1 //===- Evaluator.cpp - LLVM IR evaluator ----------------------------------===// 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 // Function evaluator for LLVM IR. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Transforms/Utils/Evaluator.h" 14 #include "llvm/ADT/DenseMap.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallPtrSet.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/Analysis/ConstantFolding.h" 19 #include "llvm/IR/BasicBlock.h" 20 #include "llvm/IR/Constant.h" 21 #include "llvm/IR/Constants.h" 22 #include "llvm/IR/DataLayout.h" 23 #include "llvm/IR/DerivedTypes.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/GlobalAlias.h" 26 #include "llvm/IR/GlobalValue.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/InstrTypes.h" 29 #include "llvm/IR/Instruction.h" 30 #include "llvm/IR/Instructions.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/Intrinsics.h" 33 #include "llvm/IR/Operator.h" 34 #include "llvm/IR/Type.h" 35 #include "llvm/IR/User.h" 36 #include "llvm/IR/Value.h" 37 #include "llvm/Support/Casting.h" 38 #include "llvm/Support/Debug.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include <iterator> 41 42 #define DEBUG_TYPE "evaluator" 43 44 using namespace llvm; 45 46 static inline bool 47 isSimpleEnoughValueToCommit(Constant *C, 48 SmallPtrSetImpl<Constant *> &SimpleConstants, 49 const DataLayout &DL); 50 51 /// Return true if the specified constant can be handled by the code generator. 52 /// We don't want to generate something like: 53 /// void *X = &X/42; 54 /// because the code generator doesn't have a relocation that can handle that. 55 /// 56 /// This function should be called if C was not found (but just got inserted) 57 /// in SimpleConstants to avoid having to rescan the same constants all the 58 /// time. 59 static bool 60 isSimpleEnoughValueToCommitHelper(Constant *C, 61 SmallPtrSetImpl<Constant *> &SimpleConstants, 62 const DataLayout &DL) { 63 // Simple global addresses are supported, do not allow dllimport or 64 // thread-local globals. 65 if (auto *GV = dyn_cast<GlobalValue>(C)) 66 return !GV->hasDLLImportStorageClass() && !GV->isThreadLocal(); 67 68 // Simple integer, undef, constant aggregate zero, etc are all supported. 69 if (C->getNumOperands() == 0 || isa<BlockAddress>(C)) 70 return true; 71 72 // Aggregate values are safe if all their elements are. 73 if (isa<ConstantAggregate>(C)) { 74 for (Value *Op : C->operands()) 75 if (!isSimpleEnoughValueToCommit(cast<Constant>(Op), SimpleConstants, DL)) 76 return false; 77 return true; 78 } 79 80 // We don't know exactly what relocations are allowed in constant expressions, 81 // so we allow &global+constantoffset, which is safe and uniformly supported 82 // across targets. 83 ConstantExpr *CE = cast<ConstantExpr>(C); 84 switch (CE->getOpcode()) { 85 case Instruction::BitCast: 86 // Bitcast is fine if the casted value is fine. 87 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 88 89 case Instruction::IntToPtr: 90 case Instruction::PtrToInt: 91 // int <=> ptr is fine if the int type is the same size as the 92 // pointer type. 93 if (DL.getTypeSizeInBits(CE->getType()) != 94 DL.getTypeSizeInBits(CE->getOperand(0)->getType())) 95 return false; 96 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 97 98 // GEP is fine if it is simple + constant offset. 99 case Instruction::GetElementPtr: 100 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 101 if (!isa<ConstantInt>(CE->getOperand(i))) 102 return false; 103 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 104 105 case Instruction::Add: 106 // We allow simple+cst. 107 if (!isa<ConstantInt>(CE->getOperand(1))) 108 return false; 109 return isSimpleEnoughValueToCommit(CE->getOperand(0), SimpleConstants, DL); 110 } 111 return false; 112 } 113 114 static inline bool 115 isSimpleEnoughValueToCommit(Constant *C, 116 SmallPtrSetImpl<Constant *> &SimpleConstants, 117 const DataLayout &DL) { 118 // If we already checked this constant, we win. 119 if (!SimpleConstants.insert(C).second) 120 return true; 121 // Check the constant. 122 return isSimpleEnoughValueToCommitHelper(C, SimpleConstants, DL); 123 } 124 125 /// Return true if this constant is simple enough for us to understand. In 126 /// particular, if it is a cast to anything other than from one pointer type to 127 /// another pointer type, we punt. We basically just support direct accesses to 128 /// globals and GEP's of globals. This should be kept up to date with 129 /// CommitValueTo. 130 static bool isSimpleEnoughPointerToCommit(Constant *C) { 131 // Conservatively, avoid aggregate types. This is because we don't 132 // want to worry about them partially overlapping other stores. 133 if (!cast<PointerType>(C->getType())->getElementType()->isSingleValueType()) 134 return false; 135 136 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 137 // Do not allow weak/*_odr/linkonce linkage or external globals. 138 return GV->hasUniqueInitializer(); 139 140 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 141 // Handle a constantexpr gep. 142 if (CE->getOpcode() == Instruction::GetElementPtr && 143 isa<GlobalVariable>(CE->getOperand(0)) && 144 cast<GEPOperator>(CE)->isInBounds()) { 145 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0)); 146 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or 147 // external globals. 148 if (!GV->hasUniqueInitializer()) 149 return false; 150 151 // The first index must be zero. 152 ConstantInt *CI = dyn_cast<ConstantInt>(*std::next(CE->op_begin())); 153 if (!CI || !CI->isZero()) return false; 154 155 // The remaining indices must be compile-time known integers within the 156 // notional bounds of the corresponding static array types. 157 if (!CE->isGEPWithNoNotionalOverIndexing()) 158 return false; 159 160 return ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE); 161 162 // A constantexpr bitcast from a pointer to another pointer is a no-op, 163 // and we know how to evaluate it by moving the bitcast from the pointer 164 // operand to the value operand. 165 } else if (CE->getOpcode() == Instruction::BitCast && 166 isa<GlobalVariable>(CE->getOperand(0))) { 167 // Do not allow weak/*_odr/linkonce/dllimport/dllexport linkage or 168 // external globals. 169 return cast<GlobalVariable>(CE->getOperand(0))->hasUniqueInitializer(); 170 } 171 } 172 173 return false; 174 } 175 176 /// Apply 'Func' to Ptr. If this returns nullptr, introspect the pointer's 177 /// type and walk down through the initial elements to obtain additional 178 /// pointers to try. Returns the first non-null return value from Func, or 179 /// nullptr if the type can't be introspected further. 180 static Constant * 181 evaluateBitcastFromPtr(Constant *Ptr, const DataLayout &DL, 182 const TargetLibraryInfo *TLI, 183 std::function<Constant *(Constant *)> Func) { 184 Constant *Val; 185 while (!(Val = Func(Ptr))) { 186 // If Ty is a non-opaque struct, we can convert the pointer to the struct 187 // into a pointer to its first member. 188 // FIXME: This could be extended to support arrays as well. 189 Type *Ty = cast<PointerType>(Ptr->getType())->getElementType(); 190 if (!isa<StructType>(Ty) || cast<StructType>(Ty)->isOpaque()) 191 break; 192 193 IntegerType *IdxTy = IntegerType::get(Ty->getContext(), 32); 194 Constant *IdxZero = ConstantInt::get(IdxTy, 0, false); 195 Constant *const IdxList[] = {IdxZero, IdxZero}; 196 197 Ptr = ConstantExpr::getGetElementPtr(Ty, Ptr, IdxList); 198 Ptr = ConstantFoldConstant(Ptr, DL, TLI); 199 } 200 return Val; 201 } 202 203 static Constant *getInitializer(Constant *C) { 204 auto *GV = dyn_cast<GlobalVariable>(C); 205 return GV && GV->hasDefinitiveInitializer() ? GV->getInitializer() : nullptr; 206 } 207 208 /// Return the value that would be computed by a load from P after the stores 209 /// reflected by 'memory' have been performed. If we can't decide, return null. 210 Constant *Evaluator::ComputeLoadResult(Constant *P) { 211 // If this memory location has been recently stored, use the stored value: it 212 // is the most up-to-date. 213 auto findMemLoc = [this](Constant *Ptr) { return MutatedMemory.lookup(Ptr); }; 214 215 if (Constant *Val = findMemLoc(P)) 216 return Val; 217 218 // Access it. 219 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) { 220 if (GV->hasDefinitiveInitializer()) 221 return GV->getInitializer(); 222 return nullptr; 223 } 224 225 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(P)) { 226 switch (CE->getOpcode()) { 227 // Handle a constantexpr getelementptr. 228 case Instruction::GetElementPtr: 229 if (auto *I = getInitializer(CE->getOperand(0))) 230 return ConstantFoldLoadThroughGEPConstantExpr(I, CE); 231 break; 232 // Handle a constantexpr bitcast. 233 case Instruction::BitCast: 234 // We're evaluating a load through a pointer that was bitcast to a 235 // different type. See if the "from" pointer has recently been stored. 236 // If it hasn't, we may still be able to find a stored pointer by 237 // introspecting the type. 238 Constant *Val = 239 evaluateBitcastFromPtr(CE->getOperand(0), DL, TLI, findMemLoc); 240 if (!Val) 241 Val = getInitializer(CE->getOperand(0)); 242 if (Val) 243 return ConstantFoldLoadThroughBitcast( 244 Val, P->getType()->getPointerElementType(), DL); 245 break; 246 } 247 } 248 249 return nullptr; // don't know how to evaluate. 250 } 251 252 static Function *getFunction(Constant *C) { 253 if (auto *Fn = dyn_cast<Function>(C)) 254 return Fn; 255 256 if (auto *Alias = dyn_cast<GlobalAlias>(C)) 257 if (auto *Fn = dyn_cast<Function>(Alias->getAliasee())) 258 return Fn; 259 return nullptr; 260 } 261 262 Function * 263 Evaluator::getCalleeWithFormalArgs(CallBase &CB, 264 SmallVectorImpl<Constant *> &Formals) { 265 auto *V = CB.getCalledOperand(); 266 if (auto *Fn = getFunction(getVal(V))) 267 return getFormalParams(CB, Fn, Formals) ? Fn : nullptr; 268 269 auto *CE = dyn_cast<ConstantExpr>(V); 270 if (!CE || CE->getOpcode() != Instruction::BitCast || 271 !getFormalParams(CB, getFunction(CE->getOperand(0)), Formals)) 272 return nullptr; 273 274 return dyn_cast<Function>( 275 ConstantFoldLoadThroughBitcast(CE, CE->getOperand(0)->getType(), DL)); 276 } 277 278 bool Evaluator::getFormalParams(CallBase &CB, Function *F, 279 SmallVectorImpl<Constant *> &Formals) { 280 if (!F) 281 return false; 282 283 auto *FTy = F->getFunctionType(); 284 if (FTy->getNumParams() > CB.getNumArgOperands()) { 285 LLVM_DEBUG(dbgs() << "Too few arguments for function.\n"); 286 return false; 287 } 288 289 auto ArgI = CB.arg_begin(); 290 for (auto ParI = FTy->param_begin(), ParE = FTy->param_end(); ParI != ParE; 291 ++ParI) { 292 auto *ArgC = ConstantFoldLoadThroughBitcast(getVal(*ArgI), *ParI, DL); 293 if (!ArgC) { 294 LLVM_DEBUG(dbgs() << "Can not convert function argument.\n"); 295 return false; 296 } 297 Formals.push_back(ArgC); 298 ++ArgI; 299 } 300 return true; 301 } 302 303 /// If call expression contains bitcast then we may need to cast 304 /// evaluated return value to a type of the call expression. 305 Constant *Evaluator::castCallResultIfNeeded(Value *CallExpr, Constant *RV) { 306 ConstantExpr *CE = dyn_cast<ConstantExpr>(CallExpr); 307 if (!RV || !CE || CE->getOpcode() != Instruction::BitCast) 308 return RV; 309 310 if (auto *FT = 311 dyn_cast<FunctionType>(CE->getType()->getPointerElementType())) { 312 RV = ConstantFoldLoadThroughBitcast(RV, FT->getReturnType(), DL); 313 if (!RV) 314 LLVM_DEBUG(dbgs() << "Failed to fold bitcast call expr\n"); 315 } 316 return RV; 317 } 318 319 /// Evaluate all instructions in block BB, returning true if successful, false 320 /// if we can't evaluate it. NewBB returns the next BB that control flows into, 321 /// or null upon return. 322 bool Evaluator::EvaluateBlock(BasicBlock::iterator CurInst, 323 BasicBlock *&NextBB) { 324 // This is the main evaluation loop. 325 while (true) { 326 Constant *InstResult = nullptr; 327 328 LLVM_DEBUG(dbgs() << "Evaluating Instruction: " << *CurInst << "\n"); 329 330 if (StoreInst *SI = dyn_cast<StoreInst>(CurInst)) { 331 if (!SI->isSimple()) { 332 LLVM_DEBUG(dbgs() << "Store is not simple! Can not evaluate.\n"); 333 return false; // no volatile/atomic accesses. 334 } 335 Constant *Ptr = getVal(SI->getOperand(1)); 336 Constant *FoldedPtr = ConstantFoldConstant(Ptr, DL, TLI); 337 if (Ptr != FoldedPtr) { 338 LLVM_DEBUG(dbgs() << "Folding constant ptr expression: " << *Ptr); 339 Ptr = FoldedPtr; 340 LLVM_DEBUG(dbgs() << "; To: " << *Ptr << "\n"); 341 } 342 if (!isSimpleEnoughPointerToCommit(Ptr)) { 343 // If this is too complex for us to commit, reject it. 344 LLVM_DEBUG( 345 dbgs() << "Pointer is too complex for us to evaluate store."); 346 return false; 347 } 348 349 Constant *Val = getVal(SI->getOperand(0)); 350 351 // If this might be too difficult for the backend to handle (e.g. the addr 352 // of one global variable divided by another) then we can't commit it. 353 if (!isSimpleEnoughValueToCommit(Val, SimpleConstants, DL)) { 354 LLVM_DEBUG(dbgs() << "Store value is too complex to evaluate store. " 355 << *Val << "\n"); 356 return false; 357 } 358 359 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) { 360 if (CE->getOpcode() == Instruction::BitCast) { 361 LLVM_DEBUG(dbgs() 362 << "Attempting to resolve bitcast on constant ptr.\n"); 363 // If we're evaluating a store through a bitcast, then we need 364 // to pull the bitcast off the pointer type and push it onto the 365 // stored value. In order to push the bitcast onto the stored value, 366 // a bitcast from the pointer's element type to Val's type must be 367 // legal. If it's not, we can try introspecting the type to find a 368 // legal conversion. 369 370 auto castValTy = [&](Constant *P) -> Constant * { 371 Type *Ty = cast<PointerType>(P->getType())->getElementType(); 372 if (Constant *FV = ConstantFoldLoadThroughBitcast(Val, Ty, DL)) { 373 Ptr = P; 374 return FV; 375 } 376 return nullptr; 377 }; 378 379 Constant *NewVal = 380 evaluateBitcastFromPtr(CE->getOperand(0), DL, TLI, castValTy); 381 if (!NewVal) { 382 LLVM_DEBUG(dbgs() << "Failed to bitcast constant ptr, can not " 383 "evaluate.\n"); 384 return false; 385 } 386 387 Val = NewVal; 388 LLVM_DEBUG(dbgs() << "Evaluated bitcast: " << *Val << "\n"); 389 } 390 } 391 392 MutatedMemory[Ptr] = Val; 393 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(CurInst)) { 394 InstResult = ConstantExpr::get(BO->getOpcode(), 395 getVal(BO->getOperand(0)), 396 getVal(BO->getOperand(1))); 397 LLVM_DEBUG(dbgs() << "Found a BinaryOperator! Simplifying: " 398 << *InstResult << "\n"); 399 } else if (CmpInst *CI = dyn_cast<CmpInst>(CurInst)) { 400 InstResult = ConstantExpr::getCompare(CI->getPredicate(), 401 getVal(CI->getOperand(0)), 402 getVal(CI->getOperand(1))); 403 LLVM_DEBUG(dbgs() << "Found a CmpInst! Simplifying: " << *InstResult 404 << "\n"); 405 } else if (CastInst *CI = dyn_cast<CastInst>(CurInst)) { 406 InstResult = ConstantExpr::getCast(CI->getOpcode(), 407 getVal(CI->getOperand(0)), 408 CI->getType()); 409 LLVM_DEBUG(dbgs() << "Found a Cast! Simplifying: " << *InstResult 410 << "\n"); 411 } else if (SelectInst *SI = dyn_cast<SelectInst>(CurInst)) { 412 InstResult = ConstantExpr::getSelect(getVal(SI->getOperand(0)), 413 getVal(SI->getOperand(1)), 414 getVal(SI->getOperand(2))); 415 LLVM_DEBUG(dbgs() << "Found a Select! Simplifying: " << *InstResult 416 << "\n"); 417 } else if (auto *EVI = dyn_cast<ExtractValueInst>(CurInst)) { 418 InstResult = ConstantExpr::getExtractValue( 419 getVal(EVI->getAggregateOperand()), EVI->getIndices()); 420 LLVM_DEBUG(dbgs() << "Found an ExtractValueInst! Simplifying: " 421 << *InstResult << "\n"); 422 } else if (auto *IVI = dyn_cast<InsertValueInst>(CurInst)) { 423 InstResult = ConstantExpr::getInsertValue( 424 getVal(IVI->getAggregateOperand()), 425 getVal(IVI->getInsertedValueOperand()), IVI->getIndices()); 426 LLVM_DEBUG(dbgs() << "Found an InsertValueInst! Simplifying: " 427 << *InstResult << "\n"); 428 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurInst)) { 429 Constant *P = getVal(GEP->getOperand(0)); 430 SmallVector<Constant*, 8> GEPOps; 431 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); 432 i != e; ++i) 433 GEPOps.push_back(getVal(*i)); 434 InstResult = 435 ConstantExpr::getGetElementPtr(GEP->getSourceElementType(), P, GEPOps, 436 cast<GEPOperator>(GEP)->isInBounds()); 437 LLVM_DEBUG(dbgs() << "Found a GEP! Simplifying: " << *InstResult << "\n"); 438 } else if (LoadInst *LI = dyn_cast<LoadInst>(CurInst)) { 439 if (!LI->isSimple()) { 440 LLVM_DEBUG( 441 dbgs() << "Found a Load! Not a simple load, can not evaluate.\n"); 442 return false; // no volatile/atomic accesses. 443 } 444 445 Constant *Ptr = getVal(LI->getOperand(0)); 446 Constant *FoldedPtr = ConstantFoldConstant(Ptr, DL, TLI); 447 if (Ptr != FoldedPtr) { 448 Ptr = FoldedPtr; 449 LLVM_DEBUG(dbgs() << "Found a constant pointer expression, constant " 450 "folding: " 451 << *Ptr << "\n"); 452 } 453 InstResult = ComputeLoadResult(Ptr); 454 if (!InstResult) { 455 LLVM_DEBUG( 456 dbgs() << "Failed to compute load result. Can not evaluate load." 457 "\n"); 458 return false; // Could not evaluate load. 459 } 460 461 LLVM_DEBUG(dbgs() << "Evaluated load: " << *InstResult << "\n"); 462 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(CurInst)) { 463 if (AI->isArrayAllocation()) { 464 LLVM_DEBUG(dbgs() << "Found an array alloca. Can not evaluate.\n"); 465 return false; // Cannot handle array allocs. 466 } 467 Type *Ty = AI->getAllocatedType(); 468 AllocaTmps.push_back(std::make_unique<GlobalVariable>( 469 Ty, false, GlobalValue::InternalLinkage, UndefValue::get(Ty), 470 AI->getName(), /*TLMode=*/GlobalValue::NotThreadLocal, 471 AI->getType()->getPointerAddressSpace())); 472 InstResult = AllocaTmps.back().get(); 473 LLVM_DEBUG(dbgs() << "Found an alloca. Result: " << *InstResult << "\n"); 474 } else if (isa<CallInst>(CurInst) || isa<InvokeInst>(CurInst)) { 475 CallBase &CB = *cast<CallBase>(&*CurInst); 476 477 // Debug info can safely be ignored here. 478 if (isa<DbgInfoIntrinsic>(CB)) { 479 LLVM_DEBUG(dbgs() << "Ignoring debug info.\n"); 480 ++CurInst; 481 continue; 482 } 483 484 // Cannot handle inline asm. 485 if (CB.isInlineAsm()) { 486 LLVM_DEBUG(dbgs() << "Found inline asm, can not evaluate.\n"); 487 return false; 488 } 489 490 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CB)) { 491 if (MemSetInst *MSI = dyn_cast<MemSetInst>(II)) { 492 if (MSI->isVolatile()) { 493 LLVM_DEBUG(dbgs() << "Can not optimize a volatile memset " 494 << "intrinsic.\n"); 495 return false; 496 } 497 Constant *Ptr = getVal(MSI->getDest()); 498 Constant *Val = getVal(MSI->getValue()); 499 Constant *DestVal = ComputeLoadResult(getVal(Ptr)); 500 if (Val->isNullValue() && DestVal && DestVal->isNullValue()) { 501 // This memset is a no-op. 502 LLVM_DEBUG(dbgs() << "Ignoring no-op memset.\n"); 503 ++CurInst; 504 continue; 505 } 506 } 507 508 if (II->isLifetimeStartOrEnd()) { 509 LLVM_DEBUG(dbgs() << "Ignoring lifetime intrinsic.\n"); 510 ++CurInst; 511 continue; 512 } 513 514 if (II->getIntrinsicID() == Intrinsic::invariant_start) { 515 // We don't insert an entry into Values, as it doesn't have a 516 // meaningful return value. 517 if (!II->use_empty()) { 518 LLVM_DEBUG(dbgs() 519 << "Found unused invariant_start. Can't evaluate.\n"); 520 return false; 521 } 522 ConstantInt *Size = cast<ConstantInt>(II->getArgOperand(0)); 523 Value *PtrArg = getVal(II->getArgOperand(1)); 524 Value *Ptr = PtrArg->stripPointerCasts(); 525 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Ptr)) { 526 Type *ElemTy = GV->getValueType(); 527 if (!Size->isMinusOne() && 528 Size->getValue().getLimitedValue() >= 529 DL.getTypeStoreSize(ElemTy)) { 530 Invariants.insert(GV); 531 LLVM_DEBUG(dbgs() << "Found a global var that is an invariant: " 532 << *GV << "\n"); 533 } else { 534 LLVM_DEBUG(dbgs() 535 << "Found a global var, but can not treat it as an " 536 "invariant.\n"); 537 } 538 } 539 // Continue even if we do nothing. 540 ++CurInst; 541 continue; 542 } else if (II->getIntrinsicID() == Intrinsic::assume) { 543 LLVM_DEBUG(dbgs() << "Skipping assume intrinsic.\n"); 544 ++CurInst; 545 continue; 546 } else if (II->getIntrinsicID() == Intrinsic::sideeffect) { 547 LLVM_DEBUG(dbgs() << "Skipping sideeffect intrinsic.\n"); 548 ++CurInst; 549 continue; 550 } else if (II->getIntrinsicID() == Intrinsic::pseudoprobe) { 551 LLVM_DEBUG(dbgs() << "Skipping pseudoprobe intrinsic.\n"); 552 ++CurInst; 553 continue; 554 } 555 556 LLVM_DEBUG(dbgs() << "Unknown intrinsic. Can not evaluate.\n"); 557 return false; 558 } 559 560 // Resolve function pointers. 561 SmallVector<Constant *, 8> Formals; 562 Function *Callee = getCalleeWithFormalArgs(CB, Formals); 563 if (!Callee || Callee->isInterposable()) { 564 LLVM_DEBUG(dbgs() << "Can not resolve function pointer.\n"); 565 return false; // Cannot resolve. 566 } 567 568 if (Callee->isDeclaration()) { 569 // If this is a function we can constant fold, do it. 570 if (Constant *C = ConstantFoldCall(&CB, Callee, Formals, TLI)) { 571 InstResult = castCallResultIfNeeded(CB.getCalledOperand(), C); 572 if (!InstResult) 573 return false; 574 LLVM_DEBUG(dbgs() << "Constant folded function call. Result: " 575 << *InstResult << "\n"); 576 } else { 577 LLVM_DEBUG(dbgs() << "Can not constant fold function call.\n"); 578 return false; 579 } 580 } else { 581 if (Callee->getFunctionType()->isVarArg()) { 582 LLVM_DEBUG(dbgs() << "Can not constant fold vararg function call.\n"); 583 return false; 584 } 585 586 Constant *RetVal = nullptr; 587 // Execute the call, if successful, use the return value. 588 ValueStack.emplace_back(); 589 if (!EvaluateFunction(Callee, RetVal, Formals)) { 590 LLVM_DEBUG(dbgs() << "Failed to evaluate function.\n"); 591 return false; 592 } 593 ValueStack.pop_back(); 594 InstResult = castCallResultIfNeeded(CB.getCalledOperand(), RetVal); 595 if (RetVal && !InstResult) 596 return false; 597 598 if (InstResult) { 599 LLVM_DEBUG(dbgs() << "Successfully evaluated function. Result: " 600 << *InstResult << "\n\n"); 601 } else { 602 LLVM_DEBUG(dbgs() 603 << "Successfully evaluated function. Result: 0\n\n"); 604 } 605 } 606 } else if (CurInst->isTerminator()) { 607 LLVM_DEBUG(dbgs() << "Found a terminator instruction.\n"); 608 609 if (BranchInst *BI = dyn_cast<BranchInst>(CurInst)) { 610 if (BI->isUnconditional()) { 611 NextBB = BI->getSuccessor(0); 612 } else { 613 ConstantInt *Cond = 614 dyn_cast<ConstantInt>(getVal(BI->getCondition())); 615 if (!Cond) return false; // Cannot determine. 616 617 NextBB = BI->getSuccessor(!Cond->getZExtValue()); 618 } 619 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurInst)) { 620 ConstantInt *Val = 621 dyn_cast<ConstantInt>(getVal(SI->getCondition())); 622 if (!Val) return false; // Cannot determine. 623 NextBB = SI->findCaseValue(Val)->getCaseSuccessor(); 624 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(CurInst)) { 625 Value *Val = getVal(IBI->getAddress())->stripPointerCasts(); 626 if (BlockAddress *BA = dyn_cast<BlockAddress>(Val)) 627 NextBB = BA->getBasicBlock(); 628 else 629 return false; // Cannot determine. 630 } else if (isa<ReturnInst>(CurInst)) { 631 NextBB = nullptr; 632 } else { 633 // invoke, unwind, resume, unreachable. 634 LLVM_DEBUG(dbgs() << "Can not handle terminator."); 635 return false; // Cannot handle this terminator. 636 } 637 638 // We succeeded at evaluating this block! 639 LLVM_DEBUG(dbgs() << "Successfully evaluated block.\n"); 640 return true; 641 } else { 642 // Did not know how to evaluate this! 643 LLVM_DEBUG( 644 dbgs() << "Failed to evaluate block due to unhandled instruction." 645 "\n"); 646 return false; 647 } 648 649 if (!CurInst->use_empty()) { 650 InstResult = ConstantFoldConstant(InstResult, DL, TLI); 651 setVal(&*CurInst, InstResult); 652 } 653 654 // If we just processed an invoke, we finished evaluating the block. 655 if (InvokeInst *II = dyn_cast<InvokeInst>(CurInst)) { 656 NextBB = II->getNormalDest(); 657 LLVM_DEBUG(dbgs() << "Found an invoke instruction. Finished Block.\n\n"); 658 return true; 659 } 660 661 // Advance program counter. 662 ++CurInst; 663 } 664 } 665 666 /// Evaluate a call to function F, returning true if successful, false if we 667 /// can't evaluate it. ActualArgs contains the formal arguments for the 668 /// function. 669 bool Evaluator::EvaluateFunction(Function *F, Constant *&RetVal, 670 const SmallVectorImpl<Constant*> &ActualArgs) { 671 // Check to see if this function is already executing (recursion). If so, 672 // bail out. TODO: we might want to accept limited recursion. 673 if (is_contained(CallStack, F)) 674 return false; 675 676 CallStack.push_back(F); 677 678 // Initialize arguments to the incoming values specified. 679 unsigned ArgNo = 0; 680 for (Function::arg_iterator AI = F->arg_begin(), E = F->arg_end(); AI != E; 681 ++AI, ++ArgNo) 682 setVal(&*AI, ActualArgs[ArgNo]); 683 684 // ExecutedBlocks - We only handle non-looping, non-recursive code. As such, 685 // we can only evaluate any one basic block at most once. This set keeps 686 // track of what we have executed so we can detect recursive cases etc. 687 SmallPtrSet<BasicBlock*, 32> ExecutedBlocks; 688 689 // CurBB - The current basic block we're evaluating. 690 BasicBlock *CurBB = &F->front(); 691 692 BasicBlock::iterator CurInst = CurBB->begin(); 693 694 while (true) { 695 BasicBlock *NextBB = nullptr; // Initialized to avoid compiler warnings. 696 LLVM_DEBUG(dbgs() << "Trying to evaluate BB: " << *CurBB << "\n"); 697 698 if (!EvaluateBlock(CurInst, NextBB)) 699 return false; 700 701 if (!NextBB) { 702 // Successfully running until there's no next block means that we found 703 // the return. Fill it the return value and pop the call stack. 704 ReturnInst *RI = cast<ReturnInst>(CurBB->getTerminator()); 705 if (RI->getNumOperands()) 706 RetVal = getVal(RI->getOperand(0)); 707 CallStack.pop_back(); 708 return true; 709 } 710 711 // Okay, we succeeded in evaluating this control flow. See if we have 712 // executed the new block before. If so, we have a looping function, 713 // which we cannot evaluate in reasonable time. 714 if (!ExecutedBlocks.insert(NextBB).second) 715 return false; // looped! 716 717 // Okay, we have never been in this block before. Check to see if there 718 // are any PHI nodes. If so, evaluate them with information about where 719 // we came from. 720 PHINode *PN = nullptr; 721 for (CurInst = NextBB->begin(); 722 (PN = dyn_cast<PHINode>(CurInst)); ++CurInst) 723 setVal(PN, getVal(PN->getIncomingValueForBlock(CurBB))); 724 725 // Advance to the next block. 726 CurBB = NextBB; 727 } 728 } 729