1 //===- ConstantFold.cpp - LLVM constant folder ----------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements folding of constants for LLVM.  This implements the
10 // (internal) ConstantFold.h interface, which is used by the
11 // ConstantExpr::get* methods to automatically fold constants when possible.
12 //
13 // The current constant folding implementation is implemented in two pieces: the
14 // pieces that don't need DataLayout, and the pieces that do. This is to avoid
15 // a dependence in IR on Target.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #include "ConstantFold.h"
20 #include "llvm/ADT/APSInt.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/GetElementPtrTypeIterator.h"
26 #include "llvm/IR/GlobalAlias.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/IR/PatternMatch.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/ManagedStatic.h"
34 #include "llvm/Support/MathExtras.h"
35 using namespace llvm;
36 using namespace llvm::PatternMatch;
37 
38 //===----------------------------------------------------------------------===//
39 //                ConstantFold*Instruction Implementations
40 //===----------------------------------------------------------------------===//
41 
42 /// Convert the specified vector Constant node to the specified vector type.
43 /// At this point, we know that the elements of the input vector constant are
44 /// all simple integer or FP values.
BitCastConstantVector(Constant * CV,VectorType * DstTy)45 static Constant *BitCastConstantVector(Constant *CV, VectorType *DstTy) {
46 
47   if (CV->isAllOnesValue()) return Constant::getAllOnesValue(DstTy);
48   if (CV->isNullValue()) return Constant::getNullValue(DstTy);
49 
50   // Do not iterate on scalable vector. The num of elements is unknown at
51   // compile-time.
52   if (isa<ScalableVectorType>(DstTy))
53     return nullptr;
54 
55   // If this cast changes element count then we can't handle it here:
56   // doing so requires endianness information.  This should be handled by
57   // Analysis/ConstantFolding.cpp
58   unsigned NumElts = cast<FixedVectorType>(DstTy)->getNumElements();
59   if (NumElts != cast<FixedVectorType>(CV->getType())->getNumElements())
60     return nullptr;
61 
62   Type *DstEltTy = DstTy->getElementType();
63   // Fast path for splatted constants.
64   if (Constant *Splat = CV->getSplatValue()) {
65     return ConstantVector::getSplat(DstTy->getElementCount(),
66                                     ConstantExpr::getBitCast(Splat, DstEltTy));
67   }
68 
69   SmallVector<Constant*, 16> Result;
70   Type *Ty = IntegerType::get(CV->getContext(), 32);
71   for (unsigned i = 0; i != NumElts; ++i) {
72     Constant *C =
73       ConstantExpr::getExtractElement(CV, ConstantInt::get(Ty, i));
74     C = ConstantExpr::getBitCast(C, DstEltTy);
75     Result.push_back(C);
76   }
77 
78   return ConstantVector::get(Result);
79 }
80 
81 /// This function determines which opcode to use to fold two constant cast
82 /// expressions together. It uses CastInst::isEliminableCastPair to determine
83 /// the opcode. Consequently its just a wrapper around that function.
84 /// Determine if it is valid to fold a cast of a cast
85 static unsigned
foldConstantCastPair(unsigned opc,ConstantExpr * Op,Type * DstTy)86 foldConstantCastPair(
87   unsigned opc,          ///< opcode of the second cast constant expression
88   ConstantExpr *Op,      ///< the first cast constant expression
89   Type *DstTy            ///< destination type of the first cast
90 ) {
91   assert(Op && Op->isCast() && "Can't fold cast of cast without a cast!");
92   assert(DstTy && DstTy->isFirstClassType() && "Invalid cast destination type");
93   assert(CastInst::isCast(opc) && "Invalid cast opcode");
94 
95   // The types and opcodes for the two Cast constant expressions
96   Type *SrcTy = Op->getOperand(0)->getType();
97   Type *MidTy = Op->getType();
98   Instruction::CastOps firstOp = Instruction::CastOps(Op->getOpcode());
99   Instruction::CastOps secondOp = Instruction::CastOps(opc);
100 
101   // Assume that pointers are never more than 64 bits wide, and only use this
102   // for the middle type. Otherwise we could end up folding away illegal
103   // bitcasts between address spaces with different sizes.
104   IntegerType *FakeIntPtrTy = Type::getInt64Ty(DstTy->getContext());
105 
106   // Let CastInst::isEliminableCastPair do the heavy lifting.
107   return CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy, DstTy,
108                                         nullptr, FakeIntPtrTy, nullptr);
109 }
110 
FoldBitCast(Constant * V,Type * DestTy)111 static Constant *FoldBitCast(Constant *V, Type *DestTy) {
112   Type *SrcTy = V->getType();
113   if (SrcTy == DestTy)
114     return V; // no-op cast
115 
116   // Check to see if we are casting a pointer to an aggregate to a pointer to
117   // the first element.  If so, return the appropriate GEP instruction.
118   if (PointerType *PTy = dyn_cast<PointerType>(V->getType()))
119     if (PointerType *DPTy = dyn_cast<PointerType>(DestTy))
120       if (PTy->getAddressSpace() == DPTy->getAddressSpace() &&
121           !PTy->isOpaque() && !DPTy->isOpaque() &&
122           PTy->getElementType()->isSized()) {
123         SmallVector<Value*, 8> IdxList;
124         Value *Zero =
125           Constant::getNullValue(Type::getInt32Ty(DPTy->getContext()));
126         IdxList.push_back(Zero);
127         Type *ElTy = PTy->getElementType();
128         while (ElTy && ElTy != DPTy->getElementType()) {
129           ElTy = GetElementPtrInst::getTypeAtIndex(ElTy, (uint64_t)0);
130           IdxList.push_back(Zero);
131         }
132 
133         if (ElTy == DPTy->getElementType())
134           // This GEP is inbounds because all indices are zero.
135           return ConstantExpr::getInBoundsGetElementPtr(PTy->getElementType(),
136                                                         V, IdxList);
137       }
138 
139   // Handle casts from one vector constant to another.  We know that the src
140   // and dest type have the same size (otherwise its an illegal cast).
141   if (VectorType *DestPTy = dyn_cast<VectorType>(DestTy)) {
142     if (VectorType *SrcTy = dyn_cast<VectorType>(V->getType())) {
143       assert(DestPTy->getPrimitiveSizeInBits() ==
144                  SrcTy->getPrimitiveSizeInBits() &&
145              "Not cast between same sized vectors!");
146       SrcTy = nullptr;
147       // First, check for null.  Undef is already handled.
148       if (isa<ConstantAggregateZero>(V))
149         return Constant::getNullValue(DestTy);
150 
151       // Handle ConstantVector and ConstantAggregateVector.
152       return BitCastConstantVector(V, DestPTy);
153     }
154 
155     // Canonicalize scalar-to-vector bitcasts into vector-to-vector bitcasts
156     // This allows for other simplifications (although some of them
157     // can only be handled by Analysis/ConstantFolding.cpp).
158     if (isa<ConstantInt>(V) || isa<ConstantFP>(V))
159       return ConstantExpr::getBitCast(ConstantVector::get(V), DestPTy);
160   }
161 
162   // Finally, implement bitcast folding now.   The code below doesn't handle
163   // bitcast right.
164   if (isa<ConstantPointerNull>(V))  // ptr->ptr cast.
165     return ConstantPointerNull::get(cast<PointerType>(DestTy));
166 
167   // Handle integral constant input.
168   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
169     if (DestTy->isIntegerTy())
170       // Integral -> Integral. This is a no-op because the bit widths must
171       // be the same. Consequently, we just fold to V.
172       return V;
173 
174     // See note below regarding the PPC_FP128 restriction.
175     if (DestTy->isFloatingPointTy() && !DestTy->isPPC_FP128Ty())
176       return ConstantFP::get(DestTy->getContext(),
177                              APFloat(DestTy->getFltSemantics(),
178                                      CI->getValue()));
179 
180     // Otherwise, can't fold this (vector?)
181     return nullptr;
182   }
183 
184   // Handle ConstantFP input: FP -> Integral.
185   if (ConstantFP *FP = dyn_cast<ConstantFP>(V)) {
186     // PPC_FP128 is really the sum of two consecutive doubles, where the first
187     // double is always stored first in memory, regardless of the target
188     // endianness. The memory layout of i128, however, depends on the target
189     // endianness, and so we can't fold this without target endianness
190     // information. This should instead be handled by
191     // Analysis/ConstantFolding.cpp
192     if (FP->getType()->isPPC_FP128Ty())
193       return nullptr;
194 
195     // Make sure dest type is compatible with the folded integer constant.
196     if (!DestTy->isIntegerTy())
197       return nullptr;
198 
199     return ConstantInt::get(FP->getContext(),
200                             FP->getValueAPF().bitcastToAPInt());
201   }
202 
203   return nullptr;
204 }
205 
206 
207 /// V is an integer constant which only has a subset of its bytes used.
208 /// The bytes used are indicated by ByteStart (which is the first byte used,
209 /// counting from the least significant byte) and ByteSize, which is the number
210 /// of bytes used.
211 ///
212 /// This function analyzes the specified constant to see if the specified byte
213 /// range can be returned as a simplified constant.  If so, the constant is
214 /// returned, otherwise null is returned.
ExtractConstantBytes(Constant * C,unsigned ByteStart,unsigned ByteSize)215 static Constant *ExtractConstantBytes(Constant *C, unsigned ByteStart,
216                                       unsigned ByteSize) {
217   assert(C->getType()->isIntegerTy() &&
218          (cast<IntegerType>(C->getType())->getBitWidth() & 7) == 0 &&
219          "Non-byte sized integer input");
220   unsigned CSize = cast<IntegerType>(C->getType())->getBitWidth()/8;
221   assert(ByteSize && "Must be accessing some piece");
222   assert(ByteStart+ByteSize <= CSize && "Extracting invalid piece from input");
223   assert(ByteSize != CSize && "Should not extract everything");
224 
225   // Constant Integers are simple.
226   if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
227     APInt V = CI->getValue();
228     if (ByteStart)
229       V.lshrInPlace(ByteStart*8);
230     V = V.trunc(ByteSize*8);
231     return ConstantInt::get(CI->getContext(), V);
232   }
233 
234   // In the input is a constant expr, we might be able to recursively simplify.
235   // If not, we definitely can't do anything.
236   ConstantExpr *CE = dyn_cast<ConstantExpr>(C);
237   if (!CE) return nullptr;
238 
239   switch (CE->getOpcode()) {
240   default: return nullptr;
241   case Instruction::Or: {
242     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
243     if (!RHS)
244       return nullptr;
245 
246     // X | -1 -> -1.
247     if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS))
248       if (RHSC->isMinusOne())
249         return RHSC;
250 
251     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
252     if (!LHS)
253       return nullptr;
254     return ConstantExpr::getOr(LHS, RHS);
255   }
256   case Instruction::And: {
257     Constant *RHS = ExtractConstantBytes(CE->getOperand(1), ByteStart,ByteSize);
258     if (!RHS)
259       return nullptr;
260 
261     // X & 0 -> 0.
262     if (RHS->isNullValue())
263       return RHS;
264 
265     Constant *LHS = ExtractConstantBytes(CE->getOperand(0), ByteStart,ByteSize);
266     if (!LHS)
267       return nullptr;
268     return ConstantExpr::getAnd(LHS, RHS);
269   }
270   case Instruction::LShr: {
271     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
272     if (!Amt)
273       return nullptr;
274     APInt ShAmt = Amt->getValue();
275     // Cannot analyze non-byte shifts.
276     if ((ShAmt & 7) != 0)
277       return nullptr;
278     ShAmt.lshrInPlace(3);
279 
280     // If the extract is known to be all zeros, return zero.
281     if (ShAmt.uge(CSize - ByteStart))
282       return Constant::getNullValue(
283           IntegerType::get(CE->getContext(), ByteSize * 8));
284     // If the extract is known to be fully in the input, extract it.
285     if (ShAmt.ule(CSize - (ByteStart + ByteSize)))
286       return ExtractConstantBytes(CE->getOperand(0),
287                                   ByteStart + ShAmt.getZExtValue(), ByteSize);
288 
289     // TODO: Handle the 'partially zero' case.
290     return nullptr;
291   }
292 
293   case Instruction::Shl: {
294     ConstantInt *Amt = dyn_cast<ConstantInt>(CE->getOperand(1));
295     if (!Amt)
296       return nullptr;
297     APInt ShAmt = Amt->getValue();
298     // Cannot analyze non-byte shifts.
299     if ((ShAmt & 7) != 0)
300       return nullptr;
301     ShAmt.lshrInPlace(3);
302 
303     // If the extract is known to be all zeros, return zero.
304     if (ShAmt.uge(ByteStart + ByteSize))
305       return Constant::getNullValue(
306           IntegerType::get(CE->getContext(), ByteSize * 8));
307     // If the extract is known to be fully in the input, extract it.
308     if (ShAmt.ule(ByteStart))
309       return ExtractConstantBytes(CE->getOperand(0),
310                                   ByteStart - ShAmt.getZExtValue(), ByteSize);
311 
312     // TODO: Handle the 'partially zero' case.
313     return nullptr;
314   }
315 
316   case Instruction::ZExt: {
317     unsigned SrcBitSize =
318       cast<IntegerType>(CE->getOperand(0)->getType())->getBitWidth();
319 
320     // If extracting something that is completely zero, return 0.
321     if (ByteStart*8 >= SrcBitSize)
322       return Constant::getNullValue(IntegerType::get(CE->getContext(),
323                                                      ByteSize*8));
324 
325     // If exactly extracting the input, return it.
326     if (ByteStart == 0 && ByteSize*8 == SrcBitSize)
327       return CE->getOperand(0);
328 
329     // If extracting something completely in the input, if the input is a
330     // multiple of 8 bits, recurse.
331     if ((SrcBitSize&7) == 0 && (ByteStart+ByteSize)*8 <= SrcBitSize)
332       return ExtractConstantBytes(CE->getOperand(0), ByteStart, ByteSize);
333 
334     // Otherwise, if extracting a subset of the input, which is not multiple of
335     // 8 bits, do a shift and trunc to get the bits.
336     if ((ByteStart+ByteSize)*8 < SrcBitSize) {
337       assert((SrcBitSize&7) && "Shouldn't get byte sized case here");
338       Constant *Res = CE->getOperand(0);
339       if (ByteStart)
340         Res = ConstantExpr::getLShr(Res,
341                                  ConstantInt::get(Res->getType(), ByteStart*8));
342       return ConstantExpr::getTrunc(Res, IntegerType::get(C->getContext(),
343                                                           ByteSize*8));
344     }
345 
346     // TODO: Handle the 'partially zero' case.
347     return nullptr;
348   }
349   }
350 }
351 
ConstantFoldCastInstruction(unsigned opc,Constant * V,Type * DestTy)352 Constant *llvm::ConstantFoldCastInstruction(unsigned opc, Constant *V,
353                                             Type *DestTy) {
354   if (isa<PoisonValue>(V))
355     return PoisonValue::get(DestTy);
356 
357   if (isa<UndefValue>(V)) {
358     // zext(undef) = 0, because the top bits will be zero.
359     // sext(undef) = 0, because the top bits will all be the same.
360     // [us]itofp(undef) = 0, because the result value is bounded.
361     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
362         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
363       return Constant::getNullValue(DestTy);
364     return UndefValue::get(DestTy);
365   }
366 
367   if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() &&
368       opc != Instruction::AddrSpaceCast)
369     return Constant::getNullValue(DestTy);
370 
371   // If the cast operand is a constant expression, there's a few things we can
372   // do to try to simplify it.
373   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
374     if (CE->isCast()) {
375       // Try hard to fold cast of cast because they are often eliminable.
376       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
377         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);
378     } else if (CE->getOpcode() == Instruction::GetElementPtr &&
379                // Do not fold addrspacecast (gep 0, .., 0). It might make the
380                // addrspacecast uncanonicalized.
381                opc != Instruction::AddrSpaceCast &&
382                // Do not fold bitcast (gep) with inrange index, as this loses
383                // information.
384                !cast<GEPOperator>(CE)->getInRangeIndex().hasValue() &&
385                // Do not fold if the gep type is a vector, as bitcasting
386                // operand 0 of a vector gep will result in a bitcast between
387                // different sizes.
388                !CE->getType()->isVectorTy()) {
389       // If all of the indexes in the GEP are null values, there is no pointer
390       // adjustment going on.  We might as well cast the source pointer.
391       bool isAllNull = true;
392       for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
393         if (!CE->getOperand(i)->isNullValue()) {
394           isAllNull = false;
395           break;
396         }
397       if (isAllNull)
398         // This is casting one pointer type to another, always BitCast
399         return ConstantExpr::getPointerCast(CE->getOperand(0), DestTy);
400     }
401   }
402 
403   // If the cast operand is a constant vector, perform the cast by
404   // operating on each element. In the cast of bitcasts, the element
405   // count may be mismatched; don't attempt to handle that here.
406   if ((isa<ConstantVector>(V) || isa<ConstantDataVector>(V)) &&
407       DestTy->isVectorTy() &&
408       cast<FixedVectorType>(DestTy)->getNumElements() ==
409           cast<FixedVectorType>(V->getType())->getNumElements()) {
410     VectorType *DestVecTy = cast<VectorType>(DestTy);
411     Type *DstEltTy = DestVecTy->getElementType();
412     // Fast path for splatted constants.
413     if (Constant *Splat = V->getSplatValue()) {
414       return ConstantVector::getSplat(
415           cast<VectorType>(DestTy)->getElementCount(),
416           ConstantExpr::getCast(opc, Splat, DstEltTy));
417     }
418     SmallVector<Constant *, 16> res;
419     Type *Ty = IntegerType::get(V->getContext(), 32);
420     for (unsigned i = 0,
421                   e = cast<FixedVectorType>(V->getType())->getNumElements();
422          i != e; ++i) {
423       Constant *C =
424         ConstantExpr::getExtractElement(V, ConstantInt::get(Ty, i));
425       res.push_back(ConstantExpr::getCast(opc, C, DstEltTy));
426     }
427     return ConstantVector::get(res);
428   }
429 
430   // We actually have to do a cast now. Perform the cast according to the
431   // opcode specified.
432   switch (opc) {
433   default:
434     llvm_unreachable("Failed to cast constant expression");
435   case Instruction::FPTrunc:
436   case Instruction::FPExt:
437     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
438       bool ignored;
439       APFloat Val = FPC->getValueAPF();
440       Val.convert(DestTy->isHalfTy() ? APFloat::IEEEhalf() :
441                   DestTy->isFloatTy() ? APFloat::IEEEsingle() :
442                   DestTy->isDoubleTy() ? APFloat::IEEEdouble() :
443                   DestTy->isX86_FP80Ty() ? APFloat::x87DoubleExtended() :
444                   DestTy->isFP128Ty() ? APFloat::IEEEquad() :
445                   DestTy->isPPC_FP128Ty() ? APFloat::PPCDoubleDouble() :
446                   APFloat::Bogus(),
447                   APFloat::rmNearestTiesToEven, &ignored);
448       return ConstantFP::get(V->getContext(), Val);
449     }
450     return nullptr; // Can't fold.
451   case Instruction::FPToUI:
452   case Instruction::FPToSI:
453     if (ConstantFP *FPC = dyn_cast<ConstantFP>(V)) {
454       const APFloat &V = FPC->getValueAPF();
455       bool ignored;
456       uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
457       APSInt IntVal(DestBitWidth, opc == Instruction::FPToUI);
458       if (APFloat::opInvalidOp ==
459           V.convertToInteger(IntVal, APFloat::rmTowardZero, &ignored)) {
460         // Undefined behavior invoked - the destination type can't represent
461         // the input constant.
462         return PoisonValue::get(DestTy);
463       }
464       return ConstantInt::get(FPC->getContext(), IntVal);
465     }
466     return nullptr; // Can't fold.
467   case Instruction::IntToPtr:   //always treated as unsigned
468     if (V->isNullValue())       // Is it an integral null value?
469       return ConstantPointerNull::get(cast<PointerType>(DestTy));
470     return nullptr;                   // Other pointer types cannot be casted
471   case Instruction::PtrToInt:   // always treated as unsigned
472     // Is it a null pointer value?
473     if (V->isNullValue())
474       return ConstantInt::get(DestTy, 0);
475     // Other pointer types cannot be casted
476     return nullptr;
477   case Instruction::UIToFP:
478   case Instruction::SIToFP:
479     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
480       const APInt &api = CI->getValue();
481       APFloat apf(DestTy->getFltSemantics(),
482                   APInt::getZero(DestTy->getPrimitiveSizeInBits()));
483       apf.convertFromAPInt(api, opc==Instruction::SIToFP,
484                            APFloat::rmNearestTiesToEven);
485       return ConstantFP::get(V->getContext(), apf);
486     }
487     return nullptr;
488   case Instruction::ZExt:
489     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
490       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
491       return ConstantInt::get(V->getContext(),
492                               CI->getValue().zext(BitWidth));
493     }
494     return nullptr;
495   case Instruction::SExt:
496     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
497       uint32_t BitWidth = cast<IntegerType>(DestTy)->getBitWidth();
498       return ConstantInt::get(V->getContext(),
499                               CI->getValue().sext(BitWidth));
500     }
501     return nullptr;
502   case Instruction::Trunc: {
503     if (V->getType()->isVectorTy())
504       return nullptr;
505 
506     uint32_t DestBitWidth = cast<IntegerType>(DestTy)->getBitWidth();
507     if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
508       return ConstantInt::get(V->getContext(),
509                               CI->getValue().trunc(DestBitWidth));
510     }
511 
512     // The input must be a constantexpr.  See if we can simplify this based on
513     // the bytes we are demanding.  Only do this if the source and dest are an
514     // even multiple of a byte.
515     if ((DestBitWidth & 7) == 0 &&
516         (cast<IntegerType>(V->getType())->getBitWidth() & 7) == 0)
517       if (Constant *Res = ExtractConstantBytes(V, 0, DestBitWidth / 8))
518         return Res;
519 
520     return nullptr;
521   }
522   case Instruction::BitCast:
523     return FoldBitCast(V, DestTy);
524   case Instruction::AddrSpaceCast:
525     return nullptr;
526   }
527 }
528 
ConstantFoldSelectInstruction(Constant * Cond,Constant * V1,Constant * V2)529 Constant *llvm::ConstantFoldSelectInstruction(Constant *Cond,
530                                               Constant *V1, Constant *V2) {
531   // Check for i1 and vector true/false conditions.
532   if (Cond->isNullValue()) return V2;
533   if (Cond->isAllOnesValue()) return V1;
534 
535   // If the condition is a vector constant, fold the result elementwise.
536   if (ConstantVector *CondV = dyn_cast<ConstantVector>(Cond)) {
537     auto *V1VTy = CondV->getType();
538     SmallVector<Constant*, 16> Result;
539     Type *Ty = IntegerType::get(CondV->getContext(), 32);
540     for (unsigned i = 0, e = V1VTy->getNumElements(); i != e; ++i) {
541       Constant *V;
542       Constant *V1Element = ConstantExpr::getExtractElement(V1,
543                                                     ConstantInt::get(Ty, i));
544       Constant *V2Element = ConstantExpr::getExtractElement(V2,
545                                                     ConstantInt::get(Ty, i));
546       auto *Cond = cast<Constant>(CondV->getOperand(i));
547       if (isa<PoisonValue>(Cond)) {
548         V = PoisonValue::get(V1Element->getType());
549       } else if (V1Element == V2Element) {
550         V = V1Element;
551       } else if (isa<UndefValue>(Cond)) {
552         V = isa<UndefValue>(V1Element) ? V1Element : V2Element;
553       } else {
554         if (!isa<ConstantInt>(Cond)) break;
555         V = Cond->isNullValue() ? V2Element : V1Element;
556       }
557       Result.push_back(V);
558     }
559 
560     // If we were able to build the vector, return it.
561     if (Result.size() == V1VTy->getNumElements())
562       return ConstantVector::get(Result);
563   }
564 
565   if (isa<PoisonValue>(Cond))
566     return PoisonValue::get(V1->getType());
567 
568   if (isa<UndefValue>(Cond)) {
569     if (isa<UndefValue>(V1)) return V1;
570     return V2;
571   }
572 
573   if (V1 == V2) return V1;
574 
575   if (isa<PoisonValue>(V1))
576     return V2;
577   if (isa<PoisonValue>(V2))
578     return V1;
579 
580   // If the true or false value is undef, we can fold to the other value as
581   // long as the other value isn't poison.
582   auto NotPoison = [](Constant *C) {
583     if (isa<PoisonValue>(C))
584       return false;
585 
586     // TODO: We can analyze ConstExpr by opcode to determine if there is any
587     //       possibility of poison.
588     if (isa<ConstantExpr>(C))
589       return false;
590 
591     if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(C) ||
592         isa<ConstantPointerNull>(C) || isa<Function>(C))
593       return true;
594 
595     if (C->getType()->isVectorTy())
596       return !C->containsPoisonElement() && !C->containsConstantExpression();
597 
598     // TODO: Recursively analyze aggregates or other constants.
599     return false;
600   };
601   if (isa<UndefValue>(V1) && NotPoison(V2)) return V2;
602   if (isa<UndefValue>(V2) && NotPoison(V1)) return V1;
603 
604   if (ConstantExpr *TrueVal = dyn_cast<ConstantExpr>(V1)) {
605     if (TrueVal->getOpcode() == Instruction::Select)
606       if (TrueVal->getOperand(0) == Cond)
607         return ConstantExpr::getSelect(Cond, TrueVal->getOperand(1), V2);
608   }
609   if (ConstantExpr *FalseVal = dyn_cast<ConstantExpr>(V2)) {
610     if (FalseVal->getOpcode() == Instruction::Select)
611       if (FalseVal->getOperand(0) == Cond)
612         return ConstantExpr::getSelect(Cond, V1, FalseVal->getOperand(2));
613   }
614 
615   return nullptr;
616 }
617 
ConstantFoldExtractElementInstruction(Constant * Val,Constant * Idx)618 Constant *llvm::ConstantFoldExtractElementInstruction(Constant *Val,
619                                                       Constant *Idx) {
620   auto *ValVTy = cast<VectorType>(Val->getType());
621 
622   // extractelt poison, C -> poison
623   // extractelt C, undef -> poison
624   if (isa<PoisonValue>(Val) || isa<UndefValue>(Idx))
625     return PoisonValue::get(ValVTy->getElementType());
626 
627   // extractelt undef, C -> undef
628   if (isa<UndefValue>(Val))
629     return UndefValue::get(ValVTy->getElementType());
630 
631   auto *CIdx = dyn_cast<ConstantInt>(Idx);
632   if (!CIdx)
633     return nullptr;
634 
635   if (auto *ValFVTy = dyn_cast<FixedVectorType>(Val->getType())) {
636     // ee({w,x,y,z}, wrong_value) -> poison
637     if (CIdx->uge(ValFVTy->getNumElements()))
638       return PoisonValue::get(ValFVTy->getElementType());
639   }
640 
641   // ee (gep (ptr, idx0, ...), idx) -> gep (ee (ptr, idx), ee (idx0, idx), ...)
642   if (auto *CE = dyn_cast<ConstantExpr>(Val)) {
643     if (auto *GEP = dyn_cast<GEPOperator>(CE)) {
644       SmallVector<Constant *, 8> Ops;
645       Ops.reserve(CE->getNumOperands());
646       for (unsigned i = 0, e = CE->getNumOperands(); i != e; ++i) {
647         Constant *Op = CE->getOperand(i);
648         if (Op->getType()->isVectorTy()) {
649           Constant *ScalarOp = ConstantExpr::getExtractElement(Op, Idx);
650           if (!ScalarOp)
651             return nullptr;
652           Ops.push_back(ScalarOp);
653         } else
654           Ops.push_back(Op);
655       }
656       return CE->getWithOperands(Ops, ValVTy->getElementType(), false,
657                                  GEP->getSourceElementType());
658     } else if (CE->getOpcode() == Instruction::InsertElement) {
659       if (const auto *IEIdx = dyn_cast<ConstantInt>(CE->getOperand(2))) {
660         if (APSInt::isSameValue(APSInt(IEIdx->getValue()),
661                                 APSInt(CIdx->getValue()))) {
662           return CE->getOperand(1);
663         } else {
664           return ConstantExpr::getExtractElement(CE->getOperand(0), CIdx);
665         }
666       }
667     }
668   }
669 
670   if (Constant *C = Val->getAggregateElement(CIdx))
671     return C;
672 
673   // Lane < Splat minimum vector width => extractelt Splat(x), Lane -> x
674   if (CIdx->getValue().ult(ValVTy->getElementCount().getKnownMinValue())) {
675     if (Constant *SplatVal = Val->getSplatValue())
676       return SplatVal;
677   }
678 
679   return nullptr;
680 }
681 
ConstantFoldInsertElementInstruction(Constant * Val,Constant * Elt,Constant * Idx)682 Constant *llvm::ConstantFoldInsertElementInstruction(Constant *Val,
683                                                      Constant *Elt,
684                                                      Constant *Idx) {
685   if (isa<UndefValue>(Idx))
686     return PoisonValue::get(Val->getType());
687 
688   ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
689   if (!CIdx) return nullptr;
690 
691   // Do not iterate on scalable vector. The num of elements is unknown at
692   // compile-time.
693   if (isa<ScalableVectorType>(Val->getType()))
694     return nullptr;
695 
696   auto *ValTy = cast<FixedVectorType>(Val->getType());
697 
698   unsigned NumElts = ValTy->getNumElements();
699   if (CIdx->uge(NumElts))
700     return PoisonValue::get(Val->getType());
701 
702   SmallVector<Constant*, 16> Result;
703   Result.reserve(NumElts);
704   auto *Ty = Type::getInt32Ty(Val->getContext());
705   uint64_t IdxVal = CIdx->getZExtValue();
706   for (unsigned i = 0; i != NumElts; ++i) {
707     if (i == IdxVal) {
708       Result.push_back(Elt);
709       continue;
710     }
711 
712     Constant *C = ConstantExpr::getExtractElement(Val, ConstantInt::get(Ty, i));
713     Result.push_back(C);
714   }
715 
716   return ConstantVector::get(Result);
717 }
718 
ConstantFoldShuffleVectorInstruction(Constant * V1,Constant * V2,ArrayRef<int> Mask)719 Constant *llvm::ConstantFoldShuffleVectorInstruction(Constant *V1, Constant *V2,
720                                                      ArrayRef<int> Mask) {
721   auto *V1VTy = cast<VectorType>(V1->getType());
722   unsigned MaskNumElts = Mask.size();
723   auto MaskEltCount =
724       ElementCount::get(MaskNumElts, isa<ScalableVectorType>(V1VTy));
725   Type *EltTy = V1VTy->getElementType();
726 
727   // Undefined shuffle mask -> undefined value.
728   if (all_of(Mask, [](int Elt) { return Elt == UndefMaskElem; })) {
729     return UndefValue::get(FixedVectorType::get(EltTy, MaskNumElts));
730   }
731 
732   // If the mask is all zeros this is a splat, no need to go through all
733   // elements.
734   if (all_of(Mask, [](int Elt) { return Elt == 0; }) &&
735       !MaskEltCount.isScalable()) {
736     Type *Ty = IntegerType::get(V1->getContext(), 32);
737     Constant *Elt =
738         ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, 0));
739     return ConstantVector::getSplat(MaskEltCount, Elt);
740   }
741   // Do not iterate on scalable vector. The num of elements is unknown at
742   // compile-time.
743   if (isa<ScalableVectorType>(V1VTy))
744     return nullptr;
745 
746   unsigned SrcNumElts = V1VTy->getElementCount().getKnownMinValue();
747 
748   // Loop over the shuffle mask, evaluating each element.
749   SmallVector<Constant*, 32> Result;
750   for (unsigned i = 0; i != MaskNumElts; ++i) {
751     int Elt = Mask[i];
752     if (Elt == -1) {
753       Result.push_back(UndefValue::get(EltTy));
754       continue;
755     }
756     Constant *InElt;
757     if (unsigned(Elt) >= SrcNumElts*2)
758       InElt = UndefValue::get(EltTy);
759     else if (unsigned(Elt) >= SrcNumElts) {
760       Type *Ty = IntegerType::get(V2->getContext(), 32);
761       InElt =
762         ConstantExpr::getExtractElement(V2,
763                                         ConstantInt::get(Ty, Elt - SrcNumElts));
764     } else {
765       Type *Ty = IntegerType::get(V1->getContext(), 32);
766       InElt = ConstantExpr::getExtractElement(V1, ConstantInt::get(Ty, Elt));
767     }
768     Result.push_back(InElt);
769   }
770 
771   return ConstantVector::get(Result);
772 }
773 
ConstantFoldExtractValueInstruction(Constant * Agg,ArrayRef<unsigned> Idxs)774 Constant *llvm::ConstantFoldExtractValueInstruction(Constant *Agg,
775                                                     ArrayRef<unsigned> Idxs) {
776   // Base case: no indices, so return the entire value.
777   if (Idxs.empty())
778     return Agg;
779 
780   if (Constant *C = Agg->getAggregateElement(Idxs[0]))
781     return ConstantFoldExtractValueInstruction(C, Idxs.slice(1));
782 
783   return nullptr;
784 }
785 
ConstantFoldInsertValueInstruction(Constant * Agg,Constant * Val,ArrayRef<unsigned> Idxs)786 Constant *llvm::ConstantFoldInsertValueInstruction(Constant *Agg,
787                                                    Constant *Val,
788                                                    ArrayRef<unsigned> Idxs) {
789   // Base case: no indices, so replace the entire value.
790   if (Idxs.empty())
791     return Val;
792 
793   unsigned NumElts;
794   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
795     NumElts = ST->getNumElements();
796   else
797     NumElts = cast<ArrayType>(Agg->getType())->getNumElements();
798 
799   SmallVector<Constant*, 32> Result;
800   for (unsigned i = 0; i != NumElts; ++i) {
801     Constant *C = Agg->getAggregateElement(i);
802     if (!C) return nullptr;
803 
804     if (Idxs[0] == i)
805       C = ConstantFoldInsertValueInstruction(C, Val, Idxs.slice(1));
806 
807     Result.push_back(C);
808   }
809 
810   if (StructType *ST = dyn_cast<StructType>(Agg->getType()))
811     return ConstantStruct::get(ST, Result);
812   return ConstantArray::get(cast<ArrayType>(Agg->getType()), Result);
813 }
814 
ConstantFoldUnaryInstruction(unsigned Opcode,Constant * C)815 Constant *llvm::ConstantFoldUnaryInstruction(unsigned Opcode, Constant *C) {
816   assert(Instruction::isUnaryOp(Opcode) && "Non-unary instruction detected");
817 
818   // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
819   // vectors are always evaluated per element.
820   bool IsScalableVector = isa<ScalableVectorType>(C->getType());
821   bool HasScalarUndefOrScalableVectorUndef =
822       (!C->getType()->isVectorTy() || IsScalableVector) && isa<UndefValue>(C);
823 
824   if (HasScalarUndefOrScalableVectorUndef) {
825     switch (static_cast<Instruction::UnaryOps>(Opcode)) {
826     case Instruction::FNeg:
827       return C; // -undef -> undef
828     case Instruction::UnaryOpsEnd:
829       llvm_unreachable("Invalid UnaryOp");
830     }
831   }
832 
833   // Constant should not be UndefValue, unless these are vector constants.
834   assert(!HasScalarUndefOrScalableVectorUndef && "Unexpected UndefValue");
835   // We only have FP UnaryOps right now.
836   assert(!isa<ConstantInt>(C) && "Unexpected Integer UnaryOp");
837 
838   if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
839     const APFloat &CV = CFP->getValueAPF();
840     switch (Opcode) {
841     default:
842       break;
843     case Instruction::FNeg:
844       return ConstantFP::get(C->getContext(), neg(CV));
845     }
846   } else if (auto *VTy = dyn_cast<FixedVectorType>(C->getType())) {
847 
848     Type *Ty = IntegerType::get(VTy->getContext(), 32);
849     // Fast path for splatted constants.
850     if (Constant *Splat = C->getSplatValue()) {
851       Constant *Elt = ConstantExpr::get(Opcode, Splat);
852       return ConstantVector::getSplat(VTy->getElementCount(), Elt);
853     }
854 
855     // Fold each element and create a vector constant from those constants.
856     SmallVector<Constant *, 16> Result;
857     for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
858       Constant *ExtractIdx = ConstantInt::get(Ty, i);
859       Constant *Elt = ConstantExpr::getExtractElement(C, ExtractIdx);
860 
861       Result.push_back(ConstantExpr::get(Opcode, Elt));
862     }
863 
864     return ConstantVector::get(Result);
865   }
866 
867   // We don't know how to fold this.
868   return nullptr;
869 }
870 
ConstantFoldBinaryInstruction(unsigned Opcode,Constant * C1,Constant * C2)871 Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode, Constant *C1,
872                                               Constant *C2) {
873   assert(Instruction::isBinaryOp(Opcode) && "Non-binary instruction detected");
874 
875   // Simplify BinOps with their identity values first. They are no-ops and we
876   // can always return the other value, including undef or poison values.
877   // FIXME: remove unnecessary duplicated identity patterns below.
878   // FIXME: Use AllowRHSConstant with getBinOpIdentity to handle additional ops,
879   //        like X << 0 = X.
880   Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, C1->getType());
881   if (Identity) {
882     if (C1 == Identity)
883       return C2;
884     if (C2 == Identity)
885       return C1;
886   }
887 
888   // Binary operations propagate poison.
889   if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2))
890     return PoisonValue::get(C1->getType());
891 
892   // Handle scalar UndefValue and scalable vector UndefValue. Fixed-length
893   // vectors are always evaluated per element.
894   bool IsScalableVector = isa<ScalableVectorType>(C1->getType());
895   bool HasScalarUndefOrScalableVectorUndef =
896       (!C1->getType()->isVectorTy() || IsScalableVector) &&
897       (isa<UndefValue>(C1) || isa<UndefValue>(C2));
898   if (HasScalarUndefOrScalableVectorUndef) {
899     switch (static_cast<Instruction::BinaryOps>(Opcode)) {
900     case Instruction::Xor:
901       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
902         // Handle undef ^ undef -> 0 special case. This is a common
903         // idiom (misuse).
904         return Constant::getNullValue(C1->getType());
905       LLVM_FALLTHROUGH;
906     case Instruction::Add:
907     case Instruction::Sub:
908       return UndefValue::get(C1->getType());
909     case Instruction::And:
910       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef & undef -> undef
911         return C1;
912       return Constant::getNullValue(C1->getType());   // undef & X -> 0
913     case Instruction::Mul: {
914       // undef * undef -> undef
915       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
916         return C1;
917       const APInt *CV;
918       // X * undef -> undef   if X is odd
919       if (match(C1, m_APInt(CV)) || match(C2, m_APInt(CV)))
920         if ((*CV)[0])
921           return UndefValue::get(C1->getType());
922 
923       // X * undef -> 0       otherwise
924       return Constant::getNullValue(C1->getType());
925     }
926     case Instruction::SDiv:
927     case Instruction::UDiv:
928       // X / undef -> poison
929       // X / 0 -> poison
930       if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
931         return PoisonValue::get(C2->getType());
932       // undef / 1 -> undef
933       if (match(C2, m_One()))
934         return C1;
935       // undef / X -> 0       otherwise
936       return Constant::getNullValue(C1->getType());
937     case Instruction::URem:
938     case Instruction::SRem:
939       // X % undef -> poison
940       // X % 0 -> poison
941       if (match(C2, m_CombineOr(m_Undef(), m_Zero())))
942         return PoisonValue::get(C2->getType());
943       // undef % X -> 0       otherwise
944       return Constant::getNullValue(C1->getType());
945     case Instruction::Or:                          // X | undef -> -1
946       if (isa<UndefValue>(C1) && isa<UndefValue>(C2)) // undef | undef -> undef
947         return C1;
948       return Constant::getAllOnesValue(C1->getType()); // undef | X -> ~0
949     case Instruction::LShr:
950       // X >>l undef -> poison
951       if (isa<UndefValue>(C2))
952         return PoisonValue::get(C2->getType());
953       // undef >>l 0 -> undef
954       if (match(C2, m_Zero()))
955         return C1;
956       // undef >>l X -> 0
957       return Constant::getNullValue(C1->getType());
958     case Instruction::AShr:
959       // X >>a undef -> poison
960       if (isa<UndefValue>(C2))
961         return PoisonValue::get(C2->getType());
962       // undef >>a 0 -> undef
963       if (match(C2, m_Zero()))
964         return C1;
965       // TODO: undef >>a X -> poison if the shift is exact
966       // undef >>a X -> 0
967       return Constant::getNullValue(C1->getType());
968     case Instruction::Shl:
969       // X << undef -> undef
970       if (isa<UndefValue>(C2))
971         return PoisonValue::get(C2->getType());
972       // undef << 0 -> undef
973       if (match(C2, m_Zero()))
974         return C1;
975       // undef << X -> 0
976       return Constant::getNullValue(C1->getType());
977     case Instruction::FSub:
978       // -0.0 - undef --> undef (consistent with "fneg undef")
979       if (match(C1, m_NegZeroFP()) && isa<UndefValue>(C2))
980         return C2;
981       LLVM_FALLTHROUGH;
982     case Instruction::FAdd:
983     case Instruction::FMul:
984     case Instruction::FDiv:
985     case Instruction::FRem:
986       // [any flop] undef, undef -> undef
987       if (isa<UndefValue>(C1) && isa<UndefValue>(C2))
988         return C1;
989       // [any flop] C, undef -> NaN
990       // [any flop] undef, C -> NaN
991       // We could potentially specialize NaN/Inf constants vs. 'normal'
992       // constants (possibly differently depending on opcode and operand). This
993       // would allow returning undef sometimes. But it is always safe to fold to
994       // NaN because we can choose the undef operand as NaN, and any FP opcode
995       // with a NaN operand will propagate NaN.
996       return ConstantFP::getNaN(C1->getType());
997     case Instruction::BinaryOpsEnd:
998       llvm_unreachable("Invalid BinaryOp");
999     }
1000   }
1001 
1002   // Neither constant should be UndefValue, unless these are vector constants.
1003   assert((!HasScalarUndefOrScalableVectorUndef) && "Unexpected UndefValue");
1004 
1005   // Handle simplifications when the RHS is a constant int.
1006   if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1007     switch (Opcode) {
1008     case Instruction::Add:
1009       if (CI2->isZero()) return C1;                             // X + 0 == X
1010       break;
1011     case Instruction::Sub:
1012       if (CI2->isZero()) return C1;                             // X - 0 == X
1013       break;
1014     case Instruction::Mul:
1015       if (CI2->isZero()) return C2;                             // X * 0 == 0
1016       if (CI2->isOne())
1017         return C1;                                              // X * 1 == X
1018       break;
1019     case Instruction::UDiv:
1020     case Instruction::SDiv:
1021       if (CI2->isOne())
1022         return C1;                                            // X / 1 == X
1023       if (CI2->isZero())
1024         return PoisonValue::get(CI2->getType());              // X / 0 == poison
1025       break;
1026     case Instruction::URem:
1027     case Instruction::SRem:
1028       if (CI2->isOne())
1029         return Constant::getNullValue(CI2->getType());        // X % 1 == 0
1030       if (CI2->isZero())
1031         return PoisonValue::get(CI2->getType());              // X % 0 == poison
1032       break;
1033     case Instruction::And:
1034       if (CI2->isZero()) return C2;                           // X & 0 == 0
1035       if (CI2->isMinusOne())
1036         return C1;                                            // X & -1 == X
1037 
1038       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1039         // (zext i32 to i64) & 4294967295 -> (zext i32 to i64)
1040         if (CE1->getOpcode() == Instruction::ZExt) {
1041           unsigned DstWidth = CI2->getType()->getBitWidth();
1042           unsigned SrcWidth =
1043             CE1->getOperand(0)->getType()->getPrimitiveSizeInBits();
1044           APInt PossiblySetBits(APInt::getLowBitsSet(DstWidth, SrcWidth));
1045           if ((PossiblySetBits & CI2->getValue()) == PossiblySetBits)
1046             return C1;
1047         }
1048 
1049         // If and'ing the address of a global with a constant, fold it.
1050         if (CE1->getOpcode() == Instruction::PtrToInt &&
1051             isa<GlobalValue>(CE1->getOperand(0))) {
1052           GlobalValue *GV = cast<GlobalValue>(CE1->getOperand(0));
1053 
1054           MaybeAlign GVAlign;
1055 
1056           if (Module *TheModule = GV->getParent()) {
1057             const DataLayout &DL = TheModule->getDataLayout();
1058             GVAlign = GV->getPointerAlignment(DL);
1059 
1060             // If the function alignment is not specified then assume that it
1061             // is 4.
1062             // This is dangerous; on x86, the alignment of the pointer
1063             // corresponds to the alignment of the function, but might be less
1064             // than 4 if it isn't explicitly specified.
1065             // However, a fix for this behaviour was reverted because it
1066             // increased code size (see https://reviews.llvm.org/D55115)
1067             // FIXME: This code should be deleted once existing targets have
1068             // appropriate defaults
1069             if (isa<Function>(GV) && !DL.getFunctionPtrAlign())
1070               GVAlign = Align(4);
1071           } else if (isa<Function>(GV)) {
1072             // Without a datalayout we have to assume the worst case: that the
1073             // function pointer isn't aligned at all.
1074             GVAlign = llvm::None;
1075           } else if (isa<GlobalVariable>(GV)) {
1076             GVAlign = cast<GlobalVariable>(GV)->getAlign();
1077           }
1078 
1079           if (GVAlign && *GVAlign > 1) {
1080             unsigned DstWidth = CI2->getType()->getBitWidth();
1081             unsigned SrcWidth = std::min(DstWidth, Log2(*GVAlign));
1082             APInt BitsNotSet(APInt::getLowBitsSet(DstWidth, SrcWidth));
1083 
1084             // If checking bits we know are clear, return zero.
1085             if ((CI2->getValue() & BitsNotSet) == CI2->getValue())
1086               return Constant::getNullValue(CI2->getType());
1087           }
1088         }
1089       }
1090       break;
1091     case Instruction::Or:
1092       if (CI2->isZero()) return C1;        // X | 0 == X
1093       if (CI2->isMinusOne())
1094         return C2;                         // X | -1 == -1
1095       break;
1096     case Instruction::Xor:
1097       if (CI2->isZero()) return C1;        // X ^ 0 == X
1098 
1099       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1100         switch (CE1->getOpcode()) {
1101         default: break;
1102         case Instruction::ICmp:
1103         case Instruction::FCmp:
1104           // cmp pred ^ true -> cmp !pred
1105           assert(CI2->isOne());
1106           CmpInst::Predicate pred = (CmpInst::Predicate)CE1->getPredicate();
1107           pred = CmpInst::getInversePredicate(pred);
1108           return ConstantExpr::getCompare(pred, CE1->getOperand(0),
1109                                           CE1->getOperand(1));
1110         }
1111       }
1112       break;
1113     case Instruction::AShr:
1114       // ashr (zext C to Ty), C2 -> lshr (zext C, CSA), C2
1115       if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1))
1116         if (CE1->getOpcode() == Instruction::ZExt)  // Top bits known zero.
1117           return ConstantExpr::getLShr(C1, C2);
1118       break;
1119     }
1120   } else if (isa<ConstantInt>(C1)) {
1121     // If C1 is a ConstantInt and C2 is not, swap the operands.
1122     if (Instruction::isCommutative(Opcode))
1123       return ConstantExpr::get(Opcode, C2, C1);
1124   }
1125 
1126   if (ConstantInt *CI1 = dyn_cast<ConstantInt>(C1)) {
1127     if (ConstantInt *CI2 = dyn_cast<ConstantInt>(C2)) {
1128       const APInt &C1V = CI1->getValue();
1129       const APInt &C2V = CI2->getValue();
1130       switch (Opcode) {
1131       default:
1132         break;
1133       case Instruction::Add:
1134         return ConstantInt::get(CI1->getContext(), C1V + C2V);
1135       case Instruction::Sub:
1136         return ConstantInt::get(CI1->getContext(), C1V - C2V);
1137       case Instruction::Mul:
1138         return ConstantInt::get(CI1->getContext(), C1V * C2V);
1139       case Instruction::UDiv:
1140         assert(!CI2->isZero() && "Div by zero handled above");
1141         return ConstantInt::get(CI1->getContext(), C1V.udiv(C2V));
1142       case Instruction::SDiv:
1143         assert(!CI2->isZero() && "Div by zero handled above");
1144         if (C2V.isAllOnes() && C1V.isMinSignedValue())
1145           return PoisonValue::get(CI1->getType());   // MIN_INT / -1 -> poison
1146         return ConstantInt::get(CI1->getContext(), C1V.sdiv(C2V));
1147       case Instruction::URem:
1148         assert(!CI2->isZero() && "Div by zero handled above");
1149         return ConstantInt::get(CI1->getContext(), C1V.urem(C2V));
1150       case Instruction::SRem:
1151         assert(!CI2->isZero() && "Div by zero handled above");
1152         if (C2V.isAllOnes() && C1V.isMinSignedValue())
1153           return PoisonValue::get(CI1->getType());   // MIN_INT % -1 -> poison
1154         return ConstantInt::get(CI1->getContext(), C1V.srem(C2V));
1155       case Instruction::And:
1156         return ConstantInt::get(CI1->getContext(), C1V & C2V);
1157       case Instruction::Or:
1158         return ConstantInt::get(CI1->getContext(), C1V | C2V);
1159       case Instruction::Xor:
1160         return ConstantInt::get(CI1->getContext(), C1V ^ C2V);
1161       case Instruction::Shl:
1162         if (C2V.ult(C1V.getBitWidth()))
1163           return ConstantInt::get(CI1->getContext(), C1V.shl(C2V));
1164         return PoisonValue::get(C1->getType()); // too big shift is poison
1165       case Instruction::LShr:
1166         if (C2V.ult(C1V.getBitWidth()))
1167           return ConstantInt::get(CI1->getContext(), C1V.lshr(C2V));
1168         return PoisonValue::get(C1->getType()); // too big shift is poison
1169       case Instruction::AShr:
1170         if (C2V.ult(C1V.getBitWidth()))
1171           return ConstantInt::get(CI1->getContext(), C1V.ashr(C2V));
1172         return PoisonValue::get(C1->getType()); // too big shift is poison
1173       }
1174     }
1175 
1176     switch (Opcode) {
1177     case Instruction::SDiv:
1178     case Instruction::UDiv:
1179     case Instruction::URem:
1180     case Instruction::SRem:
1181     case Instruction::LShr:
1182     case Instruction::AShr:
1183     case Instruction::Shl:
1184       if (CI1->isZero()) return C1;
1185       break;
1186     default:
1187       break;
1188     }
1189   } else if (ConstantFP *CFP1 = dyn_cast<ConstantFP>(C1)) {
1190     if (ConstantFP *CFP2 = dyn_cast<ConstantFP>(C2)) {
1191       const APFloat &C1V = CFP1->getValueAPF();
1192       const APFloat &C2V = CFP2->getValueAPF();
1193       APFloat C3V = C1V;  // copy for modification
1194       switch (Opcode) {
1195       default:
1196         break;
1197       case Instruction::FAdd:
1198         (void)C3V.add(C2V, APFloat::rmNearestTiesToEven);
1199         return ConstantFP::get(C1->getContext(), C3V);
1200       case Instruction::FSub:
1201         (void)C3V.subtract(C2V, APFloat::rmNearestTiesToEven);
1202         return ConstantFP::get(C1->getContext(), C3V);
1203       case Instruction::FMul:
1204         (void)C3V.multiply(C2V, APFloat::rmNearestTiesToEven);
1205         return ConstantFP::get(C1->getContext(), C3V);
1206       case Instruction::FDiv:
1207         (void)C3V.divide(C2V, APFloat::rmNearestTiesToEven);
1208         return ConstantFP::get(C1->getContext(), C3V);
1209       case Instruction::FRem:
1210         (void)C3V.mod(C2V);
1211         return ConstantFP::get(C1->getContext(), C3V);
1212       }
1213     }
1214   } else if (auto *VTy = dyn_cast<VectorType>(C1->getType())) {
1215     // Fast path for splatted constants.
1216     if (Constant *C2Splat = C2->getSplatValue()) {
1217       if (Instruction::isIntDivRem(Opcode) && C2Splat->isNullValue())
1218         return PoisonValue::get(VTy);
1219       if (Constant *C1Splat = C1->getSplatValue()) {
1220         return ConstantVector::getSplat(
1221             VTy->getElementCount(),
1222             ConstantExpr::get(Opcode, C1Splat, C2Splat));
1223       }
1224     }
1225 
1226     if (auto *FVTy = dyn_cast<FixedVectorType>(VTy)) {
1227       // Fold each element and create a vector constant from those constants.
1228       SmallVector<Constant*, 16> Result;
1229       Type *Ty = IntegerType::get(FVTy->getContext(), 32);
1230       for (unsigned i = 0, e = FVTy->getNumElements(); i != e; ++i) {
1231         Constant *ExtractIdx = ConstantInt::get(Ty, i);
1232         Constant *LHS = ConstantExpr::getExtractElement(C1, ExtractIdx);
1233         Constant *RHS = ConstantExpr::getExtractElement(C2, ExtractIdx);
1234 
1235         // If any element of a divisor vector is zero, the whole op is poison.
1236         if (Instruction::isIntDivRem(Opcode) && RHS->isNullValue())
1237           return PoisonValue::get(VTy);
1238 
1239         Result.push_back(ConstantExpr::get(Opcode, LHS, RHS));
1240       }
1241 
1242       return ConstantVector::get(Result);
1243     }
1244   }
1245 
1246   if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
1247     // There are many possible foldings we could do here.  We should probably
1248     // at least fold add of a pointer with an integer into the appropriate
1249     // getelementptr.  This will improve alias analysis a bit.
1250 
1251     // Given ((a + b) + c), if (b + c) folds to something interesting, return
1252     // (a + (b + c)).
1253     if (Instruction::isAssociative(Opcode) && CE1->getOpcode() == Opcode) {
1254       Constant *T = ConstantExpr::get(Opcode, CE1->getOperand(1), C2);
1255       if (!isa<ConstantExpr>(T) || cast<ConstantExpr>(T)->getOpcode() != Opcode)
1256         return ConstantExpr::get(Opcode, CE1->getOperand(0), T);
1257     }
1258   } else if (isa<ConstantExpr>(C2)) {
1259     // If C2 is a constant expr and C1 isn't, flop them around and fold the
1260     // other way if possible.
1261     if (Instruction::isCommutative(Opcode))
1262       return ConstantFoldBinaryInstruction(Opcode, C2, C1);
1263   }
1264 
1265   // i1 can be simplified in many cases.
1266   if (C1->getType()->isIntegerTy(1)) {
1267     switch (Opcode) {
1268     case Instruction::Add:
1269     case Instruction::Sub:
1270       return ConstantExpr::getXor(C1, C2);
1271     case Instruction::Mul:
1272       return ConstantExpr::getAnd(C1, C2);
1273     case Instruction::Shl:
1274     case Instruction::LShr:
1275     case Instruction::AShr:
1276       // We can assume that C2 == 0.  If it were one the result would be
1277       // undefined because the shift value is as large as the bitwidth.
1278       return C1;
1279     case Instruction::SDiv:
1280     case Instruction::UDiv:
1281       // We can assume that C2 == 1.  If it were zero the result would be
1282       // undefined through division by zero.
1283       return C1;
1284     case Instruction::URem:
1285     case Instruction::SRem:
1286       // We can assume that C2 == 1.  If it were zero the result would be
1287       // undefined through division by zero.
1288       return ConstantInt::getFalse(C1->getContext());
1289     default:
1290       break;
1291     }
1292   }
1293 
1294   // We don't know how to fold this.
1295   return nullptr;
1296 }
1297 
1298 /// This type is zero-sized if it's an array or structure of zero-sized types.
1299 /// The only leaf zero-sized type is an empty structure.
isMaybeZeroSizedType(Type * Ty)1300 static bool isMaybeZeroSizedType(Type *Ty) {
1301   if (StructType *STy = dyn_cast<StructType>(Ty)) {
1302     if (STy->isOpaque()) return true;  // Can't say.
1303 
1304     // If all of elements have zero size, this does too.
1305     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
1306       if (!isMaybeZeroSizedType(STy->getElementType(i))) return false;
1307     return true;
1308 
1309   } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1310     return isMaybeZeroSizedType(ATy->getElementType());
1311   }
1312   return false;
1313 }
1314 
1315 /// Compare the two constants as though they were getelementptr indices.
1316 /// This allows coercion of the types to be the same thing.
1317 ///
1318 /// If the two constants are the "same" (after coercion), return 0.  If the
1319 /// first is less than the second, return -1, if the second is less than the
1320 /// first, return 1.  If the constants are not integral, return -2.
1321 ///
IdxCompare(Constant * C1,Constant * C2,Type * ElTy)1322 static int IdxCompare(Constant *C1, Constant *C2, Type *ElTy) {
1323   if (C1 == C2) return 0;
1324 
1325   // Ok, we found a different index.  If they are not ConstantInt, we can't do
1326   // anything with them.
1327   if (!isa<ConstantInt>(C1) || !isa<ConstantInt>(C2))
1328     return -2; // don't know!
1329 
1330   // We cannot compare the indices if they don't fit in an int64_t.
1331   if (cast<ConstantInt>(C1)->getValue().getActiveBits() > 64 ||
1332       cast<ConstantInt>(C2)->getValue().getActiveBits() > 64)
1333     return -2; // don't know!
1334 
1335   // Ok, we have two differing integer indices.  Sign extend them to be the same
1336   // type.
1337   int64_t C1Val = cast<ConstantInt>(C1)->getSExtValue();
1338   int64_t C2Val = cast<ConstantInt>(C2)->getSExtValue();
1339 
1340   if (C1Val == C2Val) return 0;  // They are equal
1341 
1342   // If the type being indexed over is really just a zero sized type, there is
1343   // no pointer difference being made here.
1344   if (isMaybeZeroSizedType(ElTy))
1345     return -2; // dunno.
1346 
1347   // If they are really different, now that they are the same type, then we
1348   // found a difference!
1349   if (C1Val < C2Val)
1350     return -1;
1351   else
1352     return 1;
1353 }
1354 
1355 /// This function determines if there is anything we can decide about the two
1356 /// constants provided. This doesn't need to handle simple things like
1357 /// ConstantFP comparisons, but should instead handle ConstantExprs.
1358 /// If we can determine that the two constants have a particular relation to
1359 /// each other, we should return the corresponding FCmpInst predicate,
1360 /// otherwise return FCmpInst::BAD_FCMP_PREDICATE. This is used below in
1361 /// ConstantFoldCompareInstruction.
1362 ///
1363 /// To simplify this code we canonicalize the relation so that the first
1364 /// operand is always the most "complex" of the two.  We consider ConstantFP
1365 /// to be the simplest, and ConstantExprs to be the most complex.
evaluateFCmpRelation(Constant * V1,Constant * V2)1366 static FCmpInst::Predicate evaluateFCmpRelation(Constant *V1, Constant *V2) {
1367   assert(V1->getType() == V2->getType() &&
1368          "Cannot compare values of different types!");
1369 
1370   // We do not know if a constant expression will evaluate to a number or NaN.
1371   // Therefore, we can only say that the relation is unordered or equal.
1372   if (V1 == V2) return FCmpInst::FCMP_UEQ;
1373 
1374   if (!isa<ConstantExpr>(V1)) {
1375     if (!isa<ConstantExpr>(V2)) {
1376       // Simple case, use the standard constant folder.
1377       ConstantInt *R = nullptr;
1378       R = dyn_cast<ConstantInt>(
1379                       ConstantExpr::getFCmp(FCmpInst::FCMP_OEQ, V1, V2));
1380       if (R && !R->isZero())
1381         return FCmpInst::FCMP_OEQ;
1382       R = dyn_cast<ConstantInt>(
1383                       ConstantExpr::getFCmp(FCmpInst::FCMP_OLT, V1, V2));
1384       if (R && !R->isZero())
1385         return FCmpInst::FCMP_OLT;
1386       R = dyn_cast<ConstantInt>(
1387                       ConstantExpr::getFCmp(FCmpInst::FCMP_OGT, V1, V2));
1388       if (R && !R->isZero())
1389         return FCmpInst::FCMP_OGT;
1390 
1391       // Nothing more we can do
1392       return FCmpInst::BAD_FCMP_PREDICATE;
1393     }
1394 
1395     // If the first operand is simple and second is ConstantExpr, swap operands.
1396     FCmpInst::Predicate SwappedRelation = evaluateFCmpRelation(V2, V1);
1397     if (SwappedRelation != FCmpInst::BAD_FCMP_PREDICATE)
1398       return FCmpInst::getSwappedPredicate(SwappedRelation);
1399   } else {
1400     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1401     // constantexpr or a simple constant.
1402     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1403     switch (CE1->getOpcode()) {
1404     case Instruction::FPTrunc:
1405     case Instruction::FPExt:
1406     case Instruction::UIToFP:
1407     case Instruction::SIToFP:
1408       // We might be able to do something with these but we don't right now.
1409       break;
1410     default:
1411       break;
1412     }
1413   }
1414   // There are MANY other foldings that we could perform here.  They will
1415   // probably be added on demand, as they seem needed.
1416   return FCmpInst::BAD_FCMP_PREDICATE;
1417 }
1418 
areGlobalsPotentiallyEqual(const GlobalValue * GV1,const GlobalValue * GV2)1419 static ICmpInst::Predicate areGlobalsPotentiallyEqual(const GlobalValue *GV1,
1420                                                       const GlobalValue *GV2) {
1421   auto isGlobalUnsafeForEquality = [](const GlobalValue *GV) {
1422     if (GV->isInterposable() || GV->hasGlobalUnnamedAddr())
1423       return true;
1424     if (const auto *GVar = dyn_cast<GlobalVariable>(GV)) {
1425       Type *Ty = GVar->getValueType();
1426       // A global with opaque type might end up being zero sized.
1427       if (!Ty->isSized())
1428         return true;
1429       // A global with an empty type might lie at the address of any other
1430       // global.
1431       if (Ty->isEmptyTy())
1432         return true;
1433     }
1434     return false;
1435   };
1436   // Don't try to decide equality of aliases.
1437   if (!isa<GlobalAlias>(GV1) && !isa<GlobalAlias>(GV2))
1438     if (!isGlobalUnsafeForEquality(GV1) && !isGlobalUnsafeForEquality(GV2))
1439       return ICmpInst::ICMP_NE;
1440   return ICmpInst::BAD_ICMP_PREDICATE;
1441 }
1442 
1443 /// This function determines if there is anything we can decide about the two
1444 /// constants provided. This doesn't need to handle simple things like integer
1445 /// comparisons, but should instead handle ConstantExprs and GlobalValues.
1446 /// If we can determine that the two constants have a particular relation to
1447 /// each other, we should return the corresponding ICmp predicate, otherwise
1448 /// return ICmpInst::BAD_ICMP_PREDICATE.
1449 ///
1450 /// To simplify this code we canonicalize the relation so that the first
1451 /// operand is always the most "complex" of the two.  We consider simple
1452 /// constants (like ConstantInt) to be the simplest, followed by
1453 /// GlobalValues, followed by ConstantExpr's (the most complex).
1454 ///
evaluateICmpRelation(Constant * V1,Constant * V2,bool isSigned)1455 static ICmpInst::Predicate evaluateICmpRelation(Constant *V1, Constant *V2,
1456                                                 bool isSigned) {
1457   assert(V1->getType() == V2->getType() &&
1458          "Cannot compare different types of values!");
1459   if (V1 == V2) return ICmpInst::ICMP_EQ;
1460 
1461   if (!isa<ConstantExpr>(V1) && !isa<GlobalValue>(V1) &&
1462       !isa<BlockAddress>(V1)) {
1463     if (!isa<GlobalValue>(V2) && !isa<ConstantExpr>(V2) &&
1464         !isa<BlockAddress>(V2)) {
1465       // We distilled this down to a simple case, use the standard constant
1466       // folder.
1467       ConstantInt *R = nullptr;
1468       ICmpInst::Predicate pred = ICmpInst::ICMP_EQ;
1469       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1470       if (R && !R->isZero())
1471         return pred;
1472       pred = isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1473       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1474       if (R && !R->isZero())
1475         return pred;
1476       pred = isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1477       R = dyn_cast<ConstantInt>(ConstantExpr::getICmp(pred, V1, V2));
1478       if (R && !R->isZero())
1479         return pred;
1480 
1481       // If we couldn't figure it out, bail.
1482       return ICmpInst::BAD_ICMP_PREDICATE;
1483     }
1484 
1485     // If the first operand is simple, swap operands.
1486     ICmpInst::Predicate SwappedRelation =
1487       evaluateICmpRelation(V2, V1, isSigned);
1488     if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1489       return ICmpInst::getSwappedPredicate(SwappedRelation);
1490 
1491   } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(V1)) {
1492     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1493       ICmpInst::Predicate SwappedRelation =
1494         evaluateICmpRelation(V2, V1, isSigned);
1495       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1496         return ICmpInst::getSwappedPredicate(SwappedRelation);
1497       return ICmpInst::BAD_ICMP_PREDICATE;
1498     }
1499 
1500     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1501     // constant (which, since the types must match, means that it's a
1502     // ConstantPointerNull).
1503     if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1504       return areGlobalsPotentiallyEqual(GV, GV2);
1505     } else if (isa<BlockAddress>(V2)) {
1506       return ICmpInst::ICMP_NE; // Globals never equal labels.
1507     } else {
1508       assert(isa<ConstantPointerNull>(V2) && "Canonicalization guarantee!");
1509       // GlobalVals can never be null unless they have external weak linkage.
1510       // We don't try to evaluate aliases here.
1511       // NOTE: We should not be doing this constant folding if null pointer
1512       // is considered valid for the function. But currently there is no way to
1513       // query it from the Constant type.
1514       if (!GV->hasExternalWeakLinkage() && !isa<GlobalAlias>(GV) &&
1515           !NullPointerIsDefined(nullptr /* F */,
1516                                 GV->getType()->getAddressSpace()))
1517         return ICmpInst::ICMP_UGT;
1518     }
1519   } else if (const BlockAddress *BA = dyn_cast<BlockAddress>(V1)) {
1520     if (isa<ConstantExpr>(V2)) {  // Swap as necessary.
1521       ICmpInst::Predicate SwappedRelation =
1522         evaluateICmpRelation(V2, V1, isSigned);
1523       if (SwappedRelation != ICmpInst::BAD_ICMP_PREDICATE)
1524         return ICmpInst::getSwappedPredicate(SwappedRelation);
1525       return ICmpInst::BAD_ICMP_PREDICATE;
1526     }
1527 
1528     // Now we know that the RHS is a GlobalValue, BlockAddress or simple
1529     // constant (which, since the types must match, means that it is a
1530     // ConstantPointerNull).
1531     if (const BlockAddress *BA2 = dyn_cast<BlockAddress>(V2)) {
1532       // Block address in another function can't equal this one, but block
1533       // addresses in the current function might be the same if blocks are
1534       // empty.
1535       if (BA2->getFunction() != BA->getFunction())
1536         return ICmpInst::ICMP_NE;
1537     } else {
1538       // Block addresses aren't null, don't equal the address of globals.
1539       assert((isa<ConstantPointerNull>(V2) || isa<GlobalValue>(V2)) &&
1540              "Canonicalization guarantee!");
1541       return ICmpInst::ICMP_NE;
1542     }
1543   } else {
1544     // Ok, the LHS is known to be a constantexpr.  The RHS can be any of a
1545     // constantexpr, a global, block address, or a simple constant.
1546     ConstantExpr *CE1 = cast<ConstantExpr>(V1);
1547     Constant *CE1Op0 = CE1->getOperand(0);
1548 
1549     switch (CE1->getOpcode()) {
1550     case Instruction::Trunc:
1551     case Instruction::FPTrunc:
1552     case Instruction::FPExt:
1553     case Instruction::FPToUI:
1554     case Instruction::FPToSI:
1555       break; // We can't evaluate floating point casts or truncations.
1556 
1557     case Instruction::BitCast:
1558       // If this is a global value cast, check to see if the RHS is also a
1559       // GlobalValue.
1560       if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0))
1561         if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2))
1562           return areGlobalsPotentiallyEqual(GV, GV2);
1563       LLVM_FALLTHROUGH;
1564     case Instruction::UIToFP:
1565     case Instruction::SIToFP:
1566     case Instruction::ZExt:
1567     case Instruction::SExt:
1568       // We can't evaluate floating point casts or truncations.
1569       if (CE1Op0->getType()->isFPOrFPVectorTy())
1570         break;
1571 
1572       // If the cast is not actually changing bits, and the second operand is a
1573       // null pointer, do the comparison with the pre-casted value.
1574       if (V2->isNullValue() && CE1->getType()->isIntOrPtrTy()) {
1575         if (CE1->getOpcode() == Instruction::ZExt) isSigned = false;
1576         if (CE1->getOpcode() == Instruction::SExt) isSigned = true;
1577         return evaluateICmpRelation(CE1Op0,
1578                                     Constant::getNullValue(CE1Op0->getType()),
1579                                     isSigned);
1580       }
1581       break;
1582 
1583     case Instruction::GetElementPtr: {
1584       GEPOperator *CE1GEP = cast<GEPOperator>(CE1);
1585       // Ok, since this is a getelementptr, we know that the constant has a
1586       // pointer type.  Check the various cases.
1587       if (isa<ConstantPointerNull>(V2)) {
1588         // If we are comparing a GEP to a null pointer, check to see if the base
1589         // of the GEP equals the null pointer.
1590         if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1591           // If its not weak linkage, the GVal must have a non-zero address
1592           // so the result is greater-than
1593           if (!GV->hasExternalWeakLinkage())
1594             return ICmpInst::ICMP_UGT;
1595         } else if (isa<ConstantPointerNull>(CE1Op0)) {
1596           // If we are indexing from a null pointer, check to see if we have any
1597           // non-zero indices.
1598           for (unsigned i = 1, e = CE1->getNumOperands(); i != e; ++i)
1599             if (!CE1->getOperand(i)->isNullValue())
1600               // Offsetting from null, must not be equal.
1601               return ICmpInst::ICMP_UGT;
1602           // Only zero indexes from null, must still be zero.
1603           return ICmpInst::ICMP_EQ;
1604         }
1605         // Otherwise, we can't really say if the first operand is null or not.
1606       } else if (const GlobalValue *GV2 = dyn_cast<GlobalValue>(V2)) {
1607         if (isa<ConstantPointerNull>(CE1Op0)) {
1608           // If its not weak linkage, the GVal must have a non-zero address
1609           // so the result is less-than
1610           if (!GV2->hasExternalWeakLinkage())
1611             return ICmpInst::ICMP_ULT;
1612         } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(CE1Op0)) {
1613           if (GV == GV2) {
1614             // If this is a getelementptr of the same global, then it must be
1615             // different.  Because the types must match, the getelementptr could
1616             // only have at most one index, and because we fold getelementptr's
1617             // with a single zero index, it must be nonzero.
1618             assert(CE1->getNumOperands() == 2 &&
1619                    !CE1->getOperand(1)->isNullValue() &&
1620                    "Surprising getelementptr!");
1621             return ICmpInst::ICMP_UGT;
1622           } else {
1623             if (CE1GEP->hasAllZeroIndices())
1624               return areGlobalsPotentiallyEqual(GV, GV2);
1625             return ICmpInst::BAD_ICMP_PREDICATE;
1626           }
1627         }
1628       } else {
1629         ConstantExpr *CE2 = cast<ConstantExpr>(V2);
1630         Constant *CE2Op0 = CE2->getOperand(0);
1631 
1632         // There are MANY other foldings that we could perform here.  They will
1633         // probably be added on demand, as they seem needed.
1634         switch (CE2->getOpcode()) {
1635         default: break;
1636         case Instruction::GetElementPtr:
1637           // By far the most common case to handle is when the base pointers are
1638           // obviously to the same global.
1639           if (isa<GlobalValue>(CE1Op0) && isa<GlobalValue>(CE2Op0)) {
1640             // Don't know relative ordering, but check for inequality.
1641             if (CE1Op0 != CE2Op0) {
1642               GEPOperator *CE2GEP = cast<GEPOperator>(CE2);
1643               if (CE1GEP->hasAllZeroIndices() && CE2GEP->hasAllZeroIndices())
1644                 return areGlobalsPotentiallyEqual(cast<GlobalValue>(CE1Op0),
1645                                                   cast<GlobalValue>(CE2Op0));
1646               return ICmpInst::BAD_ICMP_PREDICATE;
1647             }
1648             // Ok, we know that both getelementptr instructions are based on the
1649             // same global.  From this, we can precisely determine the relative
1650             // ordering of the resultant pointers.
1651             unsigned i = 1;
1652 
1653             // The logic below assumes that the result of the comparison
1654             // can be determined by finding the first index that differs.
1655             // This doesn't work if there is over-indexing in any
1656             // subsequent indices, so check for that case first.
1657             if (!CE1->isGEPWithNoNotionalOverIndexing() ||
1658                 !CE2->isGEPWithNoNotionalOverIndexing())
1659                return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1660 
1661             // Compare all of the operands the GEP's have in common.
1662             gep_type_iterator GTI = gep_type_begin(CE1);
1663             for (;i != CE1->getNumOperands() && i != CE2->getNumOperands();
1664                  ++i, ++GTI)
1665               switch (IdxCompare(CE1->getOperand(i),
1666                                  CE2->getOperand(i), GTI.getIndexedType())) {
1667               case -1: return isSigned ? ICmpInst::ICMP_SLT:ICmpInst::ICMP_ULT;
1668               case 1:  return isSigned ? ICmpInst::ICMP_SGT:ICmpInst::ICMP_UGT;
1669               case -2: return ICmpInst::BAD_ICMP_PREDICATE;
1670               }
1671 
1672             // Ok, we ran out of things they have in common.  If any leftovers
1673             // are non-zero then we have a difference, otherwise we are equal.
1674             for (; i < CE1->getNumOperands(); ++i)
1675               if (!CE1->getOperand(i)->isNullValue()) {
1676                 if (isa<ConstantInt>(CE1->getOperand(i)))
1677                   return isSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1678                 else
1679                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1680               }
1681 
1682             for (; i < CE2->getNumOperands(); ++i)
1683               if (!CE2->getOperand(i)->isNullValue()) {
1684                 if (isa<ConstantInt>(CE2->getOperand(i)))
1685                   return isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1686                 else
1687                   return ICmpInst::BAD_ICMP_PREDICATE; // Might be equal.
1688               }
1689             return ICmpInst::ICMP_EQ;
1690           }
1691         }
1692       }
1693       break;
1694     }
1695     default:
1696       break;
1697     }
1698   }
1699 
1700   return ICmpInst::BAD_ICMP_PREDICATE;
1701 }
1702 
ConstantFoldCompareInstruction(unsigned short pred,Constant * C1,Constant * C2)1703 Constant *llvm::ConstantFoldCompareInstruction(unsigned short pred,
1704                                                Constant *C1, Constant *C2) {
1705   Type *ResultTy;
1706   if (VectorType *VT = dyn_cast<VectorType>(C1->getType()))
1707     ResultTy = VectorType::get(Type::getInt1Ty(C1->getContext()),
1708                                VT->getElementCount());
1709   else
1710     ResultTy = Type::getInt1Ty(C1->getContext());
1711 
1712   // Fold FCMP_FALSE/FCMP_TRUE unconditionally.
1713   if (pred == FCmpInst::FCMP_FALSE)
1714     return Constant::getNullValue(ResultTy);
1715 
1716   if (pred == FCmpInst::FCMP_TRUE)
1717     return Constant::getAllOnesValue(ResultTy);
1718 
1719   // Handle some degenerate cases first
1720   if (isa<PoisonValue>(C1) || isa<PoisonValue>(C2))
1721     return PoisonValue::get(ResultTy);
1722 
1723   if (isa<UndefValue>(C1) || isa<UndefValue>(C2)) {
1724     CmpInst::Predicate Predicate = CmpInst::Predicate(pred);
1725     bool isIntegerPredicate = ICmpInst::isIntPredicate(Predicate);
1726     // For EQ and NE, we can always pick a value for the undef to make the
1727     // predicate pass or fail, so we can return undef.
1728     // Also, if both operands are undef, we can return undef for int comparison.
1729     if (ICmpInst::isEquality(Predicate) || (isIntegerPredicate && C1 == C2))
1730       return UndefValue::get(ResultTy);
1731 
1732     // Otherwise, for integer compare, pick the same value as the non-undef
1733     // operand, and fold it to true or false.
1734     if (isIntegerPredicate)
1735       return ConstantInt::get(ResultTy, CmpInst::isTrueWhenEqual(Predicate));
1736 
1737     // Choosing NaN for the undef will always make unordered comparison succeed
1738     // and ordered comparison fails.
1739     return ConstantInt::get(ResultTy, CmpInst::isUnordered(Predicate));
1740   }
1741 
1742   // icmp eq/ne(null,GV) -> false/true
1743   if (C1->isNullValue()) {
1744     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C2))
1745       // Don't try to evaluate aliases.  External weak GV can be null.
1746       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1747           !NullPointerIsDefined(nullptr /* F */,
1748                                 GV->getType()->getAddressSpace())) {
1749         if (pred == ICmpInst::ICMP_EQ)
1750           return ConstantInt::getFalse(C1->getContext());
1751         else if (pred == ICmpInst::ICMP_NE)
1752           return ConstantInt::getTrue(C1->getContext());
1753       }
1754   // icmp eq/ne(GV,null) -> false/true
1755   } else if (C2->isNullValue()) {
1756     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C1)) {
1757       // Don't try to evaluate aliases.  External weak GV can be null.
1758       if (!isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage() &&
1759           !NullPointerIsDefined(nullptr /* F */,
1760                                 GV->getType()->getAddressSpace())) {
1761         if (pred == ICmpInst::ICMP_EQ)
1762           return ConstantInt::getFalse(C1->getContext());
1763         else if (pred == ICmpInst::ICMP_NE)
1764           return ConstantInt::getTrue(C1->getContext());
1765       }
1766     }
1767 
1768     // The caller is expected to commute the operands if the constant expression
1769     // is C2.
1770     // C1 >= 0 --> true
1771     if (pred == ICmpInst::ICMP_UGE)
1772       return Constant::getAllOnesValue(ResultTy);
1773     // C1 < 0 --> false
1774     if (pred == ICmpInst::ICMP_ULT)
1775       return Constant::getNullValue(ResultTy);
1776   }
1777 
1778   // If the comparison is a comparison between two i1's, simplify it.
1779   if (C1->getType()->isIntegerTy(1)) {
1780     switch(pred) {
1781     case ICmpInst::ICMP_EQ:
1782       if (isa<ConstantInt>(C2))
1783         return ConstantExpr::getXor(C1, ConstantExpr::getNot(C2));
1784       return ConstantExpr::getXor(ConstantExpr::getNot(C1), C2);
1785     case ICmpInst::ICMP_NE:
1786       return ConstantExpr::getXor(C1, C2);
1787     default:
1788       break;
1789     }
1790   }
1791 
1792   if (isa<ConstantInt>(C1) && isa<ConstantInt>(C2)) {
1793     const APInt &V1 = cast<ConstantInt>(C1)->getValue();
1794     const APInt &V2 = cast<ConstantInt>(C2)->getValue();
1795     switch (pred) {
1796     default: llvm_unreachable("Invalid ICmp Predicate");
1797     case ICmpInst::ICMP_EQ:  return ConstantInt::get(ResultTy, V1 == V2);
1798     case ICmpInst::ICMP_NE:  return ConstantInt::get(ResultTy, V1 != V2);
1799     case ICmpInst::ICMP_SLT: return ConstantInt::get(ResultTy, V1.slt(V2));
1800     case ICmpInst::ICMP_SGT: return ConstantInt::get(ResultTy, V1.sgt(V2));
1801     case ICmpInst::ICMP_SLE: return ConstantInt::get(ResultTy, V1.sle(V2));
1802     case ICmpInst::ICMP_SGE: return ConstantInt::get(ResultTy, V1.sge(V2));
1803     case ICmpInst::ICMP_ULT: return ConstantInt::get(ResultTy, V1.ult(V2));
1804     case ICmpInst::ICMP_UGT: return ConstantInt::get(ResultTy, V1.ugt(V2));
1805     case ICmpInst::ICMP_ULE: return ConstantInt::get(ResultTy, V1.ule(V2));
1806     case ICmpInst::ICMP_UGE: return ConstantInt::get(ResultTy, V1.uge(V2));
1807     }
1808   } else if (isa<ConstantFP>(C1) && isa<ConstantFP>(C2)) {
1809     const APFloat &C1V = cast<ConstantFP>(C1)->getValueAPF();
1810     const APFloat &C2V = cast<ConstantFP>(C2)->getValueAPF();
1811     APFloat::cmpResult R = C1V.compare(C2V);
1812     switch (pred) {
1813     default: llvm_unreachable("Invalid FCmp Predicate");
1814     case FCmpInst::FCMP_FALSE: return Constant::getNullValue(ResultTy);
1815     case FCmpInst::FCMP_TRUE:  return Constant::getAllOnesValue(ResultTy);
1816     case FCmpInst::FCMP_UNO:
1817       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered);
1818     case FCmpInst::FCMP_ORD:
1819       return ConstantInt::get(ResultTy, R!=APFloat::cmpUnordered);
1820     case FCmpInst::FCMP_UEQ:
1821       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1822                                         R==APFloat::cmpEqual);
1823     case FCmpInst::FCMP_OEQ:
1824       return ConstantInt::get(ResultTy, R==APFloat::cmpEqual);
1825     case FCmpInst::FCMP_UNE:
1826       return ConstantInt::get(ResultTy, R!=APFloat::cmpEqual);
1827     case FCmpInst::FCMP_ONE:
1828       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1829                                         R==APFloat::cmpGreaterThan);
1830     case FCmpInst::FCMP_ULT:
1831       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1832                                         R==APFloat::cmpLessThan);
1833     case FCmpInst::FCMP_OLT:
1834       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan);
1835     case FCmpInst::FCMP_UGT:
1836       return ConstantInt::get(ResultTy, R==APFloat::cmpUnordered ||
1837                                         R==APFloat::cmpGreaterThan);
1838     case FCmpInst::FCMP_OGT:
1839       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan);
1840     case FCmpInst::FCMP_ULE:
1841       return ConstantInt::get(ResultTy, R!=APFloat::cmpGreaterThan);
1842     case FCmpInst::FCMP_OLE:
1843       return ConstantInt::get(ResultTy, R==APFloat::cmpLessThan ||
1844                                         R==APFloat::cmpEqual);
1845     case FCmpInst::FCMP_UGE:
1846       return ConstantInt::get(ResultTy, R!=APFloat::cmpLessThan);
1847     case FCmpInst::FCMP_OGE:
1848       return ConstantInt::get(ResultTy, R==APFloat::cmpGreaterThan ||
1849                                         R==APFloat::cmpEqual);
1850     }
1851   } else if (auto *C1VTy = dyn_cast<VectorType>(C1->getType())) {
1852 
1853     // Fast path for splatted constants.
1854     if (Constant *C1Splat = C1->getSplatValue())
1855       if (Constant *C2Splat = C2->getSplatValue())
1856         return ConstantVector::getSplat(
1857             C1VTy->getElementCount(),
1858             ConstantExpr::getCompare(pred, C1Splat, C2Splat));
1859 
1860     // Do not iterate on scalable vector. The number of elements is unknown at
1861     // compile-time.
1862     if (isa<ScalableVectorType>(C1VTy))
1863       return nullptr;
1864 
1865     // If we can constant fold the comparison of each element, constant fold
1866     // the whole vector comparison.
1867     SmallVector<Constant*, 4> ResElts;
1868     Type *Ty = IntegerType::get(C1->getContext(), 32);
1869     // Compare the elements, producing an i1 result or constant expr.
1870     for (unsigned I = 0, E = C1VTy->getElementCount().getKnownMinValue();
1871          I != E; ++I) {
1872       Constant *C1E =
1873           ConstantExpr::getExtractElement(C1, ConstantInt::get(Ty, I));
1874       Constant *C2E =
1875           ConstantExpr::getExtractElement(C2, ConstantInt::get(Ty, I));
1876 
1877       ResElts.push_back(ConstantExpr::getCompare(pred, C1E, C2E));
1878     }
1879 
1880     return ConstantVector::get(ResElts);
1881   }
1882 
1883   if (C1->getType()->isFloatingPointTy() &&
1884       // Only call evaluateFCmpRelation if we have a constant expr to avoid
1885       // infinite recursive loop
1886       (isa<ConstantExpr>(C1) || isa<ConstantExpr>(C2))) {
1887     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1888     switch (evaluateFCmpRelation(C1, C2)) {
1889     default: llvm_unreachable("Unknown relation!");
1890     case FCmpInst::FCMP_UNO:
1891     case FCmpInst::FCMP_ORD:
1892     case FCmpInst::FCMP_UNE:
1893     case FCmpInst::FCMP_ULT:
1894     case FCmpInst::FCMP_UGT:
1895     case FCmpInst::FCMP_ULE:
1896     case FCmpInst::FCMP_UGE:
1897     case FCmpInst::FCMP_TRUE:
1898     case FCmpInst::FCMP_FALSE:
1899     case FCmpInst::BAD_FCMP_PREDICATE:
1900       break; // Couldn't determine anything about these constants.
1901     case FCmpInst::FCMP_OEQ: // We know that C1 == C2
1902       Result = (pred == FCmpInst::FCMP_UEQ || pred == FCmpInst::FCMP_OEQ ||
1903                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE ||
1904                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1905       break;
1906     case FCmpInst::FCMP_OLT: // We know that C1 < C2
1907       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1908                 pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT ||
1909                 pred == FCmpInst::FCMP_ULE || pred == FCmpInst::FCMP_OLE);
1910       break;
1911     case FCmpInst::FCMP_OGT: // We know that C1 > C2
1912       Result = (pred == FCmpInst::FCMP_UNE || pred == FCmpInst::FCMP_ONE ||
1913                 pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT ||
1914                 pred == FCmpInst::FCMP_UGE || pred == FCmpInst::FCMP_OGE);
1915       break;
1916     case FCmpInst::FCMP_OLE: // We know that C1 <= C2
1917       // We can only partially decide this relation.
1918       if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1919         Result = 0;
1920       else if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1921         Result = 1;
1922       break;
1923     case FCmpInst::FCMP_OGE: // We known that C1 >= C2
1924       // We can only partially decide this relation.
1925       if (pred == FCmpInst::FCMP_ULT || pred == FCmpInst::FCMP_OLT)
1926         Result = 0;
1927       else if (pred == FCmpInst::FCMP_UGT || pred == FCmpInst::FCMP_OGT)
1928         Result = 1;
1929       break;
1930     case FCmpInst::FCMP_ONE: // We know that C1 != C2
1931       // We can only partially decide this relation.
1932       if (pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ)
1933         Result = 0;
1934       else if (pred == FCmpInst::FCMP_ONE || pred == FCmpInst::FCMP_UNE)
1935         Result = 1;
1936       break;
1937     case FCmpInst::FCMP_UEQ: // We know that C1 == C2 || isUnordered(C1, C2).
1938       // We can only partially decide this relation.
1939       if (pred == FCmpInst::FCMP_ONE)
1940         Result = 0;
1941       else if (pred == FCmpInst::FCMP_UEQ)
1942         Result = 1;
1943       break;
1944     }
1945 
1946     // If we evaluated the result, return it now.
1947     if (Result != -1)
1948       return ConstantInt::get(ResultTy, Result);
1949 
1950   } else {
1951     // Evaluate the relation between the two constants, per the predicate.
1952     int Result = -1;  // -1 = unknown, 0 = known false, 1 = known true.
1953     switch (evaluateICmpRelation(C1, C2,
1954                                  CmpInst::isSigned((CmpInst::Predicate)pred))) {
1955     default: llvm_unreachable("Unknown relational!");
1956     case ICmpInst::BAD_ICMP_PREDICATE:
1957       break;  // Couldn't determine anything about these constants.
1958     case ICmpInst::ICMP_EQ:   // We know the constants are equal!
1959       // If we know the constants are equal, we can decide the result of this
1960       // computation precisely.
1961       Result = ICmpInst::isTrueWhenEqual((ICmpInst::Predicate)pred);
1962       break;
1963     case ICmpInst::ICMP_ULT:
1964       switch (pred) {
1965       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_ULE:
1966         Result = 1; break;
1967       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_UGE:
1968         Result = 0; break;
1969       }
1970       break;
1971     case ICmpInst::ICMP_SLT:
1972       switch (pred) {
1973       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SLE:
1974         Result = 1; break;
1975       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SGE:
1976         Result = 0; break;
1977       }
1978       break;
1979     case ICmpInst::ICMP_UGT:
1980       switch (pred) {
1981       case ICmpInst::ICMP_UGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_UGE:
1982         Result = 1; break;
1983       case ICmpInst::ICMP_ULT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_ULE:
1984         Result = 0; break;
1985       }
1986       break;
1987     case ICmpInst::ICMP_SGT:
1988       switch (pred) {
1989       case ICmpInst::ICMP_SGT: case ICmpInst::ICMP_NE: case ICmpInst::ICMP_SGE:
1990         Result = 1; break;
1991       case ICmpInst::ICMP_SLT: case ICmpInst::ICMP_EQ: case ICmpInst::ICMP_SLE:
1992         Result = 0; break;
1993       }
1994       break;
1995     case ICmpInst::ICMP_ULE:
1996       if (pred == ICmpInst::ICMP_UGT) Result = 0;
1997       if (pred == ICmpInst::ICMP_ULT || pred == ICmpInst::ICMP_ULE) Result = 1;
1998       break;
1999     case ICmpInst::ICMP_SLE:
2000       if (pred == ICmpInst::ICMP_SGT) Result = 0;
2001       if (pred == ICmpInst::ICMP_SLT || pred == ICmpInst::ICMP_SLE) Result = 1;
2002       break;
2003     case ICmpInst::ICMP_UGE:
2004       if (pred == ICmpInst::ICMP_ULT) Result = 0;
2005       if (pred == ICmpInst::ICMP_UGT || pred == ICmpInst::ICMP_UGE) Result = 1;
2006       break;
2007     case ICmpInst::ICMP_SGE:
2008       if (pred == ICmpInst::ICMP_SLT) Result = 0;
2009       if (pred == ICmpInst::ICMP_SGT || pred == ICmpInst::ICMP_SGE) Result = 1;
2010       break;
2011     case ICmpInst::ICMP_NE:
2012       if (pred == ICmpInst::ICMP_EQ) Result = 0;
2013       if (pred == ICmpInst::ICMP_NE) Result = 1;
2014       break;
2015     }
2016 
2017     // If we evaluated the result, return it now.
2018     if (Result != -1)
2019       return ConstantInt::get(ResultTy, Result);
2020 
2021     // If the right hand side is a bitcast, try using its inverse to simplify
2022     // it by moving it to the left hand side.  We can't do this if it would turn
2023     // a vector compare into a scalar compare or visa versa, or if it would turn
2024     // the operands into FP values.
2025     if (ConstantExpr *CE2 = dyn_cast<ConstantExpr>(C2)) {
2026       Constant *CE2Op0 = CE2->getOperand(0);
2027       if (CE2->getOpcode() == Instruction::BitCast &&
2028           CE2->getType()->isVectorTy() == CE2Op0->getType()->isVectorTy() &&
2029           !CE2Op0->getType()->isFPOrFPVectorTy()) {
2030         Constant *Inverse = ConstantExpr::getBitCast(C1, CE2Op0->getType());
2031         return ConstantExpr::getICmp(pred, Inverse, CE2Op0);
2032       }
2033     }
2034 
2035     // If the left hand side is an extension, try eliminating it.
2036     if (ConstantExpr *CE1 = dyn_cast<ConstantExpr>(C1)) {
2037       if ((CE1->getOpcode() == Instruction::SExt &&
2038            ICmpInst::isSigned((ICmpInst::Predicate)pred)) ||
2039           (CE1->getOpcode() == Instruction::ZExt &&
2040            !ICmpInst::isSigned((ICmpInst::Predicate)pred))){
2041         Constant *CE1Op0 = CE1->getOperand(0);
2042         Constant *CE1Inverse = ConstantExpr::getTrunc(CE1, CE1Op0->getType());
2043         if (CE1Inverse == CE1Op0) {
2044           // Check whether we can safely truncate the right hand side.
2045           Constant *C2Inverse = ConstantExpr::getTrunc(C2, CE1Op0->getType());
2046           if (ConstantExpr::getCast(CE1->getOpcode(), C2Inverse,
2047                                     C2->getType()) == C2)
2048             return ConstantExpr::getICmp(pred, CE1Inverse, C2Inverse);
2049         }
2050       }
2051     }
2052 
2053     if ((!isa<ConstantExpr>(C1) && isa<ConstantExpr>(C2)) ||
2054         (C1->isNullValue() && !C2->isNullValue())) {
2055       // If C2 is a constant expr and C1 isn't, flip them around and fold the
2056       // other way if possible.
2057       // Also, if C1 is null and C2 isn't, flip them around.
2058       pred = ICmpInst::getSwappedPredicate((ICmpInst::Predicate)pred);
2059       return ConstantExpr::getICmp(pred, C2, C1);
2060     }
2061   }
2062   return nullptr;
2063 }
2064 
2065 /// Test whether the given sequence of *normalized* indices is "inbounds".
2066 template<typename IndexTy>
isInBoundsIndices(ArrayRef<IndexTy> Idxs)2067 static bool isInBoundsIndices(ArrayRef<IndexTy> Idxs) {
2068   // No indices means nothing that could be out of bounds.
2069   if (Idxs.empty()) return true;
2070 
2071   // If the first index is zero, it's in bounds.
2072   if (cast<Constant>(Idxs[0])->isNullValue()) return true;
2073 
2074   // If the first index is one and all the rest are zero, it's in bounds,
2075   // by the one-past-the-end rule.
2076   if (auto *CI = dyn_cast<ConstantInt>(Idxs[0])) {
2077     if (!CI->isOne())
2078       return false;
2079   } else {
2080     auto *CV = cast<ConstantDataVector>(Idxs[0]);
2081     CI = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
2082     if (!CI || !CI->isOne())
2083       return false;
2084   }
2085 
2086   for (unsigned i = 1, e = Idxs.size(); i != e; ++i)
2087     if (!cast<Constant>(Idxs[i])->isNullValue())
2088       return false;
2089   return true;
2090 }
2091 
2092 /// Test whether a given ConstantInt is in-range for a SequentialType.
isIndexInRangeOfArrayType(uint64_t NumElements,const ConstantInt * CI)2093 static bool isIndexInRangeOfArrayType(uint64_t NumElements,
2094                                       const ConstantInt *CI) {
2095   // We cannot bounds check the index if it doesn't fit in an int64_t.
2096   if (CI->getValue().getMinSignedBits() > 64)
2097     return false;
2098 
2099   // A negative index or an index past the end of our sequential type is
2100   // considered out-of-range.
2101   int64_t IndexVal = CI->getSExtValue();
2102   if (IndexVal < 0 || (NumElements > 0 && (uint64_t)IndexVal >= NumElements))
2103     return false;
2104 
2105   // Otherwise, it is in-range.
2106   return true;
2107 }
2108 
2109 // Combine Indices - If the source pointer to this getelementptr instruction
2110 // is a getelementptr instruction, combine the indices of the two
2111 // getelementptr instructions into a single instruction.
foldGEPOfGEP(GEPOperator * GEP,Type * PointeeTy,bool InBounds,ArrayRef<Value * > Idxs)2112 static Constant *foldGEPOfGEP(GEPOperator *GEP, Type *PointeeTy, bool InBounds,
2113                               ArrayRef<Value *> Idxs) {
2114   if (PointeeTy != GEP->getResultElementType())
2115     return nullptr;
2116 
2117   Constant *Idx0 = cast<Constant>(Idxs[0]);
2118   if (Idx0->isNullValue()) {
2119     // Handle the simple case of a zero index.
2120     SmallVector<Value*, 16> NewIndices;
2121     NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
2122     NewIndices.append(GEP->idx_begin(), GEP->idx_end());
2123     NewIndices.append(Idxs.begin() + 1, Idxs.end());
2124     return ConstantExpr::getGetElementPtr(
2125         GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
2126         NewIndices, InBounds && GEP->isInBounds(), GEP->getInRangeIndex());
2127   }
2128 
2129   gep_type_iterator LastI = gep_type_end(GEP);
2130   for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
2131        I != E; ++I)
2132     LastI = I;
2133 
2134   // We cannot combine indices if doing so would take us outside of an
2135   // array or vector.  Doing otherwise could trick us if we evaluated such a
2136   // GEP as part of a load.
2137   //
2138   // e.g. Consider if the original GEP was:
2139   // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2140   //                    i32 0, i32 0, i64 0)
2141   //
2142   // If we then tried to offset it by '8' to get to the third element,
2143   // an i8, we should *not* get:
2144   // i8* getelementptr ({ [2 x i8], i32, i8, [3 x i8] }* @main.c,
2145   //                    i32 0, i32 0, i64 8)
2146   //
2147   // This GEP tries to index array element '8  which runs out-of-bounds.
2148   // Subsequent evaluation would get confused and produce erroneous results.
2149   //
2150   // The following prohibits such a GEP from being formed by checking to see
2151   // if the index is in-range with respect to an array.
2152   if (!LastI.isSequential())
2153     return nullptr;
2154   ConstantInt *CI = dyn_cast<ConstantInt>(Idx0);
2155   if (!CI)
2156     return nullptr;
2157   if (LastI.isBoundedSequential() &&
2158       !isIndexInRangeOfArrayType(LastI.getSequentialNumElements(), CI))
2159     return nullptr;
2160 
2161   // TODO: This code may be extended to handle vectors as well.
2162   auto *LastIdx = cast<Constant>(GEP->getOperand(GEP->getNumOperands()-1));
2163   Type *LastIdxTy = LastIdx->getType();
2164   if (LastIdxTy->isVectorTy())
2165     return nullptr;
2166 
2167   SmallVector<Value*, 16> NewIndices;
2168   NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
2169   NewIndices.append(GEP->idx_begin(), GEP->idx_end() - 1);
2170 
2171   // Add the last index of the source with the first index of the new GEP.
2172   // Make sure to handle the case when they are actually different types.
2173   if (LastIdxTy != Idx0->getType()) {
2174     unsigned CommonExtendedWidth =
2175         std::max(LastIdxTy->getIntegerBitWidth(),
2176                  Idx0->getType()->getIntegerBitWidth());
2177     CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2178 
2179     Type *CommonTy =
2180         Type::getIntNTy(LastIdxTy->getContext(), CommonExtendedWidth);
2181     Idx0 = ConstantExpr::getSExtOrBitCast(Idx0, CommonTy);
2182     LastIdx = ConstantExpr::getSExtOrBitCast(LastIdx, CommonTy);
2183   }
2184 
2185   NewIndices.push_back(ConstantExpr::get(Instruction::Add, Idx0, LastIdx));
2186   NewIndices.append(Idxs.begin() + 1, Idxs.end());
2187 
2188   // The combined GEP normally inherits its index inrange attribute from
2189   // the inner GEP, but if the inner GEP's last index was adjusted by the
2190   // outer GEP, any inbounds attribute on that index is invalidated.
2191   Optional<unsigned> IRIndex = GEP->getInRangeIndex();
2192   if (IRIndex && *IRIndex == GEP->getNumIndices() - 1)
2193     IRIndex = None;
2194 
2195   return ConstantExpr::getGetElementPtr(
2196       GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
2197       NewIndices, InBounds && GEP->isInBounds(), IRIndex);
2198 }
2199 
ConstantFoldGetElementPtr(Type * PointeeTy,Constant * C,bool InBounds,Optional<unsigned> InRangeIndex,ArrayRef<Value * > Idxs)2200 Constant *llvm::ConstantFoldGetElementPtr(Type *PointeeTy, Constant *C,
2201                                           bool InBounds,
2202                                           Optional<unsigned> InRangeIndex,
2203                                           ArrayRef<Value *> Idxs) {
2204   if (Idxs.empty()) return C;
2205 
2206   Type *GEPTy = GetElementPtrInst::getGEPReturnType(
2207       PointeeTy, C, makeArrayRef((Value *const *)Idxs.data(), Idxs.size()));
2208 
2209   if (isa<PoisonValue>(C))
2210     return PoisonValue::get(GEPTy);
2211 
2212   if (isa<UndefValue>(C))
2213     // If inbounds, we can choose an out-of-bounds pointer as a base pointer.
2214     return InBounds ? PoisonValue::get(GEPTy) : UndefValue::get(GEPTy);
2215 
2216   Constant *Idx0 = cast<Constant>(Idxs[0]);
2217   if (Idxs.size() == 1 && (Idx0->isNullValue() || isa<UndefValue>(Idx0)))
2218     return GEPTy->isVectorTy() && !C->getType()->isVectorTy()
2219                ? ConstantVector::getSplat(
2220                      cast<VectorType>(GEPTy)->getElementCount(), C)
2221                : C;
2222 
2223   if (C->isNullValue()) {
2224     bool isNull = true;
2225     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2226       if (!isa<UndefValue>(Idxs[i]) &&
2227           !cast<Constant>(Idxs[i])->isNullValue()) {
2228         isNull = false;
2229         break;
2230       }
2231     if (isNull) {
2232       PointerType *PtrTy = cast<PointerType>(C->getType()->getScalarType());
2233       Type *Ty = GetElementPtrInst::getIndexedType(PointeeTy, Idxs);
2234 
2235       assert(Ty && "Invalid indices for GEP!");
2236       Type *OrigGEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2237       Type *GEPTy = PointerType::get(Ty, PtrTy->getAddressSpace());
2238       if (VectorType *VT = dyn_cast<VectorType>(C->getType()))
2239         GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2240 
2241       // The GEP returns a vector of pointers when one of more of
2242       // its arguments is a vector.
2243       for (unsigned i = 0, e = Idxs.size(); i != e; ++i) {
2244         if (auto *VT = dyn_cast<VectorType>(Idxs[i]->getType())) {
2245           assert((!isa<VectorType>(GEPTy) || isa<ScalableVectorType>(GEPTy) ==
2246                                                  isa<ScalableVectorType>(VT)) &&
2247                  "Mismatched GEPTy vector types");
2248           GEPTy = VectorType::get(OrigGEPTy, VT->getElementCount());
2249           break;
2250         }
2251       }
2252 
2253       return Constant::getNullValue(GEPTy);
2254     }
2255   }
2256 
2257   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
2258     if (auto *GEP = dyn_cast<GEPOperator>(CE))
2259       if (Constant *C = foldGEPOfGEP(GEP, PointeeTy, InBounds, Idxs))
2260         return C;
2261 
2262     // Attempt to fold casts to the same type away.  For example, folding:
2263     //
2264     //   i32* getelementptr ([2 x i32]* bitcast ([3 x i32]* %X to [2 x i32]*),
2265     //                       i64 0, i64 0)
2266     // into:
2267     //
2268     //   i32* getelementptr ([3 x i32]* %X, i64 0, i64 0)
2269     //
2270     // Don't fold if the cast is changing address spaces.
2271     if (CE->isCast() && Idxs.size() > 1 && Idx0->isNullValue()) {
2272       PointerType *SrcPtrTy =
2273         dyn_cast<PointerType>(CE->getOperand(0)->getType());
2274       PointerType *DstPtrTy = dyn_cast<PointerType>(CE->getType());
2275       if (SrcPtrTy && DstPtrTy) {
2276         ArrayType *SrcArrayTy =
2277           dyn_cast<ArrayType>(SrcPtrTy->getElementType());
2278         ArrayType *DstArrayTy =
2279           dyn_cast<ArrayType>(DstPtrTy->getElementType());
2280         if (SrcArrayTy && DstArrayTy
2281             && SrcArrayTy->getElementType() == DstArrayTy->getElementType()
2282             && SrcPtrTy->getAddressSpace() == DstPtrTy->getAddressSpace())
2283           return ConstantExpr::getGetElementPtr(SrcArrayTy,
2284                                                 (Constant *)CE->getOperand(0),
2285                                                 Idxs, InBounds, InRangeIndex);
2286       }
2287     }
2288   }
2289 
2290   // Check to see if any array indices are not within the corresponding
2291   // notional array or vector bounds. If so, try to determine if they can be
2292   // factored out into preceding dimensions.
2293   SmallVector<Constant *, 8> NewIdxs;
2294   Type *Ty = PointeeTy;
2295   Type *Prev = C->getType();
2296   auto GEPIter = gep_type_begin(PointeeTy, Idxs);
2297   bool Unknown =
2298       !isa<ConstantInt>(Idxs[0]) && !isa<ConstantDataVector>(Idxs[0]);
2299   for (unsigned i = 1, e = Idxs.size(); i != e;
2300        Prev = Ty, Ty = (++GEPIter).getIndexedType(), ++i) {
2301     if (!isa<ConstantInt>(Idxs[i]) && !isa<ConstantDataVector>(Idxs[i])) {
2302       // We don't know if it's in range or not.
2303       Unknown = true;
2304       continue;
2305     }
2306     if (!isa<ConstantInt>(Idxs[i - 1]) && !isa<ConstantDataVector>(Idxs[i - 1]))
2307       // Skip if the type of the previous index is not supported.
2308       continue;
2309     if (InRangeIndex && i == *InRangeIndex + 1) {
2310       // If an index is marked inrange, we cannot apply this canonicalization to
2311       // the following index, as that will cause the inrange index to point to
2312       // the wrong element.
2313       continue;
2314     }
2315     if (isa<StructType>(Ty)) {
2316       // The verify makes sure that GEPs into a struct are in range.
2317       continue;
2318     }
2319     if (isa<VectorType>(Ty)) {
2320       // There can be awkward padding in after a non-power of two vector.
2321       Unknown = true;
2322       continue;
2323     }
2324     auto *STy = cast<ArrayType>(Ty);
2325     if (ConstantInt *CI = dyn_cast<ConstantInt>(Idxs[i])) {
2326       if (isIndexInRangeOfArrayType(STy->getNumElements(), CI))
2327         // It's in range, skip to the next index.
2328         continue;
2329       if (CI->isNegative()) {
2330         // It's out of range and negative, don't try to factor it.
2331         Unknown = true;
2332         continue;
2333       }
2334     } else {
2335       auto *CV = cast<ConstantDataVector>(Idxs[i]);
2336       bool InRange = true;
2337       for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
2338         auto *CI = cast<ConstantInt>(CV->getElementAsConstant(I));
2339         InRange &= isIndexInRangeOfArrayType(STy->getNumElements(), CI);
2340         if (CI->isNegative()) {
2341           Unknown = true;
2342           break;
2343         }
2344       }
2345       if (InRange || Unknown)
2346         // It's in range, skip to the next index.
2347         // It's out of range and negative, don't try to factor it.
2348         continue;
2349     }
2350     if (isa<StructType>(Prev)) {
2351       // It's out of range, but the prior dimension is a struct
2352       // so we can't do anything about it.
2353       Unknown = true;
2354       continue;
2355     }
2356     // It's out of range, but we can factor it into the prior
2357     // dimension.
2358     NewIdxs.resize(Idxs.size());
2359     // Determine the number of elements in our sequential type.
2360     uint64_t NumElements = STy->getArrayNumElements();
2361 
2362     // Expand the current index or the previous index to a vector from a scalar
2363     // if necessary.
2364     Constant *CurrIdx = cast<Constant>(Idxs[i]);
2365     auto *PrevIdx =
2366         NewIdxs[i - 1] ? NewIdxs[i - 1] : cast<Constant>(Idxs[i - 1]);
2367     bool IsCurrIdxVector = CurrIdx->getType()->isVectorTy();
2368     bool IsPrevIdxVector = PrevIdx->getType()->isVectorTy();
2369     bool UseVector = IsCurrIdxVector || IsPrevIdxVector;
2370 
2371     if (!IsCurrIdxVector && IsPrevIdxVector)
2372       CurrIdx = ConstantDataVector::getSplat(
2373           cast<FixedVectorType>(PrevIdx->getType())->getNumElements(), CurrIdx);
2374 
2375     if (!IsPrevIdxVector && IsCurrIdxVector)
2376       PrevIdx = ConstantDataVector::getSplat(
2377           cast<FixedVectorType>(CurrIdx->getType())->getNumElements(), PrevIdx);
2378 
2379     Constant *Factor =
2380         ConstantInt::get(CurrIdx->getType()->getScalarType(), NumElements);
2381     if (UseVector)
2382       Factor = ConstantDataVector::getSplat(
2383           IsPrevIdxVector
2384               ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2385               : cast<FixedVectorType>(CurrIdx->getType())->getNumElements(),
2386           Factor);
2387 
2388     NewIdxs[i] = ConstantExpr::getSRem(CurrIdx, Factor);
2389 
2390     Constant *Div = ConstantExpr::getSDiv(CurrIdx, Factor);
2391 
2392     unsigned CommonExtendedWidth =
2393         std::max(PrevIdx->getType()->getScalarSizeInBits(),
2394                  Div->getType()->getScalarSizeInBits());
2395     CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);
2396 
2397     // Before adding, extend both operands to i64 to avoid
2398     // overflow trouble.
2399     Type *ExtendedTy = Type::getIntNTy(Div->getContext(), CommonExtendedWidth);
2400     if (UseVector)
2401       ExtendedTy = FixedVectorType::get(
2402           ExtendedTy,
2403           IsPrevIdxVector
2404               ? cast<FixedVectorType>(PrevIdx->getType())->getNumElements()
2405               : cast<FixedVectorType>(CurrIdx->getType())->getNumElements());
2406 
2407     if (!PrevIdx->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2408       PrevIdx = ConstantExpr::getSExt(PrevIdx, ExtendedTy);
2409 
2410     if (!Div->getType()->isIntOrIntVectorTy(CommonExtendedWidth))
2411       Div = ConstantExpr::getSExt(Div, ExtendedTy);
2412 
2413     NewIdxs[i - 1] = ConstantExpr::getAdd(PrevIdx, Div);
2414   }
2415 
2416   // If we did any factoring, start over with the adjusted indices.
2417   if (!NewIdxs.empty()) {
2418     for (unsigned i = 0, e = Idxs.size(); i != e; ++i)
2419       if (!NewIdxs[i]) NewIdxs[i] = cast<Constant>(Idxs[i]);
2420     return ConstantExpr::getGetElementPtr(PointeeTy, C, NewIdxs, InBounds,
2421                                           InRangeIndex);
2422   }
2423 
2424   // If all indices are known integers and normalized, we can do a simple
2425   // check for the "inbounds" property.
2426   if (!Unknown && !InBounds)
2427     if (auto *GV = dyn_cast<GlobalVariable>(C))
2428       if (!GV->hasExternalWeakLinkage() && isInBoundsIndices(Idxs))
2429         return ConstantExpr::getGetElementPtr(PointeeTy, C, Idxs,
2430                                               /*InBounds=*/true, InRangeIndex);
2431 
2432   return nullptr;
2433 }
2434