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