llvm/lib/FuzzMutate/Operations.cpp

   1 //===-- Operations.cpp ----------------------------------------------------===//
   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 #include "llvm/FuzzMutate/Operations.h"
  10 #include "llvm/IR/BasicBlock.h"
  11 #include "llvm/IR/Constants.h"
  12 #include "llvm/IR/Function.h"
  13 #include "llvm/IR/Instructions.h"
  14
  15 using namespace llvm;
  16 using namespace fuzzerop;
  17
  18 void llvm::describeFuzzerIntOps(std::vector<fuzzerop::OpDescriptor> &Ops) {
  19   Ops.push_back(binOpDescriptor(1, Instruction::Add));
  20   Ops.push_back(binOpDescriptor(1, Instruction::Sub));
  21   Ops.push_back(binOpDescriptor(1, Instruction::Mul));
  22   Ops.push_back(binOpDescriptor(1, Instruction::SDiv));
  23   Ops.push_back(binOpDescriptor(1, Instruction::UDiv));
  24   Ops.push_back(binOpDescriptor(1, Instruction::SRem));
  25   Ops.push_back(binOpDescriptor(1, Instruction::URem));
  26   Ops.push_back(binOpDescriptor(1, Instruction::Shl));
  27   Ops.push_back(binOpDescriptor(1, Instruction::LShr));
  28   Ops.push_back(binOpDescriptor(1, Instruction::AShr));
  29   Ops.push_back(binOpDescriptor(1, Instruction::And));
  30   Ops.push_back(binOpDescriptor(1, Instruction::Or));
  31   Ops.push_back(binOpDescriptor(1, Instruction::Xor));
  32
  33   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_EQ));
  34   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_NE));
  35   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_UGT));
  36   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_UGE));
  37   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_ULT));
  38   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_ULE));
  39   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_SGT));
  40   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_SGE));
  41   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_SLT));
  42   Ops.push_back(cmpOpDescriptor(1, Instruction::ICmp, CmpInst::ICMP_SLE));
  43 }
  44
  45 void llvm::describeFuzzerFloatOps(std::vector<fuzzerop::OpDescriptor> &Ops) {
  46   Ops.push_back(binOpDescriptor(1, Instruction::FAdd));
  47   Ops.push_back(binOpDescriptor(1, Instruction::FSub));
  48   Ops.push_back(binOpDescriptor(1, Instruction::FMul));
  49   Ops.push_back(binOpDescriptor(1, Instruction::FDiv));
  50   Ops.push_back(binOpDescriptor(1, Instruction::FRem));
  51
  52   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_FALSE));
  53   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_OEQ));
  54   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_OGT));
  55   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_OGE));
  56   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_OLT));
  57   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_OLE));
  58   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_ONE));
  59   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_ORD));
  60   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_UNO));
  61   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_UEQ));
  62   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_UGT));
  63   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_UGE));
  64   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_ULT));
  65   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_ULE));
  66   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_UNE));
  67   Ops.push_back(cmpOpDescriptor(1, Instruction::FCmp, CmpInst::FCMP_TRUE));
  68 }
  69
  70 void llvm::describeFuzzerUnaryOperations(
  71     std::vector<fuzzerop::OpDescriptor> &Ops) {
  72   Ops.push_back(fnegDescriptor(1));
  73 }
  74
  75 void llvm::describeFuzzerControlFlowOps(
  76     std::vector<fuzzerop::OpDescriptor> &Ops) {
  77   Ops.push_back(splitBlockDescriptor(1));
  78 }
  79
  80 void llvm::describeFuzzerOtherOps(std::vector<fuzzerop::OpDescriptor> &Ops) {
  81   Ops.push_back(selectDescriptor(1));
  82 }
  83
  84 void llvm::describeFuzzerPointerOps(std::vector<fuzzerop::OpDescriptor> &Ops) {
  85   Ops.push_back(gepDescriptor(1));
  86 }
  87
  88 void llvm::describeFuzzerAggregateOps(
  89     std::vector<fuzzerop::OpDescriptor> &Ops) {
  90   Ops.push_back(extractValueDescriptor(1));
  91   Ops.push_back(insertValueDescriptor(1));
  92 }
  93
  94 void llvm::describeFuzzerVectorOps(std::vector<fuzzerop::OpDescriptor> &Ops) {
  95   Ops.push_back(extractElementDescriptor(1));
  96   Ops.push_back(insertElementDescriptor(1));
  97   Ops.push_back(shuffleVectorDescriptor(1));
  98 }
  99
 100 OpDescriptor llvm::fuzzerop::selectDescriptor(unsigned Weight) {
 101   auto buildOp = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 102     return SelectInst::Create(Srcs[0], Srcs[1], Srcs[2], "S", Inst);
 103   };
 104   return {Weight,
 105           {boolOrVecBoolType(), matchFirstLengthWAnyType(), matchSecondType()},
 106           buildOp};
 107 }
 108
 109 OpDescriptor llvm::fuzzerop::fnegDescriptor(unsigned Weight) {
 110   auto buildOp = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 111     return UnaryOperator::Create(Instruction::FNeg, Srcs[0], "F", Inst);
 112   };
 113   return {Weight, {anyFloatOrVecFloatType()}, buildOp};
 114 }
 115
 116 OpDescriptor llvm::fuzzerop::binOpDescriptor(unsigned Weight,
 117                                              Instruction::BinaryOps Op) {
 118   auto buildOp = [Op](ArrayRef<Value *> Srcs, Instruction *Inst) {
 119     return BinaryOperator::Create(Op, Srcs[0], Srcs[1], "B", Inst);
 120   };
 121   switch (Op) {
 122   case Instruction::Add:
 123   case Instruction::Sub:
 124   case Instruction::Mul:
 125   case Instruction::SDiv:
 126   case Instruction::UDiv:
 127   case Instruction::SRem:
 128   case Instruction::URem:
 129   case Instruction::Shl:
 130   case Instruction::LShr:
 131   case Instruction::AShr:
 132   case Instruction::And:
 133   case Instruction::Or:
 134   case Instruction::Xor:
 135     return {Weight, {anyIntOrVecIntType(), matchFirstType()}, buildOp};
 136   case Instruction::FAdd:
 137   case Instruction::FSub:
 138   case Instruction::FMul:
 139   case Instruction::FDiv:
 140   case Instruction::FRem:
 141     return {Weight, {anyFloatOrVecFloatType(), matchFirstType()}, buildOp};
 142   case Instruction::BinaryOpsEnd:
 143     llvm_unreachable("Value out of range of enum");
 144   }
 145   llvm_unreachable("Covered switch");
 146 }
 147
 148 OpDescriptor llvm::fuzzerop::cmpOpDescriptor(unsigned Weight,
 149                                              Instruction::OtherOps CmpOp,
 150                                              CmpInst::Predicate Pred) {
 151   auto buildOp = [CmpOp, Pred](ArrayRef<Value *> Srcs, Instruction *Inst) {
 152     return CmpInst::Create(CmpOp, Pred, Srcs[0], Srcs[1], "C", Inst);
 153   };
 154
 155   switch (CmpOp) {
 156   case Instruction::ICmp:
 157     return {Weight, {anyIntOrVecIntType(), matchFirstType()}, buildOp};
 158   case Instruction::FCmp:
 159     return {Weight, {anyFloatOrVecFloatType(), matchFirstType()}, buildOp};
 160   default:
 161     llvm_unreachable("CmpOp must be ICmp or FCmp");
 162   }
 163 }
 164
 165 OpDescriptor llvm::fuzzerop::splitBlockDescriptor(unsigned Weight) {
 166   auto buildSplitBlock = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 167     BasicBlock *Block = Inst->getParent();
 168     BasicBlock *Next = Block->splitBasicBlock(Inst, "BB");
 169
 170     // If it was an exception handling block, we are done.
 171     if (Block->isEHPad())
 172       return nullptr;
 173
 174     // Loop back on this block by replacing the unconditional forward branch
 175     // with a conditional with a backedge.
 176     if (Block != &Block->getParent()->getEntryBlock()) {
 177       BranchInst::Create(Block, Next, Srcs[0], Block->getTerminator());
 178       Block->getTerminator()->eraseFromParent();
 179
 180       // We need values for each phi in the block. Since there isn't a good way
 181       // to do a variable number of input values currently, we just fill them
 182       // with undef.
 183       for (PHINode &PHI : Block->phis())
 184         PHI.addIncoming(UndefValue::get(PHI.getType()), Block);
 185     }
 186     return nullptr;
 187   };
 188   SourcePred isInt1Ty{[](ArrayRef<Value *>, const Value *V) {
 189                         return V->getType()->isIntegerTy(1);
 190                       },
 191                       std::nullopt};
 192   return {Weight, {isInt1Ty}, buildSplitBlock};
 193 }
 194
 195 OpDescriptor llvm::fuzzerop::gepDescriptor(unsigned Weight) {
 196   auto buildGEP = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 197     // TODO: It would be better to generate a random type here, rather than
 198     // generating a random value and picking its type.
 199     Type *Ty = Srcs[0]->getType()->isOpaquePointerTy()
 200                    ? Srcs[1]->getType()
 201                    : Srcs[0]->getType()->getNonOpaquePointerElementType();
 202     auto Indices = ArrayRef(Srcs).drop_front(2);
 203     return GetElementPtrInst::Create(Ty, Srcs[0], Indices, "G", Inst);
 204   };
 205   // TODO: Handle aggregates and vectors
 206   // TODO: Support multiple indices.
 207   // TODO: Try to avoid meaningless accesses.
 208   SourcePred sizedType(
 209       [](ArrayRef<Value *>, const Value *V) { return V->getType()->isSized(); },
 210       std::nullopt);
 211   return {Weight, {sizedPtrType(), sizedType, anyIntType()}, buildGEP};
 212 }
 213
 214 static uint64_t getAggregateNumElements(Type *T) {
 215   assert(T->isAggregateType() && "Not a struct or array");
 216   if (isa<StructType>(T))
 217     return T->getStructNumElements();
 218   return T->getArrayNumElements();
 219 }
 220
 221 static SourcePred validExtractValueIndex() {
 222   auto Pred = [](ArrayRef<Value *> Cur, const Value *V) {
 223     if (auto *CI = dyn_cast<ConstantInt>(V))
 224       if (!CI->uge(getAggregateNumElements(Cur[0]->getType())))
 225         return true;
 226     return false;
 227   };
 228   auto Make = [](ArrayRef<Value *> Cur, ArrayRef<Type *> Ts) {
 229     std::vector<Constant *> Result;
 230     auto *Int32Ty = Type::getInt32Ty(Cur[0]->getContext());
 231     uint64_t N = getAggregateNumElements(Cur[0]->getType());
 232     // Create indices at the start, end, and middle, but avoid dups.
 233     Result.push_back(ConstantInt::get(Int32Ty, 0));
 234     if (N > 1)
 235       Result.push_back(ConstantInt::get(Int32Ty, N - 1));
 236     if (N > 2)
 237       Result.push_back(ConstantInt::get(Int32Ty, N / 2));
 238     return Result;
 239   };
 240   return {Pred, Make};
 241 }
 242
 243 OpDescriptor llvm::fuzzerop::extractValueDescriptor(unsigned Weight) {
 244   auto buildExtract = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 245     // TODO: It's pretty inefficient to shuffle this all through constants.
 246     unsigned Idx = cast<ConstantInt>(Srcs[1])->getZExtValue();
 247     return ExtractValueInst::Create(Srcs[0], {Idx}, "E", Inst);
 248   };
 249   // TODO: Should we handle multiple indices?
 250   return {Weight, {anyAggregateType(), validExtractValueIndex()}, buildExtract};
 251 }
 252
 253 static SourcePred matchScalarInAggregate() {
 254   auto Pred = [](ArrayRef<Value *> Cur, const Value *V) {
 255     if (auto *ArrayT = dyn_cast<ArrayType>(Cur[0]->getType()))
 256       return V->getType() == ArrayT->getElementType();
 257
 258     auto *STy = cast<StructType>(Cur[0]->getType());
 259     for (int I = 0, E = STy->getNumElements(); I < E; ++I)
 260       if (STy->getTypeAtIndex(I) == V->getType())
 261         return true;
 262     return false;
 263   };
 264   auto Make = [](ArrayRef<Value *> Cur, ArrayRef<Type *>) {
 265     if (auto *ArrayT = dyn_cast<ArrayType>(Cur[0]->getType()))
 266       return makeConstantsWithType(ArrayT->getElementType());
 267
 268     std::vector<Constant *> Result;
 269     auto *STy = cast<StructType>(Cur[0]->getType());
 270     for (int I = 0, E = STy->getNumElements(); I < E; ++I)
 271       makeConstantsWithType(STy->getTypeAtIndex(I), Result);
 272     return Result;
 273   };
 274   return {Pred, Make};
 275 }
 276
 277 static SourcePred validInsertValueIndex() {
 278   auto Pred = [](ArrayRef<Value *> Cur, const Value *V) {
 279     if (auto *CI = dyn_cast<ConstantInt>(V))
 280       if (CI->getBitWidth() == 32) {
 281         Type *Indexed = ExtractValueInst::getIndexedType(Cur[0]->getType(),
 282                                                          CI->getZExtValue());
 283         return Indexed == Cur[1]->getType();
 284       }
 285     return false;
 286   };
 287   auto Make = [](ArrayRef<Value *> Cur, ArrayRef<Type *> Ts) {
 288     std::vector<Constant *> Result;
 289     auto *Int32Ty = Type::getInt32Ty(Cur[0]->getContext());
 290     auto *BaseTy = Cur[0]->getType();
 291     int I = 0;
 292     while (Type *Indexed = ExtractValueInst::getIndexedType(BaseTy, I)) {
 293       if (Indexed == Cur[1]->getType())
 294         Result.push_back(ConstantInt::get(Int32Ty, I));
 295       ++I;
 296     }
 297     return Result;
 298   };
 299   return {Pred, Make};
 300 }
 301
 302 OpDescriptor llvm::fuzzerop::insertValueDescriptor(unsigned Weight) {
 303   auto buildInsert = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 304     // TODO: It's pretty inefficient to shuffle this all through constants.
 305     unsigned Idx = cast<ConstantInt>(Srcs[2])->getZExtValue();
 306     return InsertValueInst::Create(Srcs[0], Srcs[1], {Idx}, "I", Inst);
 307   };
 308   return {
 309       Weight,
 310       {anyAggregateType(), matchScalarInAggregate(), validInsertValueIndex()},
 311       buildInsert};
 312 }
 313
 314 OpDescriptor llvm::fuzzerop::extractElementDescriptor(unsigned Weight) {
 315   auto buildExtract = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 316     return ExtractElementInst::Create(Srcs[0], Srcs[1], "E", Inst);
 317   };
 318   // TODO: Try to avoid undefined accesses.
 319   return {Weight, {anyVectorType(), anyIntType()}, buildExtract};
 320 }
 321
 322 OpDescriptor llvm::fuzzerop::insertElementDescriptor(unsigned Weight) {
 323   auto buildInsert = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 324     return InsertElementInst::Create(Srcs[0], Srcs[1], Srcs[2], "I", Inst);
 325   };
 326   // TODO: Try to avoid undefined accesses.
 327   return {Weight,
 328           {anyVectorType(), matchScalarOfFirstType(), anyIntType()},
 329           buildInsert};
 330 }
 331
 332 static SourcePred validShuffleVectorIndex() {
 333   auto Pred = [](ArrayRef<Value *> Cur, const Value *V) {
 334     return ShuffleVectorInst::isValidOperands(Cur[0], Cur[1], V);
 335   };
 336   auto Make = [](ArrayRef<Value *> Cur, ArrayRef<Type *> Ts) {
 337     auto *FirstTy = cast<VectorType>(Cur[0]->getType());
 338     auto *Int32Ty = Type::getInt32Ty(Cur[0]->getContext());
 339     // TODO: It's straighforward to make up reasonable values, but listing them
 340     // exhaustively would be insane. Come up with a couple of sensible ones.
 341     return std::vector<Constant *>{
 342         UndefValue::get(VectorType::get(Int32Ty, FirstTy->getElementCount()))};
 343   };
 344   return {Pred, Make};
 345 }
 346
 347 OpDescriptor llvm::fuzzerop::shuffleVectorDescriptor(unsigned Weight) {
 348   auto buildShuffle = [](ArrayRef<Value *> Srcs, Instruction *Inst) {
 349     return new ShuffleVectorInst(Srcs[0], Srcs[1], Srcs[2], "S", Inst);
 350   };
 351   return {Weight,
 352           {anyVectorType(), matchFirstType(), validShuffleVectorIndex()},
 353           buildShuffle};
 354 }