand(0)->getType())
66                 ->getElementType()->isSized())
67      return false;
68
69    // If the base isn't a global+constant, we aren't either.
70    if (!IsConstantOffsetFromGlobal(CE->getOperand(0), GV, Offset, TD))
71      return false;
72
73    // Otherwise, add any offset that our operands provide.
74    gep_type_iterator GTI = gep_type_begin(CE);
75    for (User::const_op_iterator i = CE->op_begin() + 1, e = CE->op_end();
76         i != e; ++i, ++GTI) {
77      ConstantInt *CI = dyn_cast<ConstantInt>(*i);
78      if (!CI) return false;  // Index isn't a simple constant?
79      if (CI->getZExtValue() == 0) continue;  // Not adding anything.
80
81      if (const StructType *ST = dyn_cast<StructType>(*GTI)) {
82        // N = N + Offset
83        Offset += TD.getStructLayout(ST)->getElementOffset(CI->getZExtValue());
84      } else {
85        const SequentialType *SQT = cast<SequentialType>(*GTI);
86        Offset += TD.getTypeAllocSize(SQT->getElementType())*CI->getSExtValue();
87      }
88    }
89    return true;
90  }
91
92  return false;
93}
94
95
96/// SymbolicallyEvaluateBinop - One of Op0/Op1 is a constant expression.
97/// Attempt to symbolically evaluate the result of a binary operator merging
98/// these together.  If target data info is available, it is provided as TD,
99/// otherwise TD is null.
100static Constant *SymbolicallyEvaluateBinop(unsigned Opc, Constant *Op0,
101                                           Constant *Op1, const TargetData *TD,
102                                           LLVMContext &Context){
103  // SROA
104
105  // Fold (and 0xffffffff00000000, (shl x, 32)) -> shl.
106  // Fold (lshr (or X, Y), 32) -> (lshr [X/Y], 32) if one doesn't contribute
107  // bits.
108
109
110  // If the constant expr is something like &A[123] - &A[4].f, fold this into a
111  // constant.  This happens frequently when iterating over a global array.
112  if (Opc == Instruction::Sub && TD) {
113    GlobalValue *GV1, *GV2;
114    int64_t Offs1, Offs2;
115
116    if (IsConstantOffsetFromGlobal(Op0, GV1, Offs1, *TD))
117      if (IsConstantOffsetFromGlobal(Op1, GV2, Offs2, *TD) &&
118          GV1 == GV2) {
119        // (&GV+C1) - (&GV+C2) -> C1-C2, pointer arithmetic cannot overflow.
120        return ConstantInt::get(Op0->getType(), Offs1-Offs2);
121      }
122  }
123
124  return 0;
125}
126
127/// SymbolicallyEvaluateGEP - If we can symbolically evaluate the specified GEP
128/// constant expression, do so.
129static Constant *SymbolicallyEvaluateGEP(Constant* const* Ops, unsigned NumOps,
130                                         const Type *ResultTy,
131                                         LLVMContext &Context,
132                                         const TargetData *TD) {
133  Constant *Ptr = Ops[0];
134  if (!TD || !cast<PointerType>(Ptr->getType())->getElementType()->isSized())
135    return 0;
136
137  unsigned BitWidth = TD->getTypeSizeInBits(TD->getIntPtrType(Context));
138  APInt BasePtr(BitWidth, 0);
139  bool BaseIsInt = true;
140  if (!Ptr->isNullValue()) {
141    // If this is a inttoptr from a constant int, we can fold this as the base,
142    // otherwise we can't.
143    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
144      if (CE->getOpcode() == Instruction::IntToPtr)
145        if (ConstantInt *Base = dyn_cast<ConstantInt>(CE->getOperand(0))) {
146          BasePtr = Base->getValue();
147          BasePtr.zextOrTrunc(BitWidth);
148        }
149
150    if (BasePtr == 0)
151      BaseIsInt = false;
152  }
153
154  // If this is a constant expr gep that is effectively computing an
155  // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
156  for (unsigned i = 1; i != NumOps; ++i)
157    if (!isa<ConstantInt>(Ops[i]))
158      return 0;
159
160  APInt Offset = APInt(BitWidth,
161                       TD->getIndexedOffset(Ptr->getType(),
162                                            (Value**)Ops+1, NumOps-1));
163  // If the base value for this address is a literal integer value, fold the
164  // getelementptr to the resulting integer value casted to the pointer type.
165  if (BaseIsInt) {
166    Constant *C = ConstantInt::get(Context, Offset+BasePtr);
167    return ConstantExpr::getIntToPtr(C, ResultTy);
168  }
169
170  // Otherwise form a regular getelementptr. Recompute the indices so that
171  // we eliminate over-indexing of the notional static type array bounds.
172  // This makes it easy to determine if the getelementptr is "inbounds".
173  // Also, this helps GlobalOpt do SROA on GlobalVariables.
174  const Type *Ty = Ptr->getType();
175  SmallVector<Constant*, 32> NewIdxs;
176  do {
177    if (const SequentialType *ATy = dyn_cast<SequentialType>(Ty)) {
178      // The only pointer indexing we'll do is on the first index of the GEP.
179      if (isa<PointerType>(ATy) && !NewIdxs.empty())
180        break;
181      // Determine which element of the array the offset points into.
182      APInt ElemSize(BitWidth, TD->getTypeAllocSize(ATy->getElementType()));
183      if (ElemSize == 0)
184        return 0;
185      APInt NewIdx = Offset.udiv(ElemSize);
186      Offset -= NewIdx * ElemSize;
187      NewIdxs.push_back(ConstantInt::get(TD->getIntPtrType(Context), NewIdx));
188      Ty = ATy->getElementType();
189    } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
190      // Determine which field of the struct the offset points into. The
191      // getZExtValue is at least as safe as the StructLayout API because we
192      // know the offset is within the struct at this point.
193      const StructLayout &SL = *TD->getStructLayout(STy);
194      unsigned ElIdx = SL.getElementContainingOffset(Offset.getZExtValue());
195      NewIdxs.push_back(ConstantInt::get(Type::getInt32Ty(Context), ElIdx));
196      Offset -= APInt(BitWidth, SL.getElementOffset(ElIdx));
197      Ty = STy->getTypeAtIndex(ElIdx);
198    } else {
199      // We've reached some non-indexable type.
200      break;
201    }
202  } while (Ty != cast<PointerType>(ResultTy)->getElementType());
203
204  // If we haven't used up the entire offset by descending the static
205  // type, then the offset is pointing into the middle of an indivisible
206  // member, so we can't simplify it.
207  if (Offset != 0)
208    return 0;
209
210  // Create a GEP.
211  Constant *C =
212    ConstantExpr::getGetElementPtr(Ptr, &NewIdxs[0], NewIdxs.size());
213  assert(cast<PointerType>(C->getType())->getElementType() == Ty &&
214         "Computed GetElementPtr has unexpected type!");
215
216  // If we ended up indexing a member with a type that doesn't match
217  // the type of what the original indices indexed, add a cast.
218  if (Ty != cast<PointerType>(ResultTy)->getElementType())
219    C = ConstantExpr::getBitCast(C, ResultTy);
220
221  return C;
222}
223
224/// FoldBitCast - Constant fold bitcast, symbolically evaluating it with
225/// targetdata.  Return 0 if unfoldable.
226static Constant *FoldBitCast(Constant *C, const Type *DestTy,
227                             const TargetData &TD, LLVMContext &Context) {
228  // If this is a bitcast from constant vector -> vector, fold it.
229  if (ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
230    if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
231      // If the element types match, VMCore can fold it.
232      unsigned NumDstElt = DestVTy->getNumElements();
233      unsigned NumSrcElt = CV->getNumOperands();
234      if (NumDstElt == NumSrcElt)
235        return 0;
236
237      const Type *SrcEltTy = CV->getType()->getElementType();
238      const Type *DstEltTy = DestVTy->getElementType();
239
240      // Otherwise, we're changing the number of elements in a vector, which
241      // requires endianness information to do the right thing.  For example,
242      //    bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
243      // folds to (little endian):
244      //    <4 x i32> <i32 0, i32 0, i32 1, i32 0>
245      // and to (big endian):
246      //    <4 x i32> <i32 0, i32 0, i32 0, i32 1>
247
248      // First thing is first.  We only want to think about integer here, so if
249      // we have something in FP form, recast it as integer.
250      if (DstEltTy->isFloatingPoint()) {
251        // Fold to an vector of integers with same size as our FP type.
252        unsigned FPWidth = DstEltTy->getPrimitiveSizeInBits();
253        const Type *DestIVTy = VectorType::get(
254                                 IntegerType::get(Context, FPWidth), NumDstElt);
255        // Recursively handle this integer conversion, if possible.
256        C = FoldBitCast(C, DestIVTy, TD, Context);
257        if (!C) return 0;
258
259        // Finally, VMCore can handle this now that #elts line up.
260        return ConstantExpr::getBitCast(C, DestTy);
261      }
262
263      // Okay, we know the destination is integer, if the input is FP, convert
264      // it to integer first.
265      if (SrcEltTy->isFloatingPoint()) {
266        unsigned FPWidth = SrcEltTy->getPrimitiveSizeInBits();
267        const Type *SrcIVTy = VectorType::get(
268                                 IntegerType::get(Context, FPWidth), NumSrcElt);
269        // Ask VMCore to do the conversion now that #elts line up.
270        C = ConstantExpr::getBitCast(C, SrcIVTy);
271        CV = dyn_cast<ConstantVector>(C);
272        if (!CV) return 0;  // If VMCore wasn't able to fold it, bail out.
273      }
274
275      // Now we know that the input and output vectors are both integer vectors
276      // of the same size, and that their #elements is not the same.  Do the
277      // conversion here, which depends on whether the input or output has
278      // more elements.
279      bool isLittleEndian = TD.isLittleEndian();
280
281      SmallVector<Constant*, 32> Result;
282      if (NumDstElt < NumSrcElt) {
283        // Handle: bitcast (<4 x i32> <i32 0, i32 1, i32 2, i32 3> to <2 x i64>)
284        Constant *Zero = Constant::getNullValue(DstEltTy);
285        unsigned Ratio = NumSrcElt/NumDstElt;
286        unsigned SrcBitSize = SrcEltTy->getPrimitiveSizeInBits();
287        unsigned SrcElt = 0;
288        for (unsigned i = 0; i != NumDstElt; ++i) {
289          // Build each element of the result.
290          Constant *Elt = Zero;
291          unsigned ShiftAmt = isLittleEndian ? 0 : SrcBitSize*(Ratio-1);
292          for (unsigned j = 0; j != Ratio; ++j) {
293            Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(SrcElt++));
294            if (!Src) return 0;  // Reject constantexpr elements.
295
296            // Zero extend the element to the right size.
297            Src = ConstantExpr::getZExt(Src, Elt->getType());
298
299            // Shift it to the right place, depending on endianness.
300            Src = ConstantExpr::getShl(Src,
301                             ConstantInt::get(Src->getType(), ShiftAmt));
302            ShiftAmt += isLittleEndian ? SrcBitSize : -SrcBitSize;
303
304            // Mix it in.
305            Elt = ConstantExpr::getOr(Elt, Src);
306          }
307          Result.push_back(Elt);
308        }
309      } else {
310        // Handle: bitcast (<2 x i64> <i64 0, i64 1> to <4 x i32>)
311        unsigned Ratio = NumDstElt/NumSrcElt;
312        unsigned DstBitSize = DstEltTy->getPrimitiveSizeInBits();
313
314        // Loop over each source value, expanding into multiple results.
315        for (unsigned i = 0; i != NumSrcElt; ++i) {
316          Constant *Src = dyn_cast<ConstantInt>(CV->getOperand(i));
317          if (!Src) return 0;  // Reject constantexpr elements.
318
319          unsigned ShiftAmt = isLittleEndian ? 0 : DstBitSize*(Ratio-1);
320          for (unsigned j = 0; j != Ratio; ++j) {
321            // Shift the piece of the value into the right place, depending on
322            // endianness.
323            Constant *Elt = ConstantExpr::getLShr(Src,
324                            ConstantInt::get(Src->getType(), ShiftAmt));
325            ShiftAmt += isLittleEndian ? DstBitSize : -DstBitSize;
326
327            // Truncate and remember this piece.
328            Result.push_back(ConstantExpr::getTrunc(Elt, DstEltTy));
329          }
330        }
331      }
332
333      return ConstantVector::get(Result.data(), Result.size());
334    }
335  }
336
337  return 0;
338}
339
340
341//===----------------------------------------------------------------------===//
342// Constant Folding public APIs
343//===----------------------------------------------------------------------===//
344
345
346/// ConstantFoldInstruction - Attempt to constant fold the specified
347/// instruction.  If successful, the constant result is returned, if not, null
348/// is returned.  Note that this function can only fail when attempting to fold
349/// instructions like loads and stores, which have no constant expression form.
350///
351Constant *llvm::ConstantFoldInstruction(Instruction *I, LLVMContext &Context,
352                                        const TargetData *TD) {
353  if (PHINode *PN = dyn_cast<PHINode>(I)) {
354    if (PN->getNumIncomingValues() == 0)
355      return UndefValue::get(PN->getType());
356
357    Constant *Result = dyn_cast<Constant>(PN->getIncomingValue(0));
358    if (Result == 0) return 0;
359
360    // Handle PHI nodes specially here...
361    for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i)
362      if (PN->getIncomingValue(i) != Result && PN->getIncomingValue(i) != PN)
363        return 0;   // Not all the same incoming constants...
364
365    // If we reach here, all incoming values are the same constant.
366    return Result;
367  }
368
369  // Scan the operand list, checking to see if they are all constants, if so,
370  // hand off to ConstantFoldInstOperands.
371  SmallVector<Constant*, 8> Ops;
372  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
373    if (Constant *Op = dyn_cast<Constant>(*i))
374      Ops.push_back(Op);
375    else
376      return 0;  // All operands not constant!
377
378  if (const CmpInst *CI = dyn_cast<CmpInst>(I))
379    return ConstantFoldCompareInstOperands(CI->getPredicate(),
380                                           Ops.data(), Ops.size(),
381                                           Context, TD);
382
383  return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
384                                  Ops.data(), Ops.size(), Context, TD);
385}
386
387/// ConstantFoldConstantExpression - Attempt to fold the constant expression
388/// using the specified TargetData.  If successful, the constant result is
389/// result is returned, if not, null is returned.
390Constant *llvm::ConstantFoldConstantExpression(ConstantExpr *CE,
391                                               LLVMContext &Context,
392                                               const TargetData *TD) {
393  SmallVector<Constant*, 8> Ops;
394  for (User::op_iterator i = CE->op_begin(), e = CE->op_end(); i != e; ++i)
395    Ops.push_back(cast<Constant>(*i));
396
397  if (CE->isCompare())
398    return ConstantFoldCompareInstOperands(CE->getPredicate(),
399                                           Ops.data(), Ops.size(),
400                                           Context, TD);
401  return ConstantFoldInstOperands(CE->getOpcode(), CE->getType(),
402                                  Ops.data(), Ops.size(), Context, TD);
403}
404
405/// ConstantFoldInstOperands - Attempt to constant fold an instruction with the
406/// specified opcode and operands.  If successful, the constant result is
407/// returned, if not, null is returned.  Note that this function can fail when
408/// attempting to fold instructions like loads and stores, which have no
409/// constant expression form.
410///
411Constant *llvm::ConstantFoldInstOperands(unsigned Opcode, const Type *DestTy,
412                                         Constant* const* Ops, unsigned NumOps,
413                                         LLVMContext &Context,
414                                         const TargetData *TD) {
415  // Handle easy binops first.
416  if (Instruction::isBinaryOp(Opcode)) {
417    if (isa<ConstantExpr>(Ops[0]) || isa<ConstantExpr>(Ops[1]))
418      if (Constant *C = SymbolicallyEvaluateBinop(Opcode, Ops[0], Ops[1], TD,
419                                                  Context))
420        return C;
421
422    return ConstantExpr::get(Opcode, Ops[0], Ops[1]);
423  }
424
425  switch (Opcode) {
426  default: return 0;
427  case Instruction::Call:
428    if (Function *F = dyn_cast<Function>(Ops[0]))
429      if (canConstantFoldCallTo(F))
430        return ConstantFoldCall(F, Ops+1, NumOps-1);
431    return 0;
432  case Instruction::ICmp:
433  case Instruction::FCmp:
434    llvm_unreachable("This function is invalid for compares: no predicate specified");
435  case Instruction::PtrToInt:
436    // If the input is a inttoptr, eliminate the pair.  This requires knowing
437    // the width of a pointer, so it can't be done in ConstantExpr::getCast.
438    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
439      if (TD && CE->getOpcode() == Instruction::IntToPtr) {
440        Constant *Input = CE->getOperand(0);
441        unsigned InWidth = Input->getType()->getScalarSizeInBits();
442        if (TD->getPointerSizeInBits() < InWidth) {
443          Constant *Mask =
444            ConstantInt::get(Context, APInt::getLowBitsSet(InWidth,
445                                                  TD->getPointerSizeInBits()));
446          Input = ConstantExpr::getAnd(Input, Mask);
447        }
448        // Do a zext or trunc to get to the dest size.
449        return ConstantExpr::getIntegerCast(Input, DestTy, false);
450      }
451    }
452    return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
453  case Instruction::IntToPtr:
454    // If the input is a ptrtoint, turn the pair into a ptr to ptr bitcast if
455    // the int size is >= the ptr size.  This requires knowing the width of a
456    // pointer, so it can't be done in ConstantExpr::getCast.
457    if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ops[0])) {
458      if (TD &&
459          TD->getPointerSizeInBits() <=
460          CE->getType()->getScalarSizeInBits()) {
461        if (CE->getOpcode() == Instruction::PtrToInt) {
462          Constant *Input = CE->getOperand(0);
463          Constant *C = FoldBitCast(Input, DestTy, *TD, Context);
464          return C ? C : ConstantExpr::getBitCast(Input, DestTy);
465        }
466        // If there's a constant offset added to the integer value before
467        // it is casted back to a pointer, see if the expression can be
468        // converted into a GEP.
469        if (CE->getOpcode() == Instruction::Add)
470          if (ConstantInt *L = dyn_cast<ConstantInt>(CE->getOperand(0)))
471            if (ConstantExpr *R = dyn_cast<ConstantExpr>(CE->getOperand(1)))
472              if (R->getOpcode() == Instruction::PtrToInt)
473                if (GlobalVariable *GV =
474                      dyn_cast<GlobalVariable>(R->getOperand(0))) {
475                  const PointerType *GVTy = cast<PointerType>(GV->getType());
476                  if (const ArrayType *AT =
477                        dyn_cast<ArrayType>(GVTy->getElementType())) {
478                    const Type *ElTy = AT->getElementType();
479                    uint64_t AllocSize = TD->getTypeAllocSize(ElTy);
480                    APInt PSA(L->getValue().getBitWidth(), AllocSize);
481                    if (ElTy == cast<PointerType>(DestTy)->getElementType() &&
482                        L->getValue().urem(PSA) == 0) {
483                      APInt ElemIdx = L->getValue().udiv(PSA);
484                      if (ElemIdx.ult(APInt(ElemIdx.getBitWidth(),
485                                            AT->getNumElements()))) {
486                        Constant *Index[] = {
487                          Constant::getNullValue(CE->getType()),
488                          ConstantInt::get(Context, ElemIdx)
489                        };
490                        return
491                        ConstantExpr::getGetElementPtr(GV, &Index[0], 2);
492                      }
493                    }
494                  }
495                }
496      }
497    }
498    return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
499  case Instruction::Trunc:
500  case Instruction::ZExt:
501  case Instruction::SExt:
502  case Instruction::FPTrunc:
503  case Instruction::FPExt:
504  case Instruction::UIToFP:
505  case Instruction::SIToFP:
506  case Instruction::FPToUI:
507  case Instruction::FPToSI:
508      return ConstantExpr::getCast(Opcode, Ops[0], DestTy);
509  case Instruction::BitCast:
510    if (TD)
511      if (Constant *C = FoldBitCast(Ops[0], DestTy, *TD, Context))
512        return C;
513    return ConstantExpr::getBitCast(Ops[0], DestTy);
514  case Instruction::Select:
515    return ConstantExpr::getSelect(Ops[0], Ops[1], Ops[2]);
516  case Instruction::ExtractElement:
517    return ConstantExpr::getExtractElement(Ops[0], Ops[1]);
518  case Instruction::InsertElement:
519    return ConstantExpr::getInsertElement(Ops[0], Ops[1], Ops[2]);
520  case Instruction::ShuffleVector:
521    return ConstantExpr::getShuffleVector(Ops[0], Ops[1], Ops[2]);
522  case Instruction::GetElementPtr:
523    if (Constant *C = SymbolicallyEvaluateGEP(Ops, NumOps, DestTy, Context, TD))
524      return C;
525
526    return ConstantExpr::getGetElementPtr(Ops[0], Ops+1, NumOps-1);
527  }
528}
529
530/// ConstantFoldCompareInstOperands - Attempt to constant fold a compare
531/// instruction (icmp/fcmp) with the specified operands.  If it fails, it
532/// returns a constant expression of the specified operands.
533///
534Constant *llvm::ConstantFoldCompareInstOperands(unsigned Predicate,
535                                                Constant*const * Ops,
536                                                unsigned NumOps,
537                                                LLVMContext &Context,
538                                                const TargetData *TD) {
539  // fold: icmp (inttoptr x), null         -> icmp x, 0
540  // fold: icmp (ptrtoint x), 0            -> icmp x, null
541  // fold: icmp (inttoptr x), (inttoptr y) -> icmp trunc/zext x, trunc/zext y
542  // fold: icmp (ptrtoint x), (ptrtoint y) -> icmp x, y
543  //
544  // ConstantExpr::getCompare cannot do this, because it doesn't have TD
545  // around to know if bit truncation is happening.
546  if (ConstantExpr *CE0 = dyn_cast<ConstantExpr>(Ops[0])) {
547    if (TD && Ops[1]->isNullValue()) {
548      const Type *IntPtrTy = TD->getIntPtrType(Context);
549      if (CE0->getOpcode() == Instruction::IntToPtr) {
550        // Convert the integer value to the right size to ensure we get the
551        // proper extension or truncation.
552        Constant *C = ConstantExpr::getIntegerCast(CE0->getOperand(0),
553                                                   IntPtrTy, false);
554        Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
555        return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
556                                               Context, TD);
557      }
558
559      // Only do this transformation if the int is intptrty in size, otherwise
560      // there is a truncation or extension that we aren't modeling.
561      if (CE0->getOpcode() == Instruction::PtrToInt &&
562          CE0->getType() == IntPtrTy) {
563        Constant *C = CE0->getOperand(0);
564        Constant *NewOps[] = { C, Constant::getNullValue(C->getType()) };
565        // FIXME!
566        return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
567                                               Context, TD);
568      }
569    }
570
571    if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(Ops[1])) {
572      if (TD && CE0->getOpcode() == CE1->getOpcode()) {
573        const Type *IntPtrTy = TD->getIntPtrType(Context);
574
575        if (CE0->getOpcode() == Instruction::IntToPtr) {
576          // Convert the integer value to the right size to ensure we get the
577          // proper extension or truncation.
578          Constant *C0 = ConstantExpr::getIntegerCast(CE0->getOperand(0),
579                                                      IntPtrTy, false);
580          Constant *C1 = ConstantExpr::getIntegerCast(CE1->getOperand(0),
581                                                      IntPtrTy, false);
582          Constant *NewOps[] = { C0, C1 };
583          return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
584                                                 Context, TD);
585        }
586
587        // Only do this transformation if the int is intptrty in size, otherwise
588        // there is a truncation or extension that we aren't modeling.
589        if ((CE0->getOpcode() == Instruction::PtrToInt &&
590             CE0->getType() == IntPtrTy &&
591             CE0->getOperand(0)->getType() == CE1->getOperand(0)->getType())) {
592          Constant *NewOps[] = {
593            CE0->getOperand(0), CE1->getOperand(0)
594          };
595          return ConstantFoldCompareInstOperands(Predicate, NewOps, 2,
596                                                 Context, TD);
597        }
598      }
599    }
600  }
601  return ConstantExpr::getCompare(Predicate, Ops[0], Ops[1]);
602}
603
604
605/// ConstantFoldLoadThroughGEPConstantExpr - Given a constant and a
606/// getelementptr constantexpr, return the constant value being addressed by the
607/// constant expression, or null if something is funny and we can't decide.
608Constant *llvm::ConstantFoldLoadThroughGEPConstantExpr(Constant *C,
609                                                       ConstantExpr *CE,
610                                                       LLVMContext &Context) {
611  if (CE->getOperand(1) != Constant::getNullValue(CE->getOperand(1)->getType()))
612    return 0;  // Do not allow stepping over the value!
613
614  // Loop over all of the operands, tracking down which value we are
615  // addressing...
616  gep_type_iterator I = gep_type_begin(CE), E = gep_type_end(CE);
617  for (++I; I != E; ++I)
618    if (const StructType *STy = dyn_cast<StructType>(*I)) {
619      ConstantInt *CU = cast<ConstantInt>(I.getOperand());
620      assert(CU->getZExtValue() < STy->getNumElements() &&
621             "Struct index out of range!");
622      unsigned El = (unsigned)CU->getZExtValue();
623      if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C)) {
624        C = CS->getOperand(El);
625      } else if (isa<ConstantAggregateZero>(C)) {
626        C = Constant::getNullValue(STy->getElementType(El));
627      } else if (isa<UndefValue>(C)) {
628        C = UndefValue::get(STy->getElementType(El));
629      } else {
630        return 0;
631      }
632    } else if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) {
633      if (const ArrayType *ATy = dyn_cast<ArrayType>(*I)) {
634        if (CI->getZExtValue() >= ATy->getNumElements())
635         return 0;
636        if (ConstantArray *CA = dyn_cast<ConstantArray>(C))
637          C = CA->getOperand(CI->getZExtValue());
638        else if (isa<ConstantAggregateZero>(C))
639          C = Constant::getNullValue(ATy->getElementType());
640        else if (isa<UndefValue>(C))
641          C = UndefValue::get(ATy->getElementType());
642        else
643          return 0;
644      } else if (const VectorType *PTy = dyn_cast<VectorType>(*I)) {
645        if (CI->getZExtValue() >= PTy->getNumElements())
646          return 0;
647        if (ConstantVector *CP = dyn_cast<ConstantVector>(C))
648          C = CP->getOperand(CI->getZExtValue());
649        else if (isa<ConstantAggregateZero>(C))
650          C = Constant::getNullValue(PTy->getElementType());
651        else if (isa<UndefValue>(C))
652          C = UndefValue::get(PTy->getElementType());
653        else
654          return 0;
655      } else {
656        return 0;
657      }
658    } else {
659      return 0;
660    }
661  return C;
662}
663
664
665//===----------------------------------------------------------------------===//
666//  Constant Folding for Calls
667//
668
669/// canConstantFoldCallTo - Return true if its even possible to fold a call to
670/// the specified function.
671bool
672llvm::canConstantFoldCallTo(const Function *F) {
673  switch (F->getIntrinsicID()) {
674  case Intrinsic::sqrt:
675  case Intrinsic::powi:
676  case Intrinsic::bswap:
677  case Intrinsic::ctpop:
678  case Intrinsic::ctlz:
679  case Intrinsic::cttz:
680    return true;
681  default: break;
682  }
683
684  if (!F->hasName()) return false;
685  StringRef Name = F->getName();
686
687  // In these cases, the check of the length is required.  We don't want to
688  // return true for a name like "cos\0blah" which strcmp would return equal to
689  // "cos", but has length 8.
690  switch (Name[0]) {
691  default: return false;
692  case 'a':
693    return Name == "acos" || Name == "asin" ||
694      Name == "atan" || Name == "atan2";
695  case 'c':
696    return Name == "cos" || Name == "ceil" || Name == "cosf" || Name == "cosh";
697  case 'e':
698    return Name == "exp";
699  case 'f':
700    return Name == "fabs" || Name == "fmod" || Name == "floor";
701  case 'l':
702    return Name == "log" || Name == "log10";
703  case 'p':
704    return Name == "pow";
705  case 's':
706    return Name == "sin" || Name == "sinh" || Name == "sqrt" ||
707      Name == "sinf" || Name == "sqrtf";
708  case 't':
709    return Name == "tan" || Name == "tanh";
710  }
711}
712
713static Constant *ConstantFoldFP(double (*NativeFP)(double), double V,
714                                const Type *Ty, LLVMContext &Context) {
715  errno = 0;
716  V = NativeFP(V);
717  if (errno != 0) {
718    errno = 0;
719    return 0;
720  }
721
722  if (Ty == Type::getFloatTy(Context))
723    return ConstantFP::get(Context, APFloat((float)V));
724  if (Ty == Type::getDoubleTy(Context))
725    return ConstantFP::get(Context, APFloat(V));
726  llvm_unreachable("Can only constant fold float/double");
727  return 0; // dummy return to suppress warning
728}
729
730static Constant *ConstantFoldBinaryFP(double (*NativeFP)(double, double),
731                                      double V, double W,
732                                      const Type *Ty,
733                                      LLVMContext &Context) {
734  errno = 0;
735  V = NativeFP(V, W);
736  if (errno != 0) {
737    errno = 0;
738    return 0;
739  }
740
741  if (Ty == Type::getFloatTy(Context))
742    return ConstantFP::get(Context, APFloat((float)V));
743  if (Ty == Type::getDoubleTy(Context))
744    return ConstantFP::get(Context, APFloat(V));
745  llvm_unreachable("Can only constant fold float/double");
746  return 0; // dummy return to suppress warning
747}
748
749/// ConstantFoldCall - Attempt to constant fold a call to the specified function
750/// with the specified arguments, returning null if unsuccessful.
751
752Constant *
753llvm::ConstantFoldCall(Function *F,
754                       Constant* const* Operands, unsigned NumOperands) {
755  if (!F->hasName()) return 0;
756  LLVMContext &Context = F->getContext();
757  StringRef Name = F->getName();
758
759  const Type *Ty = F->getReturnType();
760  if (NumOperands == 1) {
761    if (ConstantFP *Op = dyn_cast<ConstantFP>(Operands[0])) {
762      if (Ty!=Type::getFloatTy(F->getContext()) &&
763          Ty!=Type::getDoubleTy(Context))
764        return 0;
765      /// Currently APFloat versions of these functions do not exist, so we use
766      /// the host native double versions.  Float versions are not called
767      /// directly but for all these it is true (float)(f((double)arg)) ==
768      /// f(arg).  Long double not supported yet.
769      double V = Ty==Type::getFloatTy(F->getContext()) ?
770                                     (double)Op->getValueAPF().convertToFloat():
771                                     Op->getValueAPF().convertToDouble();
772      switch (Name[0]) {
773      case 'a':
774        if (Name == "acos")
775          return ConstantFoldFP(acos, V, Ty, Context);
776        else if (Name == "asin")
777          return ConstantFoldFP(asin, V, Ty, Context);
778        else if (Name == "atan")
779          return ConstantFoldFP(atan, V, Ty, Context);
780        break;
781      case 'c':
782        if (Name == "ceil")
783          return ConstantFoldFP(ceil, V, Ty, Context);
784        else if (Name == "cos")
785          return ConstantFoldFP(cos, V, Ty, Context);
786        else if (Name == "cosh")
787          return ConstantFoldFP(cosh, V, Ty, Context);
788        else if (Name == "cosf")
789          return ConstantFoldFP(cos, V, Ty, Context);
790        break;
791      case 'e':
792        if (Name == "exp")
793          return ConstantFoldFP(exp, V, Ty, Context);
794        break;
795      case 'f':
796        if (Name == "fabs")
797          return ConstantFoldFP(fabs, V, Ty, Context);
798        else if (Name == "floor")
799          return ConstantFoldFP(floor, V, Ty, Context);
800        break;
801      case 'l':
802        if (Name == "log" && V > 0)
803          return ConstantFoldFP(log, V, Ty, Context);
804        else if (Name == "log10" && V > 0)
805          return ConstantFoldFP(log10, V, Ty, Context);
806        else if (Name == "llvm.sqrt.f32" ||
807                 Name == "llvm.sqrt.f64") {
808          if (V >= -0.0)
809            return ConstantFoldFP(sqrt, V, Ty, Context);
810          else // Undefined
811            return Constant::getNullValue(Ty);
812        }
813        break;
814      case 's':
815        if (Name == "sin")
816          return ConstantFoldFP(sin, V, Ty, Context);
817        else if (Name == "sinh")
818          return ConstantFoldFP(sinh, V, Ty, Context);
819        else if (Name == "sqrt" && V >= 0)
820          return ConstantFoldFP(sqrt, V, Ty, Context);
821        else if (Name == "sqrtf" && V >= 0)
822          return ConstantFoldFP(sqrt, V, Ty, Context);
823        else if (Name == "sinf")
824          return ConstantFoldFP(sin, V, Ty, Context);
825        break;
826      case 't':
827        if (Name == "tan")
828          return ConstantFoldFP(tan, V, Ty, Context);
829        else if (Name == "tanh")
830          return ConstantFoldFP(tanh, V, Ty, Context);
831        break;
832      default:
833        break;
834      }
835    } else if (ConstantInt *Op = dyn_cast<ConstantInt>(Operands[0])) {
836      if (Name.startswith("llvm.bswap"))
837        return ConstantInt::get(Context, Op->getValue().byteSwap());
838      else if (Name.startswith("llvm.ctpop"))
839        return ConstantInt::get(Ty, Op->getValue().countPopulation());
840      else if (Name.startswith("llvm.cttz"))
841        return ConstantInt::get(Ty, Op->getValue().countTrailingZeros());
842      else if (Name.startswith("llvm.ctlz"))
843        return ConstantInt::get(Ty, Op->getValue().countLeadingZeros());
844    }
845  } else if (NumOperands == 2) {
846    if (ConstantFP *Op1 = dyn_cast<ConstantFP>(Operands[0])) {
847      if (Ty!=Type::getFloatTy(F->getContext()) &&
848          Ty!=Type::getDoubleTy(Context))
849        return 0;
850      double Op1V = Ty==Type::getFloatTy(F->getContext()) ?
851                      (double)Op1->getValueAPF().convertToFloat():
852                      Op1->getValueAPF().convertToDouble();
853      if (ConstantFP *Op2 = dyn_cast<ConstantFP>(Operands[1])) {
854        double Op2V = Ty==Type::getFloatTy(F->getContext()) ?
855                      (double)Op2->getValueAPF().convertToFloat():
856                      Op2->getValueAPF().convertToDouble();
857
858        if (Name == "pow") {
859          return ConstantFoldBinaryFP(pow, Op1V, Op2V, Ty, Context);
860        } else if (Name == "fmod") {
861          return ConstantFoldBinaryFP(fmod, Op1V, Op2V, Ty, Context);
862        } else if (Name == "atan2") {
863          return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty, Context);
864        }
865      } else if (ConstantInt *Op2C = dyn_cast<ConstantInt>(Operands[1])) {
866        if (Name == "llvm.powi.f32") {
867          return ConstantFP::get(Context, APFloat((float)std::pow((float)Op1V,
868                                                 (int)Op2C->getZExtValue())));
869        } else if (Name == "llvm.powi.f64") {
870          return ConstantFP::get(Context, APFloat((double)std::pow((double)Op1V,
871                                                 (int)Op2C->getZExtValue())));
872        }
873      }
874    }
875  }
876  return 0;
877}
878
879