MemoryBuiltins.cpp revision 239462
1198892Srdivacky//===------ MemoryBuiltins.cpp - Identify calls to memory builtins --------===// 2198892Srdivacky// 3198892Srdivacky// The LLVM Compiler Infrastructure 4198892Srdivacky// 5198892Srdivacky// This file is distributed under the University of Illinois Open Source 6198892Srdivacky// License. See LICENSE.TXT for details. 7198892Srdivacky// 8198892Srdivacky//===----------------------------------------------------------------------===// 9198892Srdivacky// 10198892Srdivacky// This family of functions identifies calls to builtin functions that allocate 11198892Srdivacky// or free memory. 12198892Srdivacky// 13198892Srdivacky//===----------------------------------------------------------------------===// 14198892Srdivacky 15239462Sdim#define DEBUG_TYPE "memory-builtins" 16239462Sdim#include "llvm/ADT/Statistic.h" 17239462Sdim#include "llvm/ADT/STLExtras.h" 18198892Srdivacky#include "llvm/Analysis/MemoryBuiltins.h" 19239462Sdim#include "llvm/GlobalVariable.h" 20198892Srdivacky#include "llvm/Instructions.h" 21239462Sdim#include "llvm/Intrinsics.h" 22239462Sdim#include "llvm/Metadata.h" 23198892Srdivacky#include "llvm/Module.h" 24199481Srdivacky#include "llvm/Analysis/ValueTracking.h" 25239462Sdim#include "llvm/Support/Debug.h" 26239462Sdim#include "llvm/Support/MathExtras.h" 27239462Sdim#include "llvm/Support/raw_ostream.h" 28198953Srdivacky#include "llvm/Target/TargetData.h" 29239462Sdim#include "llvm/Transforms/Utils/Local.h" 30198892Srdivackyusing namespace llvm; 31198892Srdivacky 32239462Sdimenum AllocType { 33239462Sdim MallocLike = 1<<0, // allocates 34239462Sdim CallocLike = 1<<1, // allocates + bzero 35239462Sdim ReallocLike = 1<<2, // reallocates 36239462Sdim StrDupLike = 1<<3, 37239462Sdim AllocLike = MallocLike | CallocLike | StrDupLike, 38239462Sdim AnyAlloc = MallocLike | CallocLike | ReallocLike | StrDupLike 39239462Sdim}; 40198892Srdivacky 41239462Sdimstruct AllocFnsTy { 42239462Sdim const char *Name; 43239462Sdim AllocType AllocTy; 44239462Sdim unsigned char NumParams; 45239462Sdim // First and Second size parameters (or -1 if unused) 46239462Sdim signed char FstParam, SndParam; 47239462Sdim}; 48239462Sdim 49239462Sdim// FIXME: certain users need more information. E.g., SimplifyLibCalls needs to 50239462Sdim// know which functions are nounwind, noalias, nocapture parameters, etc. 51239462Sdimstatic const AllocFnsTy AllocationFnData[] = { 52239462Sdim {"malloc", MallocLike, 1, 0, -1}, 53239462Sdim {"valloc", MallocLike, 1, 0, -1}, 54239462Sdim {"_Znwj", MallocLike, 1, 0, -1}, // new(unsigned int) 55239462Sdim {"_ZnwjRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new(unsigned int, nothrow) 56239462Sdim {"_Znwm", MallocLike, 1, 0, -1}, // new(unsigned long) 57239462Sdim {"_ZnwmRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new(unsigned long, nothrow) 58239462Sdim {"_Znaj", MallocLike, 1, 0, -1}, // new[](unsigned int) 59239462Sdim {"_ZnajRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new[](unsigned int, nothrow) 60239462Sdim {"_Znam", MallocLike, 1, 0, -1}, // new[](unsigned long) 61239462Sdim {"_ZnamRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new[](unsigned long, nothrow) 62239462Sdim {"posix_memalign", MallocLike, 3, 2, -1}, 63239462Sdim {"calloc", CallocLike, 2, 0, 1}, 64239462Sdim {"realloc", ReallocLike, 2, 1, -1}, 65239462Sdim {"reallocf", ReallocLike, 2, 1, -1}, 66239462Sdim {"strdup", StrDupLike, 1, -1, -1}, 67239462Sdim {"strndup", StrDupLike, 2, 1, -1} 68239462Sdim}; 69239462Sdim 70239462Sdim 71239462Sdimstatic Function *getCalledFunction(const Value *V, bool LookThroughBitCast) { 72239462Sdim if (LookThroughBitCast) 73239462Sdim V = V->stripPointerCasts(); 74239462Sdim 75239462Sdim CallSite CS(const_cast<Value*>(V)); 76239462Sdim if (!CS.getInstruction()) 77239462Sdim return 0; 78239462Sdim 79239462Sdim Function *Callee = CS.getCalledFunction(); 80239462Sdim if (!Callee || !Callee->isDeclaration()) 81239462Sdim return 0; 82239462Sdim return Callee; 83198892Srdivacky} 84198892Srdivacky 85239462Sdim/// \brief Returns the allocation data for the given value if it is a call to a 86239462Sdim/// known allocation function, and NULL otherwise. 87239462Sdimstatic const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy, 88239462Sdim bool LookThroughBitCast = false) { 89239462Sdim Function *Callee = getCalledFunction(V, LookThroughBitCast); 90239462Sdim if (!Callee) 91239462Sdim return 0; 92198892Srdivacky 93239462Sdim unsigned i = 0; 94239462Sdim bool found = false; 95239462Sdim for ( ; i < array_lengthof(AllocationFnData); ++i) { 96239462Sdim if (Callee->getName() == AllocationFnData[i].Name) { 97239462Sdim found = true; 98239462Sdim break; 99239462Sdim } 100239462Sdim } 101239462Sdim if (!found) 102239462Sdim return 0; 103198892Srdivacky 104239462Sdim const AllocFnsTy *FnData = &AllocationFnData[i]; 105239462Sdim if ((FnData->AllocTy & AllocTy) == 0) 106239462Sdim return 0; 107239462Sdim 108239462Sdim // Check function prototype. 109239462Sdim // FIXME: Check the nobuiltin metadata?? (PR5130) 110239462Sdim int FstParam = FnData->FstParam; 111239462Sdim int SndParam = FnData->SndParam; 112226633Sdim FunctionType *FTy = Callee->getFunctionType(); 113239462Sdim 114239462Sdim if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) && 115239462Sdim FTy->getNumParams() == FnData->NumParams && 116239462Sdim (FstParam < 0 || 117239462Sdim (FTy->getParamType(FstParam)->isIntegerTy(32) || 118239462Sdim FTy->getParamType(FstParam)->isIntegerTy(64))) && 119239462Sdim (SndParam < 0 || 120239462Sdim FTy->getParamType(SndParam)->isIntegerTy(32) || 121239462Sdim FTy->getParamType(SndParam)->isIntegerTy(64))) 122239462Sdim return FnData; 123239462Sdim return 0; 124198892Srdivacky} 125198892Srdivacky 126239462Sdimstatic bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) { 127239462Sdim ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V); 128239462Sdim return CS && CS.hasFnAttr(Attribute::NoAlias); 129198892Srdivacky} 130198892Srdivacky 131239462Sdim 132239462Sdim/// \brief Tests if a value is a call or invoke to a library function that 133239462Sdim/// allocates or reallocates memory (either malloc, calloc, realloc, or strdup 134239462Sdim/// like). 135239462Sdimbool llvm::isAllocationFn(const Value *V, bool LookThroughBitCast) { 136239462Sdim return getAllocationData(V, AnyAlloc, LookThroughBitCast); 137198892Srdivacky} 138198892Srdivacky 139239462Sdim/// \brief Tests if a value is a call or invoke to a function that returns a 140239462Sdim/// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions). 141239462Sdimbool llvm::isNoAliasFn(const Value *V, bool LookThroughBitCast) { 142239462Sdim // it's safe to consider realloc as noalias since accessing the original 143239462Sdim // pointer is undefined behavior 144239462Sdim return isAllocationFn(V, LookThroughBitCast) || 145239462Sdim hasNoAliasAttr(V, LookThroughBitCast); 146198892Srdivacky} 147198892Srdivacky 148239462Sdim/// \brief Tests if a value is a call or invoke to a library function that 149239462Sdim/// allocates uninitialized memory (such as malloc). 150239462Sdimbool llvm::isMallocLikeFn(const Value *V, bool LookThroughBitCast) { 151239462Sdim return getAllocationData(V, MallocLike, LookThroughBitCast); 152198892Srdivacky} 153198892Srdivacky 154239462Sdim/// \brief Tests if a value is a call or invoke to a library function that 155239462Sdim/// allocates zero-filled memory (such as calloc). 156239462Sdimbool llvm::isCallocLikeFn(const Value *V, bool LookThroughBitCast) { 157239462Sdim return getAllocationData(V, CallocLike, LookThroughBitCast); 158198892Srdivacky} 159198892Srdivacky 160239462Sdim/// \brief Tests if a value is a call or invoke to a library function that 161239462Sdim/// allocates memory (either malloc, calloc, or strdup like). 162239462Sdimbool llvm::isAllocLikeFn(const Value *V, bool LookThroughBitCast) { 163239462Sdim return getAllocationData(V, AllocLike, LookThroughBitCast); 164239462Sdim} 165239462Sdim 166239462Sdim/// \brief Tests if a value is a call or invoke to a library function that 167239462Sdim/// reallocates memory (such as realloc). 168239462Sdimbool llvm::isReallocLikeFn(const Value *V, bool LookThroughBitCast) { 169239462Sdim return getAllocationData(V, ReallocLike, LookThroughBitCast); 170239462Sdim} 171239462Sdim 172239462Sdim/// extractMallocCall - Returns the corresponding CallInst if the instruction 173239462Sdim/// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we 174239462Sdim/// ignore InvokeInst here. 175239462Sdimconst CallInst *llvm::extractMallocCall(const Value *I) { 176239462Sdim return isMallocLikeFn(I) ? dyn_cast<CallInst>(I) : 0; 177239462Sdim} 178239462Sdim 179199481Srdivackystatic Value *computeArraySize(const CallInst *CI, const TargetData *TD, 180199481Srdivacky bool LookThroughSExt = false) { 181198892Srdivacky if (!CI) 182198892Srdivacky return NULL; 183198892Srdivacky 184198953Srdivacky // The size of the malloc's result type must be known to determine array size. 185226633Sdim Type *T = getMallocAllocatedType(CI); 186198953Srdivacky if (!T || !T->isSized() || !TD) 187198892Srdivacky return NULL; 188198892Srdivacky 189199481Srdivacky unsigned ElementSize = TD->getTypeAllocSize(T); 190226633Sdim if (StructType *ST = dyn_cast<StructType>(T)) 191199481Srdivacky ElementSize = TD->getStructLayout(ST)->getSizeInBytes(); 192198892Srdivacky 193210299Sed // If malloc call's arg can be determined to be a multiple of ElementSize, 194199481Srdivacky // return the multiple. Otherwise, return NULL. 195210299Sed Value *MallocArg = CI->getArgOperand(0); 196199481Srdivacky Value *Multiple = NULL; 197199481Srdivacky if (ComputeMultiple(MallocArg, ElementSize, Multiple, 198199481Srdivacky LookThroughSExt)) 199199481Srdivacky return Multiple; 200198892Srdivacky 201198892Srdivacky return NULL; 202198892Srdivacky} 203198892Srdivacky 204198892Srdivacky/// isArrayMalloc - Returns the corresponding CallInst if the instruction 205198892Srdivacky/// is a call to malloc whose array size can be determined and the array size 206198892Srdivacky/// is not constant 1. Otherwise, return NULL. 207199481Srdivackyconst CallInst *llvm::isArrayMalloc(const Value *I, const TargetData *TD) { 208198892Srdivacky const CallInst *CI = extractMallocCall(I); 209199481Srdivacky Value *ArraySize = computeArraySize(CI, TD); 210198892Srdivacky 211198892Srdivacky if (ArraySize && 212210299Sed ArraySize != ConstantInt::get(CI->getArgOperand(0)->getType(), 1)) 213198892Srdivacky return CI; 214198892Srdivacky 215198892Srdivacky // CI is a non-array malloc or we can't figure out that it is an array malloc. 216198892Srdivacky return NULL; 217198892Srdivacky} 218198892Srdivacky 219198892Srdivacky/// getMallocType - Returns the PointerType resulting from the malloc call. 220198953Srdivacky/// The PointerType depends on the number of bitcast uses of the malloc call: 221198953Srdivacky/// 0: PointerType is the calls' return type. 222198953Srdivacky/// 1: PointerType is the bitcast's result type. 223198953Srdivacky/// >1: Unique PointerType cannot be determined, return NULL. 224226633SdimPointerType *llvm::getMallocType(const CallInst *CI) { 225239462Sdim assert(isMallocLikeFn(CI) && "getMallocType and not malloc call"); 226198892Srdivacky 227226633Sdim PointerType *MallocType = NULL; 228198953Srdivacky unsigned NumOfBitCastUses = 0; 229198953Srdivacky 230198892Srdivacky // Determine if CallInst has a bitcast use. 231206083Srdivacky for (Value::const_use_iterator UI = CI->use_begin(), E = CI->use_end(); 232198892Srdivacky UI != E; ) 233198953Srdivacky if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) { 234198953Srdivacky MallocType = cast<PointerType>(BCI->getDestTy()); 235198953Srdivacky NumOfBitCastUses++; 236198953Srdivacky } 237198892Srdivacky 238198953Srdivacky // Malloc call has 1 bitcast use, so type is the bitcast's destination type. 239198953Srdivacky if (NumOfBitCastUses == 1) 240198953Srdivacky return MallocType; 241198892Srdivacky 242198892Srdivacky // Malloc call was not bitcast, so type is the malloc function's return type. 243198953Srdivacky if (NumOfBitCastUses == 0) 244198892Srdivacky return cast<PointerType>(CI->getType()); 245198892Srdivacky 246198892Srdivacky // Type could not be determined. 247198892Srdivacky return NULL; 248198892Srdivacky} 249198892Srdivacky 250198953Srdivacky/// getMallocAllocatedType - Returns the Type allocated by malloc call. 251198953Srdivacky/// The Type depends on the number of bitcast uses of the malloc call: 252198953Srdivacky/// 0: PointerType is the malloc calls' return type. 253198953Srdivacky/// 1: PointerType is the bitcast's result type. 254198953Srdivacky/// >1: Unique PointerType cannot be determined, return NULL. 255226633SdimType *llvm::getMallocAllocatedType(const CallInst *CI) { 256226633Sdim PointerType *PT = getMallocType(CI); 257198892Srdivacky return PT ? PT->getElementType() : NULL; 258198892Srdivacky} 259198892Srdivacky 260198892Srdivacky/// getMallocArraySize - Returns the array size of a malloc call. If the 261198892Srdivacky/// argument passed to malloc is a multiple of the size of the malloced type, 262198892Srdivacky/// then return that multiple. For non-array mallocs, the multiple is 263198892Srdivacky/// constant 1. Otherwise, return NULL for mallocs whose array size cannot be 264198892Srdivacky/// determined. 265199481SrdivackyValue *llvm::getMallocArraySize(CallInst *CI, const TargetData *TD, 266199481Srdivacky bool LookThroughSExt) { 267239462Sdim assert(isMallocLikeFn(CI) && "getMallocArraySize and not malloc call"); 268199481Srdivacky return computeArraySize(CI, TD, LookThroughSExt); 269198892Srdivacky} 270198892Srdivacky 271198892Srdivacky 272239462Sdim/// extractCallocCall - Returns the corresponding CallInst if the instruction 273239462Sdim/// is a calloc call. 274239462Sdimconst CallInst *llvm::extractCallocCall(const Value *I) { 275239462Sdim return isCallocLikeFn(I) ? cast<CallInst>(I) : 0; 276239462Sdim} 277239462Sdim 278239462Sdim 279210299Sed/// isFreeCall - Returns non-null if the value is a call to the builtin free() 280210299Sedconst CallInst *llvm::isFreeCall(const Value *I) { 281198892Srdivacky const CallInst *CI = dyn_cast<CallInst>(I); 282198892Srdivacky if (!CI) 283210299Sed return 0; 284198892Srdivacky Function *Callee = CI->getCalledFunction(); 285221345Sdim if (Callee == 0 || !Callee->isDeclaration()) 286210299Sed return 0; 287198892Srdivacky 288221345Sdim if (Callee->getName() != "free" && 289221345Sdim Callee->getName() != "_ZdlPv" && // operator delete(void*) 290221345Sdim Callee->getName() != "_ZdaPv") // operator delete[](void*) 291221345Sdim return 0; 292221345Sdim 293198892Srdivacky // Check free prototype. 294198892Srdivacky // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 295198892Srdivacky // attribute will exist. 296226633Sdim FunctionType *FTy = Callee->getFunctionType(); 297198892Srdivacky if (!FTy->getReturnType()->isVoidTy()) 298210299Sed return 0; 299198892Srdivacky if (FTy->getNumParams() != 1) 300210299Sed return 0; 301224145Sdim if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext())) 302210299Sed return 0; 303198892Srdivacky 304210299Sed return CI; 305198892Srdivacky} 306239462Sdim 307239462Sdim 308239462Sdim 309239462Sdim//===----------------------------------------------------------------------===// 310239462Sdim// Utility functions to compute size of objects. 311239462Sdim// 312239462Sdim 313239462Sdim 314239462Sdim/// \brief Compute the size of the object pointed by Ptr. Returns true and the 315239462Sdim/// object size in Size if successful, and false otherwise. 316239462Sdim/// If RoundToAlign is true, then Size is rounded up to the aligment of allocas, 317239462Sdim/// byval arguments, and global variables. 318239462Sdimbool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const TargetData *TD, 319239462Sdim bool RoundToAlign) { 320239462Sdim if (!TD) 321239462Sdim return false; 322239462Sdim 323239462Sdim ObjectSizeOffsetVisitor Visitor(TD, Ptr->getContext(), RoundToAlign); 324239462Sdim SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 325239462Sdim if (!Visitor.bothKnown(Data)) 326239462Sdim return false; 327239462Sdim 328239462Sdim APInt ObjSize = Data.first, Offset = Data.second; 329239462Sdim // check for overflow 330239462Sdim if (Offset.slt(0) || ObjSize.ult(Offset)) 331239462Sdim Size = 0; 332239462Sdim else 333239462Sdim Size = (ObjSize - Offset).getZExtValue(); 334239462Sdim return true; 335239462Sdim} 336239462Sdim 337239462Sdim 338239462SdimSTATISTIC(ObjectVisitorArgument, 339239462Sdim "Number of arguments with unsolved size and offset"); 340239462SdimSTATISTIC(ObjectVisitorLoad, 341239462Sdim "Number of load instructions with unsolved size and offset"); 342239462Sdim 343239462Sdim 344239462SdimAPInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Align) { 345239462Sdim if (RoundToAlign && Align) 346239462Sdim return APInt(IntTyBits, RoundUpToAlignment(Size.getZExtValue(), Align)); 347239462Sdim return Size; 348239462Sdim} 349239462Sdim 350239462SdimObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const TargetData *TD, 351239462Sdim LLVMContext &Context, 352239462Sdim bool RoundToAlign) 353239462Sdim: TD(TD), RoundToAlign(RoundToAlign) { 354239462Sdim IntegerType *IntTy = TD->getIntPtrType(Context); 355239462Sdim IntTyBits = IntTy->getBitWidth(); 356239462Sdim Zero = APInt::getNullValue(IntTyBits); 357239462Sdim} 358239462Sdim 359239462SdimSizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 360239462Sdim V = V->stripPointerCasts(); 361239462Sdim 362239462Sdim if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 363239462Sdim return visitGEPOperator(*GEP); 364239462Sdim if (Instruction *I = dyn_cast<Instruction>(V)) 365239462Sdim return visit(*I); 366239462Sdim if (Argument *A = dyn_cast<Argument>(V)) 367239462Sdim return visitArgument(*A); 368239462Sdim if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 369239462Sdim return visitConstantPointerNull(*P); 370239462Sdim if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 371239462Sdim return visitGlobalVariable(*GV); 372239462Sdim if (UndefValue *UV = dyn_cast<UndefValue>(V)) 373239462Sdim return visitUndefValue(*UV); 374239462Sdim if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 375239462Sdim if (CE->getOpcode() == Instruction::IntToPtr) 376239462Sdim return unknown(); // clueless 377239462Sdim 378239462Sdim DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " << *V 379239462Sdim << '\n'); 380239462Sdim return unknown(); 381239462Sdim} 382239462Sdim 383239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 384239462Sdim if (!I.getAllocatedType()->isSized()) 385239462Sdim return unknown(); 386239462Sdim 387239462Sdim APInt Size(IntTyBits, TD->getTypeAllocSize(I.getAllocatedType())); 388239462Sdim if (!I.isArrayAllocation()) 389239462Sdim return std::make_pair(align(Size, I.getAlignment()), Zero); 390239462Sdim 391239462Sdim Value *ArraySize = I.getArraySize(); 392239462Sdim if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 393239462Sdim Size *= C->getValue().zextOrSelf(IntTyBits); 394239462Sdim return std::make_pair(align(Size, I.getAlignment()), Zero); 395239462Sdim } 396239462Sdim return unknown(); 397239462Sdim} 398239462Sdim 399239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 400239462Sdim // no interprocedural analysis is done at the moment 401239462Sdim if (!A.hasByValAttr()) { 402239462Sdim ++ObjectVisitorArgument; 403239462Sdim return unknown(); 404239462Sdim } 405239462Sdim PointerType *PT = cast<PointerType>(A.getType()); 406239462Sdim APInt Size(IntTyBits, TD->getTypeAllocSize(PT->getElementType())); 407239462Sdim return std::make_pair(align(Size, A.getParamAlignment()), Zero); 408239462Sdim} 409239462Sdim 410239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) { 411239462Sdim const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc); 412239462Sdim if (!FnData) 413239462Sdim return unknown(); 414239462Sdim 415239462Sdim // handle strdup-like functions separately 416239462Sdim if (FnData->AllocTy == StrDupLike) { 417239462Sdim APInt Size(IntTyBits, GetStringLength(CS.getArgument(0))); 418239462Sdim if (!Size) 419239462Sdim return unknown(); 420239462Sdim 421239462Sdim // strndup limits strlen 422239462Sdim if (FnData->FstParam > 0) { 423239462Sdim ConstantInt *Arg= dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 424239462Sdim if (!Arg) 425239462Sdim return unknown(); 426239462Sdim 427239462Sdim APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits); 428239462Sdim if (Size.ugt(MaxSize)) 429239462Sdim Size = MaxSize + 1; 430239462Sdim } 431239462Sdim return std::make_pair(Size, Zero); 432239462Sdim } 433239462Sdim 434239462Sdim ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 435239462Sdim if (!Arg) 436239462Sdim return unknown(); 437239462Sdim 438239462Sdim APInt Size = Arg->getValue().zextOrSelf(IntTyBits); 439239462Sdim // size determined by just 1 parameter 440239462Sdim if (FnData->SndParam < 0) 441239462Sdim return std::make_pair(Size, Zero); 442239462Sdim 443239462Sdim Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam)); 444239462Sdim if (!Arg) 445239462Sdim return unknown(); 446239462Sdim 447239462Sdim Size *= Arg->getValue().zextOrSelf(IntTyBits); 448239462Sdim return std::make_pair(Size, Zero); 449239462Sdim 450239462Sdim // TODO: handle more standard functions (+ wchar cousins): 451239462Sdim // - strdup / strndup 452239462Sdim // - strcpy / strncpy 453239462Sdim // - strcat / strncat 454239462Sdim // - memcpy / memmove 455239462Sdim // - strcat / strncat 456239462Sdim // - memset 457239462Sdim} 458239462Sdim 459239462SdimSizeOffsetType 460239462SdimObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull&) { 461239462Sdim return std::make_pair(Zero, Zero); 462239462Sdim} 463239462Sdim 464239462SdimSizeOffsetType 465239462SdimObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 466239462Sdim return unknown(); 467239462Sdim} 468239462Sdim 469239462SdimSizeOffsetType 470239462SdimObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 471239462Sdim // Easy cases were already folded by previous passes. 472239462Sdim return unknown(); 473239462Sdim} 474239462Sdim 475239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) { 476239462Sdim // Ignore self-referencing GEPs, they can occur in unreachable code. 477239462Sdim if (&GEP == GEP.getPointerOperand()) 478239462Sdim return unknown(); 479239462Sdim 480239462Sdim SizeOffsetType PtrData = compute(GEP.getPointerOperand()); 481239462Sdim if (!bothKnown(PtrData) || !GEP.hasAllConstantIndices()) 482239462Sdim return unknown(); 483239462Sdim 484239462Sdim SmallVector<Value*, 8> Ops(GEP.idx_begin(), GEP.idx_end()); 485239462Sdim APInt Offset(IntTyBits,TD->getIndexedOffset(GEP.getPointerOperandType(),Ops)); 486239462Sdim return std::make_pair(PtrData.first, PtrData.second + Offset); 487239462Sdim} 488239462Sdim 489239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 490239462Sdim if (!GV.hasDefinitiveInitializer()) 491239462Sdim return unknown(); 492239462Sdim 493239462Sdim APInt Size(IntTyBits, TD->getTypeAllocSize(GV.getType()->getElementType())); 494239462Sdim return std::make_pair(align(Size, GV.getAlignment()), Zero); 495239462Sdim} 496239462Sdim 497239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 498239462Sdim // clueless 499239462Sdim return unknown(); 500239462Sdim} 501239462Sdim 502239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) { 503239462Sdim ++ObjectVisitorLoad; 504239462Sdim return unknown(); 505239462Sdim} 506239462Sdim 507239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) { 508239462Sdim // too complex to analyze statically. 509239462Sdim return unknown(); 510239462Sdim} 511239462Sdim 512239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 513239462Sdim // ignore malformed self-looping selects 514239462Sdim if (I.getTrueValue() == &I || I.getFalseValue() == &I) 515239462Sdim return unknown(); 516239462Sdim 517239462Sdim SizeOffsetType TrueSide = compute(I.getTrueValue()); 518239462Sdim SizeOffsetType FalseSide = compute(I.getFalseValue()); 519239462Sdim if (bothKnown(TrueSide) && bothKnown(FalseSide) && TrueSide == FalseSide) 520239462Sdim return TrueSide; 521239462Sdim return unknown(); 522239462Sdim} 523239462Sdim 524239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 525239462Sdim return std::make_pair(Zero, Zero); 526239462Sdim} 527239462Sdim 528239462SdimSizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 529239462Sdim DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I << '\n'); 530239462Sdim return unknown(); 531239462Sdim} 532239462Sdim 533239462Sdim 534239462SdimObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator(const TargetData *TD, 535239462Sdim LLVMContext &Context) 536239462Sdim: TD(TD), Context(Context), Builder(Context, TargetFolder(TD)), 537239462SdimVisitor(TD, Context) { 538239462Sdim IntTy = TD->getIntPtrType(Context); 539239462Sdim Zero = ConstantInt::get(IntTy, 0); 540239462Sdim} 541239462Sdim 542239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 543239462Sdim SizeOffsetEvalType Result = compute_(V); 544239462Sdim 545239462Sdim if (!bothKnown(Result)) { 546239462Sdim // erase everything that was computed in this iteration from the cache, so 547239462Sdim // that no dangling references are left behind. We could be a bit smarter if 548239462Sdim // we kept a dependency graph. It's probably not worth the complexity. 549239462Sdim for (PtrSetTy::iterator I=SeenVals.begin(), E=SeenVals.end(); I != E; ++I) { 550239462Sdim CacheMapTy::iterator CacheIt = CacheMap.find(*I); 551239462Sdim // non-computable results can be safely cached 552239462Sdim if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 553239462Sdim CacheMap.erase(CacheIt); 554239462Sdim } 555239462Sdim } 556239462Sdim 557239462Sdim SeenVals.clear(); 558239462Sdim return Result; 559239462Sdim} 560239462Sdim 561239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 562239462Sdim SizeOffsetType Const = Visitor.compute(V); 563239462Sdim if (Visitor.bothKnown(Const)) 564239462Sdim return std::make_pair(ConstantInt::get(Context, Const.first), 565239462Sdim ConstantInt::get(Context, Const.second)); 566239462Sdim 567239462Sdim V = V->stripPointerCasts(); 568239462Sdim 569239462Sdim // check cache 570239462Sdim CacheMapTy::iterator CacheIt = CacheMap.find(V); 571239462Sdim if (CacheIt != CacheMap.end()) 572239462Sdim return CacheIt->second; 573239462Sdim 574239462Sdim // always generate code immediately before the instruction being 575239462Sdim // processed, so that the generated code dominates the same BBs 576239462Sdim Instruction *PrevInsertPoint = Builder.GetInsertPoint(); 577239462Sdim if (Instruction *I = dyn_cast<Instruction>(V)) 578239462Sdim Builder.SetInsertPoint(I); 579239462Sdim 580239462Sdim // record the pointers that were handled in this run, so that they can be 581239462Sdim // cleaned later if something fails 582239462Sdim SeenVals.insert(V); 583239462Sdim 584239462Sdim // now compute the size and offset 585239462Sdim SizeOffsetEvalType Result; 586239462Sdim if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 587239462Sdim Result = visitGEPOperator(*GEP); 588239462Sdim } else if (Instruction *I = dyn_cast<Instruction>(V)) { 589239462Sdim Result = visit(*I); 590239462Sdim } else if (isa<Argument>(V) || 591239462Sdim (isa<ConstantExpr>(V) && 592239462Sdim cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 593239462Sdim isa<GlobalVariable>(V)) { 594239462Sdim // ignore values where we cannot do more than what ObjectSizeVisitor can 595239462Sdim Result = unknown(); 596239462Sdim } else { 597239462Sdim DEBUG(dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " 598239462Sdim << *V << '\n'); 599239462Sdim Result = unknown(); 600239462Sdim } 601239462Sdim 602239462Sdim if (PrevInsertPoint) 603239462Sdim Builder.SetInsertPoint(PrevInsertPoint); 604239462Sdim 605239462Sdim // Don't reuse CacheIt since it may be invalid at this point. 606239462Sdim CacheMap[V] = Result; 607239462Sdim return Result; 608239462Sdim} 609239462Sdim 610239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 611239462Sdim if (!I.getAllocatedType()->isSized()) 612239462Sdim return unknown(); 613239462Sdim 614239462Sdim // must be a VLA 615239462Sdim assert(I.isArrayAllocation()); 616239462Sdim Value *ArraySize = I.getArraySize(); 617239462Sdim Value *Size = ConstantInt::get(ArraySize->getType(), 618239462Sdim TD->getTypeAllocSize(I.getAllocatedType())); 619239462Sdim Size = Builder.CreateMul(Size, ArraySize); 620239462Sdim return std::make_pair(Size, Zero); 621239462Sdim} 622239462Sdim 623239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) { 624239462Sdim const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc); 625239462Sdim if (!FnData) 626239462Sdim return unknown(); 627239462Sdim 628239462Sdim // handle strdup-like functions separately 629239462Sdim if (FnData->AllocTy == StrDupLike) { 630239462Sdim // TODO 631239462Sdim return unknown(); 632239462Sdim } 633239462Sdim 634239462Sdim Value *FirstArg = CS.getArgument(FnData->FstParam); 635239462Sdim FirstArg = Builder.CreateZExt(FirstArg, IntTy); 636239462Sdim if (FnData->SndParam < 0) 637239462Sdim return std::make_pair(FirstArg, Zero); 638239462Sdim 639239462Sdim Value *SecondArg = CS.getArgument(FnData->SndParam); 640239462Sdim SecondArg = Builder.CreateZExt(SecondArg, IntTy); 641239462Sdim Value *Size = Builder.CreateMul(FirstArg, SecondArg); 642239462Sdim return std::make_pair(Size, Zero); 643239462Sdim 644239462Sdim // TODO: handle more standard functions (+ wchar cousins): 645239462Sdim // - strdup / strndup 646239462Sdim // - strcpy / strncpy 647239462Sdim // - strcat / strncat 648239462Sdim // - memcpy / memmove 649239462Sdim // - strcat / strncat 650239462Sdim // - memset 651239462Sdim} 652239462Sdim 653239462SdimSizeOffsetEvalType 654239462SdimObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 655239462Sdim return unknown(); 656239462Sdim} 657239462Sdim 658239462SdimSizeOffsetEvalType 659239462SdimObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 660239462Sdim return unknown(); 661239462Sdim} 662239462Sdim 663239462SdimSizeOffsetEvalType 664239462SdimObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 665239462Sdim SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 666239462Sdim if (!bothKnown(PtrData)) 667239462Sdim return unknown(); 668239462Sdim 669239462Sdim Value *Offset = EmitGEPOffset(&Builder, *TD, &GEP, /*NoAssumptions=*/true); 670239462Sdim Offset = Builder.CreateAdd(PtrData.second, Offset); 671239462Sdim return std::make_pair(PtrData.first, Offset); 672239462Sdim} 673239462Sdim 674239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 675239462Sdim // clueless 676239462Sdim return unknown(); 677239462Sdim} 678239462Sdim 679239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) { 680239462Sdim return unknown(); 681239462Sdim} 682239462Sdim 683239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 684239462Sdim // create 2 PHIs: one for size and another for offset 685239462Sdim PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 686239462Sdim PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 687239462Sdim 688239462Sdim // insert right away in the cache to handle recursive PHIs 689239462Sdim CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 690239462Sdim 691239462Sdim // compute offset/size for each PHI incoming pointer 692239462Sdim for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 693239462Sdim Builder.SetInsertPoint(PHI.getIncomingBlock(i)->getFirstInsertionPt()); 694239462Sdim SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 695239462Sdim 696239462Sdim if (!bothKnown(EdgeData)) { 697239462Sdim OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy)); 698239462Sdim OffsetPHI->eraseFromParent(); 699239462Sdim SizePHI->replaceAllUsesWith(UndefValue::get(IntTy)); 700239462Sdim SizePHI->eraseFromParent(); 701239462Sdim return unknown(); 702239462Sdim } 703239462Sdim SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 704239462Sdim OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 705239462Sdim } 706239462Sdim 707239462Sdim Value *Size = SizePHI, *Offset = OffsetPHI, *Tmp; 708239462Sdim if ((Tmp = SizePHI->hasConstantValue())) { 709239462Sdim Size = Tmp; 710239462Sdim SizePHI->replaceAllUsesWith(Size); 711239462Sdim SizePHI->eraseFromParent(); 712239462Sdim } 713239462Sdim if ((Tmp = OffsetPHI->hasConstantValue())) { 714239462Sdim Offset = Tmp; 715239462Sdim OffsetPHI->replaceAllUsesWith(Offset); 716239462Sdim OffsetPHI->eraseFromParent(); 717239462Sdim } 718239462Sdim return std::make_pair(Size, Offset); 719239462Sdim} 720239462Sdim 721239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 722239462Sdim // ignore malformed self-looping selects 723239462Sdim if (I.getTrueValue() == &I || I.getFalseValue() == &I) 724239462Sdim return unknown(); 725239462Sdim 726239462Sdim SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 727239462Sdim SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 728239462Sdim 729239462Sdim if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 730239462Sdim return unknown(); 731239462Sdim if (TrueSide == FalseSide) 732239462Sdim return TrueSide; 733239462Sdim 734239462Sdim Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 735239462Sdim FalseSide.first); 736239462Sdim Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 737239462Sdim FalseSide.second); 738239462Sdim return std::make_pair(Size, Offset); 739239462Sdim} 740239462Sdim 741239462SdimSizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 742239462Sdim DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I <<'\n'); 743239462Sdim return unknown(); 744239462Sdim} 745