LLVM API Documentation
00001 //===- Reader.cpp - Code to read bytecode files ---------------------------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file was developed by the LLVM research group and is distributed under 00006 // the University of Illinois Open Source License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This library implements the functionality defined in llvm/Bytecode/Reader.h 00011 // 00012 // Note that this library should be as fast as possible, reentrant, and 00013 // threadsafe!! 00014 // 00015 // TODO: Allow passing in an option to ignore the symbol table 00016 // 00017 //===----------------------------------------------------------------------===// 00018 00019 #include "Reader.h" 00020 #include "llvm/Assembly/AutoUpgrade.h" 00021 #include "llvm/Bytecode/BytecodeHandler.h" 00022 #include "llvm/BasicBlock.h" 00023 #include "llvm/CallingConv.h" 00024 #include "llvm/Constants.h" 00025 #include "llvm/InlineAsm.h" 00026 #include "llvm/Instructions.h" 00027 #include "llvm/SymbolTable.h" 00028 #include "llvm/Bytecode/Format.h" 00029 #include "llvm/Config/alloca.h" 00030 #include "llvm/Support/GetElementPtrTypeIterator.h" 00031 #include "llvm/Support/Compressor.h" 00032 #include "llvm/Support/MathExtras.h" 00033 #include "llvm/ADT/StringExtras.h" 00034 #include <sstream> 00035 #include <algorithm> 00036 using namespace llvm; 00037 00038 namespace { 00039 /// @brief A class for maintaining the slot number definition 00040 /// as a placeholder for the actual definition for forward constants defs. 00041 class ConstantPlaceHolder : public ConstantExpr { 00042 ConstantPlaceHolder(); // DO NOT IMPLEMENT 00043 void operator=(const ConstantPlaceHolder &); // DO NOT IMPLEMENT 00044 public: 00045 Use Op; 00046 ConstantPlaceHolder(const Type *Ty) 00047 : ConstantExpr(Ty, Instruction::UserOp1, &Op, 1), 00048 Op(UndefValue::get(Type::IntTy), this) { 00049 } 00050 }; 00051 } 00052 00053 // Provide some details on error 00054 inline void BytecodeReader::error(std::string err) { 00055 err += " (Vers=" ; 00056 err += itostr(RevisionNum) ; 00057 err += ", Pos=" ; 00058 err += itostr(At-MemStart); 00059 err += ")"; 00060 throw err; 00061 } 00062 00063 //===----------------------------------------------------------------------===// 00064 // Bytecode Reading Methods 00065 //===----------------------------------------------------------------------===// 00066 00067 /// Determine if the current block being read contains any more data. 00068 inline bool BytecodeReader::moreInBlock() { 00069 return At < BlockEnd; 00070 } 00071 00072 /// Throw an error if we've read past the end of the current block 00073 inline void BytecodeReader::checkPastBlockEnd(const char * block_name) { 00074 if (At > BlockEnd) 00075 error(std::string("Attempt to read past the end of ") + block_name + 00076 " block."); 00077 } 00078 00079 /// Align the buffer position to a 32 bit boundary 00080 inline void BytecodeReader::align32() { 00081 if (hasAlignment) { 00082 BufPtr Save = At; 00083 At = (const unsigned char *)((intptr_t)(At+3) & (~3UL)); 00084 if (At > Save) 00085 if (Handler) Handler->handleAlignment(At - Save); 00086 if (At > BlockEnd) 00087 error("Ran out of data while aligning!"); 00088 } 00089 } 00090 00091 /// Read a whole unsigned integer 00092 inline unsigned BytecodeReader::read_uint() { 00093 if (At+4 > BlockEnd) 00094 error("Ran out of data reading uint!"); 00095 At += 4; 00096 return At[-4] | (At[-3] << 8) | (At[-2] << 16) | (At[-1] << 24); 00097 } 00098 00099 /// Read a variable-bit-rate encoded unsigned integer 00100 inline unsigned BytecodeReader::read_vbr_uint() { 00101 unsigned Shift = 0; 00102 unsigned Result = 0; 00103 BufPtr Save = At; 00104 00105 do { 00106 if (At == BlockEnd) 00107 error("Ran out of data reading vbr_uint!"); 00108 Result |= (unsigned)((*At++) & 0x7F) << Shift; 00109 Shift += 7; 00110 } while (At[-1] & 0x80); 00111 if (Handler) Handler->handleVBR32(At-Save); 00112 return Result; 00113 } 00114 00115 /// Read a variable-bit-rate encoded unsigned 64-bit integer. 00116 inline uint64_t BytecodeReader::read_vbr_uint64() { 00117 unsigned Shift = 0; 00118 uint64_t Result = 0; 00119 BufPtr Save = At; 00120 00121 do { 00122 if (At == BlockEnd) 00123 error("Ran out of data reading vbr_uint64!"); 00124 Result |= (uint64_t)((*At++) & 0x7F) << Shift; 00125 Shift += 7; 00126 } while (At[-1] & 0x80); 00127 if (Handler) Handler->handleVBR64(At-Save); 00128 return Result; 00129 } 00130 00131 /// Read a variable-bit-rate encoded signed 64-bit integer. 00132 inline int64_t BytecodeReader::read_vbr_int64() { 00133 uint64_t R = read_vbr_uint64(); 00134 if (R & 1) { 00135 if (R != 1) 00136 return -(int64_t)(R >> 1); 00137 else // There is no such thing as -0 with integers. "-0" really means 00138 // 0x8000000000000000. 00139 return 1LL << 63; 00140 } else 00141 return (int64_t)(R >> 1); 00142 } 00143 00144 /// Read a pascal-style string (length followed by text) 00145 inline std::string BytecodeReader::read_str() { 00146 unsigned Size = read_vbr_uint(); 00147 const unsigned char *OldAt = At; 00148 At += Size; 00149 if (At > BlockEnd) // Size invalid? 00150 error("Ran out of data reading a string!"); 00151 return std::string((char*)OldAt, Size); 00152 } 00153 00154 /// Read an arbitrary block of data 00155 inline void BytecodeReader::read_data(void *Ptr, void *End) { 00156 unsigned char *Start = (unsigned char *)Ptr; 00157 unsigned Amount = (unsigned char *)End - Start; 00158 if (At+Amount > BlockEnd) 00159 error("Ran out of data!"); 00160 std::copy(At, At+Amount, Start); 00161 At += Amount; 00162 } 00163 00164 /// Read a float value in little-endian order 00165 inline void BytecodeReader::read_float(float& FloatVal) { 00166 /// FIXME: This isn't optimal, it has size problems on some platforms 00167 /// where FP is not IEEE. 00168 FloatVal = BitsToFloat(At[0] | (At[1] << 8) | (At[2] << 16) | (At[3] << 24)); 00169 At+=sizeof(uint32_t); 00170 } 00171 00172 /// Read a double value in little-endian order 00173 inline void BytecodeReader::read_double(double& DoubleVal) { 00174 /// FIXME: This isn't optimal, it has size problems on some platforms 00175 /// where FP is not IEEE. 00176 DoubleVal = BitsToDouble((uint64_t(At[0]) << 0) | (uint64_t(At[1]) << 8) | 00177 (uint64_t(At[2]) << 16) | (uint64_t(At[3]) << 24) | 00178 (uint64_t(At[4]) << 32) | (uint64_t(At[5]) << 40) | 00179 (uint64_t(At[6]) << 48) | (uint64_t(At[7]) << 56)); 00180 At+=sizeof(uint64_t); 00181 } 00182 00183 /// Read a block header and obtain its type and size 00184 inline void BytecodeReader::read_block(unsigned &Type, unsigned &Size) { 00185 if ( hasLongBlockHeaders ) { 00186 Type = read_uint(); 00187 Size = read_uint(); 00188 switch (Type) { 00189 case BytecodeFormat::Reserved_DoNotUse : 00190 error("Reserved_DoNotUse used as Module Type?"); 00191 Type = BytecodeFormat::ModuleBlockID; break; 00192 case BytecodeFormat::Module: 00193 Type = BytecodeFormat::ModuleBlockID; break; 00194 case BytecodeFormat::Function: 00195 Type = BytecodeFormat::FunctionBlockID; break; 00196 case BytecodeFormat::ConstantPool: 00197 Type = BytecodeFormat::ConstantPoolBlockID; break; 00198 case BytecodeFormat::SymbolTable: 00199 Type = BytecodeFormat::SymbolTableBlockID; break; 00200 case BytecodeFormat::ModuleGlobalInfo: 00201 Type = BytecodeFormat::ModuleGlobalInfoBlockID; break; 00202 case BytecodeFormat::GlobalTypePlane: 00203 Type = BytecodeFormat::GlobalTypePlaneBlockID; break; 00204 case BytecodeFormat::InstructionList: 00205 Type = BytecodeFormat::InstructionListBlockID; break; 00206 case BytecodeFormat::CompactionTable: 00207 Type = BytecodeFormat::CompactionTableBlockID; break; 00208 case BytecodeFormat::BasicBlock: 00209 /// This block type isn't used after version 1.1. However, we have to 00210 /// still allow the value in case this is an old bc format file. 00211 /// We just let its value creep thru. 00212 break; 00213 default: 00214 error("Invalid block id found: " + utostr(Type)); 00215 break; 00216 } 00217 } else { 00218 Size = read_uint(); 00219 Type = Size & 0x1F; // mask low order five bits 00220 Size >>= 5; // get rid of five low order bits, leaving high 27 00221 } 00222 BlockStart = At; 00223 if (At + Size > BlockEnd) 00224 error("Attempt to size a block past end of memory"); 00225 BlockEnd = At + Size; 00226 if (Handler) Handler->handleBlock(Type, BlockStart, Size); 00227 } 00228 00229 00230 /// In LLVM 1.2 and before, Types were derived from Value and so they were 00231 /// written as part of the type planes along with any other Value. In LLVM 00232 /// 1.3 this changed so that Type does not derive from Value. Consequently, 00233 /// the BytecodeReader's containers for Values can't contain Types because 00234 /// there's no inheritance relationship. This means that the "Type Type" 00235 /// plane is defunct along with the Type::TypeTyID TypeID. In LLVM 1.3 00236 /// whenever a bytecode construct must have both types and values together, 00237 /// the types are always read/written first and then the Values. Furthermore 00238 /// since Type::TypeTyID no longer exists, its value (12) now corresponds to 00239 /// Type::LabelTyID. In order to overcome this we must "sanitize" all the 00240 /// type TypeIDs we encounter. For LLVM 1.3 bytecode files, there's no change. 00241 /// For LLVM 1.2 and before, this function will decrement the type id by 00242 /// one to account for the missing Type::TypeTyID enumerator if the value is 00243 /// larger than 12 (Type::LabelTyID). If the value is exactly 12, then this 00244 /// function returns true, otherwise false. This helps detect situations 00245 /// where the pre 1.3 bytecode is indicating that what follows is a type. 00246 /// @returns true iff type id corresponds to pre 1.3 "type type" 00247 inline bool BytecodeReader::sanitizeTypeId(unsigned &TypeId) { 00248 if (hasTypeDerivedFromValue) { /// do nothing if 1.3 or later 00249 if (TypeId == Type::LabelTyID) { 00250 TypeId = Type::VoidTyID; // sanitize it 00251 return true; // indicate we got TypeTyID in pre 1.3 bytecode 00252 } else if (TypeId > Type::LabelTyID) 00253 --TypeId; // shift all planes down because type type plane is missing 00254 } 00255 return false; 00256 } 00257 00258 /// Reads a vbr uint to read in a type id and does the necessary 00259 /// conversion on it by calling sanitizeTypeId. 00260 /// @returns true iff \p TypeId read corresponds to a pre 1.3 "type type" 00261 /// @see sanitizeTypeId 00262 inline bool BytecodeReader::read_typeid(unsigned &TypeId) { 00263 TypeId = read_vbr_uint(); 00264 if ( !has32BitTypes ) 00265 if ( TypeId == 0x00FFFFFF ) 00266 TypeId = read_vbr_uint(); 00267 return sanitizeTypeId(TypeId); 00268 } 00269 00270 //===----------------------------------------------------------------------===// 00271 // IR Lookup Methods 00272 //===----------------------------------------------------------------------===// 00273 00274 /// Determine if a type id has an implicit null value 00275 inline bool BytecodeReader::hasImplicitNull(unsigned TyID) { 00276 if (!hasExplicitPrimitiveZeros) 00277 return TyID != Type::LabelTyID && TyID != Type::VoidTyID; 00278 return TyID >= Type::FirstDerivedTyID; 00279 } 00280 00281 /// Obtain a type given a typeid and account for things like compaction tables, 00282 /// function level vs module level, and the offsetting for the primitive types. 00283 const Type *BytecodeReader::getType(unsigned ID) { 00284 if (ID < Type::FirstDerivedTyID) 00285 if (const Type *T = Type::getPrimitiveType((Type::TypeID)ID)) 00286 return T; // Asked for a primitive type... 00287 00288 // Otherwise, derived types need offset... 00289 ID -= Type::FirstDerivedTyID; 00290 00291 if (!CompactionTypes.empty()) { 00292 if (ID >= CompactionTypes.size()) 00293 error("Type ID out of range for compaction table!"); 00294 return CompactionTypes[ID].first; 00295 } 00296 00297 // Is it a module-level type? 00298 if (ID < ModuleTypes.size()) 00299 return ModuleTypes[ID].get(); 00300 00301 // Nope, is it a function-level type? 00302 ID -= ModuleTypes.size(); 00303 if (ID < FunctionTypes.size()) 00304 return FunctionTypes[ID].get(); 00305 00306 error("Illegal type reference!"); 00307 return Type::VoidTy; 00308 } 00309 00310 /// Get a sanitized type id. This just makes sure that the \p ID 00311 /// is both sanitized and not the "type type" of pre-1.3 bytecode. 00312 /// @see sanitizeTypeId 00313 inline const Type* BytecodeReader::getSanitizedType(unsigned& ID) { 00314 if (sanitizeTypeId(ID)) 00315 error("Invalid type id encountered"); 00316 return getType(ID); 00317 } 00318 00319 /// This method just saves some coding. It uses read_typeid to read 00320 /// in a sanitized type id, errors that its not the type type, and 00321 /// then calls getType to return the type value. 00322 inline const Type* BytecodeReader::readSanitizedType() { 00323 unsigned ID; 00324 if (read_typeid(ID)) 00325 error("Invalid type id encountered"); 00326 return getType(ID); 00327 } 00328 00329 /// Get the slot number associated with a type accounting for primitive 00330 /// types, compaction tables, and function level vs module level. 00331 unsigned BytecodeReader::getTypeSlot(const Type *Ty) { 00332 if (Ty->isPrimitiveType()) 00333 return Ty->getTypeID(); 00334 00335 // Scan the compaction table for the type if needed. 00336 if (!CompactionTypes.empty()) { 00337 for (unsigned i = 0, e = CompactionTypes.size(); i != e; ++i) 00338 if (CompactionTypes[i].first == Ty) 00339 return Type::FirstDerivedTyID + i; 00340 00341 error("Couldn't find type specified in compaction table!"); 00342 } 00343 00344 // Check the function level types first... 00345 TypeListTy::iterator I = std::find(FunctionTypes.begin(), 00346 FunctionTypes.end(), Ty); 00347 00348 if (I != FunctionTypes.end()) 00349 return Type::FirstDerivedTyID + ModuleTypes.size() + 00350 (&*I - &FunctionTypes[0]); 00351 00352 // If we don't have our cache yet, build it now. 00353 if (ModuleTypeIDCache.empty()) { 00354 unsigned N = 0; 00355 ModuleTypeIDCache.reserve(ModuleTypes.size()); 00356 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end(); 00357 I != E; ++I, ++N) 00358 ModuleTypeIDCache.push_back(std::make_pair(*I, N)); 00359 00360 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end()); 00361 } 00362 00363 // Binary search the cache for the entry. 00364 std::vector<std::pair<const Type*, unsigned> >::iterator IT = 00365 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(), 00366 std::make_pair(Ty, 0U)); 00367 if (IT == ModuleTypeIDCache.end() || IT->first != Ty) 00368 error("Didn't find type in ModuleTypes."); 00369 00370 return Type::FirstDerivedTyID + IT->second; 00371 } 00372 00373 /// This is just like getType, but when a compaction table is in use, it is 00374 /// ignored. It also ignores function level types. 00375 /// @see getType 00376 const Type *BytecodeReader::getGlobalTableType(unsigned Slot) { 00377 if (Slot < Type::FirstDerivedTyID) { 00378 const Type *Ty = Type::getPrimitiveType((Type::TypeID)Slot); 00379 if (!Ty) 00380 error("Not a primitive type ID?"); 00381 return Ty; 00382 } 00383 Slot -= Type::FirstDerivedTyID; 00384 if (Slot >= ModuleTypes.size()) 00385 error("Illegal compaction table type reference!"); 00386 return ModuleTypes[Slot]; 00387 } 00388 00389 /// This is just like getTypeSlot, but when a compaction table is in use, it 00390 /// is ignored. It also ignores function level types. 00391 unsigned BytecodeReader::getGlobalTableTypeSlot(const Type *Ty) { 00392 if (Ty->isPrimitiveType()) 00393 return Ty->getTypeID(); 00394 00395 // If we don't have our cache yet, build it now. 00396 if (ModuleTypeIDCache.empty()) { 00397 unsigned N = 0; 00398 ModuleTypeIDCache.reserve(ModuleTypes.size()); 00399 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end(); 00400 I != E; ++I, ++N) 00401 ModuleTypeIDCache.push_back(std::make_pair(*I, N)); 00402 00403 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end()); 00404 } 00405 00406 // Binary search the cache for the entry. 00407 std::vector<std::pair<const Type*, unsigned> >::iterator IT = 00408 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(), 00409 std::make_pair(Ty, 0U)); 00410 if (IT == ModuleTypeIDCache.end() || IT->first != Ty) 00411 error("Didn't find type in ModuleTypes."); 00412 00413 return Type::FirstDerivedTyID + IT->second; 00414 } 00415 00416 /// Retrieve a value of a given type and slot number, possibly creating 00417 /// it if it doesn't already exist. 00418 Value * BytecodeReader::getValue(unsigned type, unsigned oNum, bool Create) { 00419 assert(type != Type::LabelTyID && "getValue() cannot get blocks!"); 00420 unsigned Num = oNum; 00421 00422 // If there is a compaction table active, it defines the low-level numbers. 00423 // If not, the module values define the low-level numbers. 00424 if (CompactionValues.size() > type && !CompactionValues[type].empty()) { 00425 if (Num < CompactionValues[type].size()) 00426 return CompactionValues[type][Num]; 00427 Num -= CompactionValues[type].size(); 00428 } else { 00429 // By default, the global type id is the type id passed in 00430 unsigned GlobalTyID = type; 00431 00432 // If the type plane was compactified, figure out the global type ID by 00433 // adding the derived type ids and the distance. 00434 if (!CompactionTypes.empty() && type >= Type::FirstDerivedTyID) 00435 GlobalTyID = CompactionTypes[type-Type::FirstDerivedTyID].second; 00436 00437 if (hasImplicitNull(GlobalTyID)) { 00438 const Type *Ty = getType(type); 00439 if (!isa<OpaqueType>(Ty)) { 00440 if (Num == 0) 00441 return Constant::getNullValue(Ty); 00442 --Num; 00443 } 00444 } 00445 00446 if (GlobalTyID < ModuleValues.size() && ModuleValues[GlobalTyID]) { 00447 if (Num < ModuleValues[GlobalTyID]->size()) 00448 return ModuleValues[GlobalTyID]->getOperand(Num); 00449 Num -= ModuleValues[GlobalTyID]->size(); 00450 } 00451 } 00452 00453 if (FunctionValues.size() > type && 00454 FunctionValues[type] && 00455 Num < FunctionValues[type]->size()) 00456 return FunctionValues[type]->getOperand(Num); 00457 00458 if (!Create) return 0; // Do not create a placeholder? 00459 00460 // Did we already create a place holder? 00461 std::pair<unsigned,unsigned> KeyValue(type, oNum); 00462 ForwardReferenceMap::iterator I = ForwardReferences.lower_bound(KeyValue); 00463 if (I != ForwardReferences.end() && I->first == KeyValue) 00464 return I->second; // We have already created this placeholder 00465 00466 // If the type exists (it should) 00467 if (const Type* Ty = getType(type)) { 00468 // Create the place holder 00469 Value *Val = new Argument(Ty); 00470 ForwardReferences.insert(I, std::make_pair(KeyValue, Val)); 00471 return Val; 00472 } 00473 throw "Can't create placeholder for value of type slot #" + utostr(type); 00474 } 00475 00476 /// This is just like getValue, but when a compaction table is in use, it 00477 /// is ignored. Also, no forward references or other fancy features are 00478 /// supported. 00479 Value* BytecodeReader::getGlobalTableValue(unsigned TyID, unsigned SlotNo) { 00480 if (SlotNo == 0) 00481 return Constant::getNullValue(getType(TyID)); 00482 00483 if (!CompactionTypes.empty() && TyID >= Type::FirstDerivedTyID) { 00484 TyID -= Type::FirstDerivedTyID; 00485 if (TyID >= CompactionTypes.size()) 00486 error("Type ID out of range for compaction table!"); 00487 TyID = CompactionTypes[TyID].second; 00488 } 00489 00490 --SlotNo; 00491 00492 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0 || 00493 SlotNo >= ModuleValues[TyID]->size()) { 00494 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0) 00495 error("Corrupt compaction table entry!" 00496 + utostr(TyID) + ", " + utostr(SlotNo) + ": " 00497 + utostr(ModuleValues.size())); 00498 else 00499 error("Corrupt compaction table entry!" 00500 + utostr(TyID) + ", " + utostr(SlotNo) + ": " 00501 + utostr(ModuleValues.size()) + ", " 00502 + utohexstr(reinterpret_cast<uint64_t>(((void*)ModuleValues[TyID]))) 00503 + ", " 00504 + utostr(ModuleValues[TyID]->size())); 00505 } 00506 return ModuleValues[TyID]->getOperand(SlotNo); 00507 } 00508 00509 /// Just like getValue, except that it returns a null pointer 00510 /// only on error. It always returns a constant (meaning that if the value is 00511 /// defined, but is not a constant, that is an error). If the specified 00512 /// constant hasn't been parsed yet, a placeholder is defined and used. 00513 /// Later, after the real value is parsed, the placeholder is eliminated. 00514 Constant* BytecodeReader::getConstantValue(unsigned TypeSlot, unsigned Slot) { 00515 if (Value *V = getValue(TypeSlot, Slot, false)) 00516 if (Constant *C = dyn_cast<Constant>(V)) 00517 return C; // If we already have the value parsed, just return it 00518 else 00519 error("Value for slot " + utostr(Slot) + 00520 " is expected to be a constant!"); 00521 00522 std::pair<unsigned, unsigned> Key(TypeSlot, Slot); 00523 ConstantRefsType::iterator I = ConstantFwdRefs.lower_bound(Key); 00524 00525 if (I != ConstantFwdRefs.end() && I->first == Key) { 00526 return I->second; 00527 } else { 00528 // Create a placeholder for the constant reference and 00529 // keep track of the fact that we have a forward ref to recycle it 00530 Constant *C = new ConstantPlaceHolder(getType(TypeSlot)); 00531 00532 // Keep track of the fact that we have a forward ref to recycle it 00533 ConstantFwdRefs.insert(I, std::make_pair(Key, C)); 00534 return C; 00535 } 00536 } 00537 00538 //===----------------------------------------------------------------------===// 00539 // IR Construction Methods 00540 //===----------------------------------------------------------------------===// 00541 00542 /// As values are created, they are inserted into the appropriate place 00543 /// with this method. The ValueTable argument must be one of ModuleValues 00544 /// or FunctionValues data members of this class. 00545 unsigned BytecodeReader::insertValue(Value *Val, unsigned type, 00546 ValueTable &ValueTab) { 00547 if (ValueTab.size() <= type) 00548 ValueTab.resize(type+1); 00549 00550 if (!ValueTab[type]) ValueTab[type] = new ValueList(); 00551 00552 ValueTab[type]->push_back(Val); 00553 00554 bool HasOffset = hasImplicitNull(type) && !isa<OpaqueType>(Val->getType()); 00555 return ValueTab[type]->size()-1 + HasOffset; 00556 } 00557 00558 /// Insert the arguments of a function as new values in the reader. 00559 void BytecodeReader::insertArguments(Function* F) { 00560 const FunctionType *FT = F->getFunctionType(); 00561 Function::arg_iterator AI = F->arg_begin(); 00562 for (FunctionType::param_iterator It = FT->param_begin(); 00563 It != FT->param_end(); ++It, ++AI) 00564 insertValue(AI, getTypeSlot(AI->getType()), FunctionValues); 00565 } 00566 00567 //===----------------------------------------------------------------------===// 00568 // Bytecode Parsing Methods 00569 //===----------------------------------------------------------------------===// 00570 00571 /// This method parses a single instruction. The instruction is 00572 /// inserted at the end of the \p BB provided. The arguments of 00573 /// the instruction are provided in the \p Oprnds vector. 00574 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds, 00575 BasicBlock* BB) { 00576 BufPtr SaveAt = At; 00577 00578 // Clear instruction data 00579 Oprnds.clear(); 00580 unsigned iType = 0; 00581 unsigned Opcode = 0; 00582 unsigned Op = read_uint(); 00583 00584 // bits Instruction format: Common to all formats 00585 // -------------------------- 00586 // 01-00: Opcode type, fixed to 1. 00587 // 07-02: Opcode 00588 Opcode = (Op >> 2) & 63; 00589 Oprnds.resize((Op >> 0) & 03); 00590 00591 // Extract the operands 00592 switch (Oprnds.size()) { 00593 case 1: 00594 // bits Instruction format: 00595 // -------------------------- 00596 // 19-08: Resulting type plane 00597 // 31-20: Operand #1 (if set to (2^12-1), then zero operands) 00598 // 00599 iType = (Op >> 8) & 4095; 00600 Oprnds[0] = (Op >> 20) & 4095; 00601 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands... 00602 Oprnds.resize(0); 00603 break; 00604 case 2: 00605 // bits Instruction format: 00606 // -------------------------- 00607 // 15-08: Resulting type plane 00608 // 23-16: Operand #1 00609 // 31-24: Operand #2 00610 // 00611 iType = (Op >> 8) & 255; 00612 Oprnds[0] = (Op >> 16) & 255; 00613 Oprnds[1] = (Op >> 24) & 255; 00614 break; 00615 case 3: 00616 // bits Instruction format: 00617 // -------------------------- 00618 // 13-08: Resulting type plane 00619 // 19-14: Operand #1 00620 // 25-20: Operand #2 00621 // 31-26: Operand #3 00622 // 00623 iType = (Op >> 8) & 63; 00624 Oprnds[0] = (Op >> 14) & 63; 00625 Oprnds[1] = (Op >> 20) & 63; 00626 Oprnds[2] = (Op >> 26) & 63; 00627 break; 00628 case 0: 00629 At -= 4; // Hrm, try this again... 00630 Opcode = read_vbr_uint(); 00631 Opcode >>= 2; 00632 iType = read_vbr_uint(); 00633 00634 unsigned NumOprnds = read_vbr_uint(); 00635 Oprnds.resize(NumOprnds); 00636 00637 if (NumOprnds == 0) 00638 error("Zero-argument instruction found; this is invalid."); 00639 00640 for (unsigned i = 0; i != NumOprnds; ++i) 00641 Oprnds[i] = read_vbr_uint(); 00642 align32(); 00643 break; 00644 } 00645 00646 const Type *InstTy = getSanitizedType(iType); 00647 00648 // We have enough info to inform the handler now. 00649 if (Handler) Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt); 00650 00651 // Declare the resulting instruction we'll build. 00652 Instruction *Result = 0; 00653 00654 // If this is a bytecode format that did not include the unreachable 00655 // instruction, bump up all opcodes numbers to make space. 00656 if (hasNoUnreachableInst) { 00657 if (Opcode >= Instruction::Unreachable && 00658 Opcode < 62) { 00659 ++Opcode; 00660 } 00661 } 00662 00663 // Handle binary operators 00664 if (Opcode >= Instruction::BinaryOpsBegin && 00665 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2) 00666 Result = BinaryOperator::create((Instruction::BinaryOps)Opcode, 00667 getValue(iType, Oprnds[0]), 00668 getValue(iType, Oprnds[1])); 00669 00670 bool isCall = false; 00671 switch (Opcode) { 00672 default: 00673 if (Result == 0) 00674 error("Illegal instruction read!"); 00675 break; 00676 case Instruction::VAArg: 00677 Result = new VAArgInst(getValue(iType, Oprnds[0]), 00678 getSanitizedType(Oprnds[1])); 00679 break; 00680 case 32: { //VANext_old 00681 const Type* ArgTy = getValue(iType, Oprnds[0])->getType(); 00682 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy, 00683 (Type *)0); 00684 00685 //b = vanext a, t -> 00686 //foo = alloca 1 of t 00687 //bar = vacopy a 00688 //store bar -> foo 00689 //tmp = vaarg foo, t 00690 //b = load foo 00691 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vanext.fix"); 00692 BB->getInstList().push_back(foo); 00693 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0])); 00694 BB->getInstList().push_back(bar); 00695 BB->getInstList().push_back(new StoreInst(bar, foo)); 00696 Instruction* tmp = new VAArgInst(foo, getSanitizedType(Oprnds[1])); 00697 BB->getInstList().push_back(tmp); 00698 Result = new LoadInst(foo); 00699 break; 00700 } 00701 case 33: { //VAArg_old 00702 const Type* ArgTy = getValue(iType, Oprnds[0])->getType(); 00703 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy, 00704 (Type *)0); 00705 00706 //b = vaarg a, t -> 00707 //foo = alloca 1 of t 00708 //bar = vacopy a 00709 //store bar -> foo 00710 //b = vaarg foo, t 00711 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vaarg.fix"); 00712 BB->getInstList().push_back(foo); 00713 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0])); 00714 BB->getInstList().push_back(bar); 00715 BB->getInstList().push_back(new StoreInst(bar, foo)); 00716 Result = new VAArgInst(foo, getSanitizedType(Oprnds[1])); 00717 break; 00718 } 00719 case Instruction::ExtractElement: { 00720 if (Oprnds.size() != 2) 00721 throw std::string("Invalid extractelement instruction!"); 00722 Value *V1 = getValue(iType, Oprnds[0]); 00723 Value *V2 = getValue(Type::UIntTyID, Oprnds[1]); 00724 00725 if (!ExtractElementInst::isValidOperands(V1, V2)) 00726 throw std::string("Invalid extractelement instruction!"); 00727 00728 Result = new ExtractElementInst(V1, V2); 00729 break; 00730 } 00731 case Instruction::InsertElement: { 00732 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy); 00733 if (!PackedTy || Oprnds.size() != 3) 00734 throw std::string("Invalid insertelement instruction!"); 00735 00736 Value *V1 = getValue(iType, Oprnds[0]); 00737 Value *V2 = getValue(getTypeSlot(PackedTy->getElementType()), Oprnds[1]); 00738 Value *V3 = getValue(Type::UIntTyID, Oprnds[2]); 00739 00740 if (!InsertElementInst::isValidOperands(V1, V2, V3)) 00741 throw std::string("Invalid insertelement instruction!"); 00742 Result = new InsertElementInst(V1, V2, V3); 00743 break; 00744 } 00745 case Instruction::ShuffleVector: { 00746 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy); 00747 if (!PackedTy || Oprnds.size() != 3) 00748 throw std::string("Invalid shufflevector instruction!"); 00749 Value *V1 = getValue(iType, Oprnds[0]); 00750 Value *V2 = getValue(iType, Oprnds[1]); 00751 const PackedType *EltTy = 00752 PackedType::get(Type::UIntTy, PackedTy->getNumElements()); 00753 Value *V3 = getValue(getTypeSlot(EltTy), Oprnds[2]); 00754 if (!ShuffleVectorInst::isValidOperands(V1, V2, V3)) 00755 throw std::string("Invalid shufflevector instruction!"); 00756 Result = new ShuffleVectorInst(V1, V2, V3); 00757 break; 00758 } 00759 case Instruction::Cast: 00760 Result = new CastInst(getValue(iType, Oprnds[0]), 00761 getSanitizedType(Oprnds[1])); 00762 break; 00763 case Instruction::Select: 00764 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]), 00765 getValue(iType, Oprnds[1]), 00766 getValue(iType, Oprnds[2])); 00767 break; 00768 case Instruction::PHI: { 00769 if (Oprnds.size() == 0 || (Oprnds.size() & 1)) 00770 error("Invalid phi node encountered!"); 00771 00772 PHINode *PN = new PHINode(InstTy); 00773 PN->reserveOperandSpace(Oprnds.size()); 00774 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2) 00775 PN->addIncoming(getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1])); 00776 Result = PN; 00777 break; 00778 } 00779 00780 case Instruction::Shl: 00781 case Instruction::Shr: 00782 Result = new ShiftInst((Instruction::OtherOps)Opcode, 00783 getValue(iType, Oprnds[0]), 00784 getValue(Type::UByteTyID, Oprnds[1])); 00785 break; 00786 case Instruction::Ret: 00787 if (Oprnds.size() == 0) 00788 Result = new ReturnInst(); 00789 else if (Oprnds.size() == 1) 00790 Result = new ReturnInst(getValue(iType, Oprnds[0])); 00791 else 00792 error("Unrecognized instruction!"); 00793 break; 00794 00795 case Instruction::Br: 00796 if (Oprnds.size() == 1) 00797 Result = new BranchInst(getBasicBlock(Oprnds[0])); 00798 else if (Oprnds.size() == 3) 00799 Result = new BranchInst(getBasicBlock(Oprnds[0]), 00800 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2])); 00801 else 00802 error("Invalid number of operands for a 'br' instruction!"); 00803 break; 00804 case Instruction::Switch: { 00805 if (Oprnds.size() & 1) 00806 error("Switch statement with odd number of arguments!"); 00807 00808 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]), 00809 getBasicBlock(Oprnds[1]), 00810 Oprnds.size()/2-1); 00811 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2) 00812 I->addCase(cast<ConstantInt>(getValue(iType, Oprnds[i])), 00813 getBasicBlock(Oprnds[i+1])); 00814 Result = I; 00815 break; 00816 } 00817 00818 case 58: // Call with extra operand for calling conv 00819 case 59: // tail call, Fast CC 00820 case 60: // normal call, Fast CC 00821 case 61: // tail call, C Calling Conv 00822 case Instruction::Call: { // Normal Call, C Calling Convention 00823 if (Oprnds.size() == 0) 00824 error("Invalid call instruction encountered!"); 00825 00826 Value *F = getValue(iType, Oprnds[0]); 00827 00828 unsigned CallingConv = CallingConv::C; 00829 bool isTailCall = false; 00830 00831 if (Opcode == 61 || Opcode == 59) 00832 isTailCall = true; 00833 00834 if (Opcode == 58) { 00835 isTailCall = Oprnds.back() & 1; 00836 CallingConv = Oprnds.back() >> 1; 00837 Oprnds.pop_back(); 00838 } else if (Opcode == 59 || Opcode == 60) { 00839 CallingConv = CallingConv::Fast; 00840 } 00841 00842 // Check to make sure we have a pointer to function type 00843 const PointerType *PTy = dyn_cast<PointerType>(F->getType()); 00844 if (PTy == 0) error("Call to non function pointer value!"); 00845 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType()); 00846 if (FTy == 0) error("Call to non function pointer value!"); 00847 00848 std::vector<Value *> Params; 00849 if (!FTy->isVarArg()) { 00850 FunctionType::param_iterator It = FTy->param_begin(); 00851 00852 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) { 00853 if (It == FTy->param_end()) 00854 error("Invalid call instruction!"); 00855 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i])); 00856 } 00857 if (It != FTy->param_end()) 00858 error("Invalid call instruction!"); 00859 } else { 00860 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1); 00861 00862 unsigned FirstVariableOperand; 00863 if (Oprnds.size() < FTy->getNumParams()) 00864 error("Call instruction missing operands!"); 00865 00866 // Read all of the fixed arguments 00867 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i) 00868 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i])); 00869 00870 FirstVariableOperand = FTy->getNumParams(); 00871 00872 if ((Oprnds.size()-FirstVariableOperand) & 1) 00873 error("Invalid call instruction!"); // Must be pairs of type/value 00874 00875 for (unsigned i = FirstVariableOperand, e = Oprnds.size(); 00876 i != e; i += 2) 00877 Params.push_back(getValue(Oprnds[i], Oprnds[i+1])); 00878 } 00879 00880 Result = new CallInst(F, Params); 00881 if (isTailCall) cast<CallInst>(Result)->setTailCall(); 00882 if (CallingConv) cast<CallInst>(Result)->setCallingConv(CallingConv); 00883 break; 00884 } 00885 case 56: // Invoke with encoded CC 00886 case 57: // Invoke Fast CC 00887 case Instruction::Invoke: { // Invoke C CC 00888 if (Oprnds.size() < 3) 00889 error("Invalid invoke instruction!"); 00890 Value *F = getValue(iType, Oprnds[0]); 00891 00892 // Check to make sure we have a pointer to function type 00893 const PointerType *PTy = dyn_cast<PointerType>(F->getType()); 00894 if (PTy == 0) 00895 error("Invoke to non function pointer value!"); 00896 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType()); 00897 if (FTy == 0) 00898 error("Invoke to non function pointer value!"); 00899 00900 std::vector<Value *> Params; 00901 BasicBlock *Normal, *Except; 00902 unsigned CallingConv = CallingConv::C; 00903 00904 if (Opcode == 57) 00905 CallingConv = CallingConv::Fast; 00906 else if (Opcode == 56) { 00907 CallingConv = Oprnds.back(); 00908 Oprnds.pop_back(); 00909 } 00910 00911 if (!FTy->isVarArg()) { 00912 Normal = getBasicBlock(Oprnds[1]); 00913 Except = getBasicBlock(Oprnds[2]); 00914 00915 FunctionType::param_iterator It = FTy->param_begin(); 00916 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) { 00917 if (It == FTy->param_end()) 00918 error("Invalid invoke instruction!"); 00919 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i])); 00920 } 00921 if (It != FTy->param_end()) 00922 error("Invalid invoke instruction!"); 00923 } else { 00924 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1); 00925 00926 Normal = getBasicBlock(Oprnds[0]); 00927 Except = getBasicBlock(Oprnds[1]); 00928 00929 unsigned FirstVariableArgument = FTy->getNumParams()+2; 00930 for (unsigned i = 2; i != FirstVariableArgument; ++i) 00931 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)), 00932 Oprnds[i])); 00933 00934 if (Oprnds.size()-FirstVariableArgument & 1) // Must be type/value pairs 00935 error("Invalid invoke instruction!"); 00936 00937 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2) 00938 Params.push_back(getValue(Oprnds[i], Oprnds[i+1])); 00939 } 00940 00941 Result = new InvokeInst(F, Normal, Except, Params); 00942 if (CallingConv) cast<InvokeInst>(Result)->setCallingConv(CallingConv); 00943 break; 00944 } 00945 case Instruction::Malloc: { 00946 unsigned Align = 0; 00947 if (Oprnds.size() == 2) 00948 Align = (1 << Oprnds[1]) >> 1; 00949 else if (Oprnds.size() > 2) 00950 error("Invalid malloc instruction!"); 00951 if (!isa<PointerType>(InstTy)) 00952 error("Invalid malloc instruction!"); 00953 00954 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(), 00955 getValue(Type::UIntTyID, Oprnds[0]), Align); 00956 break; 00957 } 00958 00959 case Instruction::Alloca: { 00960 unsigned Align = 0; 00961 if (Oprnds.size() == 2) 00962 Align = (1 << Oprnds[1]) >> 1; 00963 else if (Oprnds.size() > 2) 00964 error("Invalid alloca instruction!"); 00965 if (!isa<PointerType>(InstTy)) 00966 error("Invalid alloca instruction!"); 00967 00968 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(), 00969 getValue(Type::UIntTyID, Oprnds[0]), Align); 00970 break; 00971 } 00972 case Instruction::Free: 00973 if (!isa<PointerType>(InstTy)) 00974 error("Invalid free instruction!"); 00975 Result = new FreeInst(getValue(iType, Oprnds[0])); 00976 break; 00977 case Instruction::GetElementPtr: { 00978 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy)) 00979 error("Invalid getelementptr instruction!"); 00980 00981 std::vector<Value*> Idx; 00982 00983 const Type *NextTy = InstTy; 00984 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) { 00985 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy); 00986 if (!TopTy) 00987 error("Invalid getelementptr instruction!"); 00988 00989 unsigned ValIdx = Oprnds[i]; 00990 unsigned IdxTy = 0; 00991 if (!hasRestrictedGEPTypes) { 00992 // Struct indices are always uints, sequential type indices can be any 00993 // of the 32 or 64-bit integer types. The actual choice of type is 00994 // encoded in the low two bits of the slot number. 00995 if (isa<StructType>(TopTy)) 00996 IdxTy = Type::UIntTyID; 00997 else { 00998 switch (ValIdx & 3) { 00999 default: 01000 case 0: IdxTy = Type::UIntTyID; break; 01001 case 1: IdxTy = Type::IntTyID; break; 01002 case 2: IdxTy = Type::ULongTyID; break; 01003 case 3: IdxTy = Type::LongTyID; break; 01004 } 01005 ValIdx >>= 2; 01006 } 01007 } else { 01008 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID; 01009 } 01010 01011 Idx.push_back(getValue(IdxTy, ValIdx)); 01012 01013 // Convert ubyte struct indices into uint struct indices. 01014 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes) 01015 if (ConstantUInt *C = dyn_cast<ConstantUInt>(Idx.back())) 01016 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy); 01017 01018 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true); 01019 } 01020 01021 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx); 01022 break; 01023 } 01024 01025 case 62: // volatile load 01026 case Instruction::Load: 01027 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy)) 01028 error("Invalid load instruction!"); 01029 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62); 01030 break; 01031 01032 case 63: // volatile store 01033 case Instruction::Store: { 01034 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2) 01035 error("Invalid store instruction!"); 01036 01037 Value *Ptr = getValue(iType, Oprnds[1]); 01038 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType(); 01039 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr, 01040 Opcode == 63); 01041 break; 01042 } 01043 case Instruction::Unwind: 01044 if (Oprnds.size() != 0) error("Invalid unwind instruction!"); 01045 Result = new UnwindInst(); 01046 break; 01047 case Instruction::Unreachable: 01048 if (Oprnds.size() != 0) error("Invalid unreachable instruction!"); 01049 Result = new UnreachableInst(); 01050 break; 01051 } // end switch(Opcode) 01052 01053 BB->getInstList().push_back(Result); 01054 01055 unsigned TypeSlot; 01056 if (Result->getType() == InstTy) 01057 TypeSlot = iType; 01058 else 01059 TypeSlot = getTypeSlot(Result->getType()); 01060 01061 insertValue(Result, TypeSlot, FunctionValues); 01062 } 01063 01064 /// Get a particular numbered basic block, which might be a forward reference. 01065 /// This works together with ParseBasicBlock to handle these forward references 01066 /// in a clean manner. This function is used when constructing phi, br, switch, 01067 /// and other instructions that reference basic blocks. Blocks are numbered 01068 /// sequentially as they appear in the function. 01069 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) { 01070 // Make sure there is room in the table... 01071 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1); 01072 01073 // First check to see if this is a backwards reference, i.e., ParseBasicBlock 01074 // has already created this block, or if the forward reference has already 01075 // been created. 01076 if (ParsedBasicBlocks[ID]) 01077 return ParsedBasicBlocks[ID]; 01078 01079 // Otherwise, the basic block has not yet been created. Do so and add it to 01080 // the ParsedBasicBlocks list. 01081 return ParsedBasicBlocks[ID] = new BasicBlock(); 01082 } 01083 01084 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time. 01085 /// This method reads in one of the basicblock packets. This method is not used 01086 /// for bytecode files after LLVM 1.0 01087 /// @returns The basic block constructed. 01088 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) { 01089 if (Handler) Handler->handleBasicBlockBegin(BlockNo); 01090 01091 BasicBlock *BB = 0; 01092 01093 if (ParsedBasicBlocks.size() == BlockNo) 01094 ParsedBasicBlocks.push_back(BB = new BasicBlock()); 01095 else if (ParsedBasicBlocks[BlockNo] == 0) 01096 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock(); 01097 else 01098 BB = ParsedBasicBlocks[BlockNo]; 01099 01100 std::vector<unsigned> Operands; 01101 while (moreInBlock()) 01102 ParseInstruction(Operands, BB); 01103 01104 if (Handler) Handler->handleBasicBlockEnd(BlockNo); 01105 return BB; 01106 } 01107 01108 /// Parse all of the BasicBlock's & Instruction's in the body of a function. 01109 /// In post 1.0 bytecode files, we no longer emit basic block individually, 01110 /// in order to avoid per-basic-block overhead. 01111 /// @returns Rhe number of basic blocks encountered. 01112 unsigned BytecodeReader::ParseInstructionList(Function* F) { 01113 unsigned BlockNo = 0; 01114 std::vector<unsigned> Args; 01115 01116 while (moreInBlock()) { 01117 if (Handler) Handler->handleBasicBlockBegin(BlockNo); 01118 BasicBlock *BB; 01119 if (ParsedBasicBlocks.size() == BlockNo) 01120 ParsedBasicBlocks.push_back(BB = new BasicBlock()); 01121 else if (ParsedBasicBlocks[BlockNo] == 0) 01122 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock(); 01123 else 01124 BB = ParsedBasicBlocks[BlockNo]; 01125 ++BlockNo; 01126 F->getBasicBlockList().push_back(BB); 01127 01128 // Read instructions into this basic block until we get to a terminator 01129 while (moreInBlock() && !BB->getTerminator()) 01130 ParseInstruction(Args, BB); 01131 01132 if (!BB->getTerminator()) 01133 error("Non-terminated basic block found!"); 01134 01135 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1); 01136 } 01137 01138 return BlockNo; 01139 } 01140 01141 /// Parse a symbol table. This works for both module level and function 01142 /// level symbol tables. For function level symbol tables, the CurrentFunction 01143 /// parameter must be non-zero and the ST parameter must correspond to 01144 /// CurrentFunction's symbol table. For Module level symbol tables, the 01145 /// CurrentFunction argument must be zero. 01146 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction, 01147 SymbolTable *ST) { 01148 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST); 01149 01150 // Allow efficient basic block lookup by number. 01151 std::vector<BasicBlock*> BBMap; 01152 if (CurrentFunction) 01153 for (Function::iterator I = CurrentFunction->begin(), 01154 E = CurrentFunction->end(); I != E; ++I) 01155 BBMap.push_back(I); 01156 01157 /// In LLVM 1.3 we write types separately from values so 01158 /// The types are always first in the symbol table. This is 01159 /// because Type no longer derives from Value. 01160 if (!hasTypeDerivedFromValue) { 01161 // Symtab block header: [num entries] 01162 unsigned NumEntries = read_vbr_uint(); 01163 for (unsigned i = 0; i < NumEntries; ++i) { 01164 // Symtab entry: [def slot #][name] 01165 unsigned slot = read_vbr_uint(); 01166 std::string Name = read_str(); 01167 const Type* T = getType(slot); 01168 ST->insert(Name, T); 01169 } 01170 } 01171 01172 while (moreInBlock()) { 01173 // Symtab block header: [num entries][type id number] 01174 unsigned NumEntries = read_vbr_uint(); 01175 unsigned Typ = 0; 01176 bool isTypeType = read_typeid(Typ); 01177 const Type *Ty = getType(Typ); 01178 01179 for (unsigned i = 0; i != NumEntries; ++i) { 01180 // Symtab entry: [def slot #][name] 01181 unsigned slot = read_vbr_uint(); 01182 std::string Name = read_str(); 01183 01184 // if we're reading a pre 1.3 bytecode file and the type plane 01185 // is the "type type", handle it here 01186 if (isTypeType) { 01187 const Type* T = getType(slot); 01188 if (T == 0) 01189 error("Failed type look-up for name '" + Name + "'"); 01190 ST->insert(Name, T); 01191 continue; // code below must be short circuited 01192 } else { 01193 Value *V = 0; 01194 if (Typ == Type::LabelTyID) { 01195 if (slot < BBMap.size()) 01196 V = BBMap[slot]; 01197 } else { 01198 V = getValue(Typ, slot, false); // Find mapping... 01199 } 01200 if (V == 0) 01201 error("Failed value look-up for name '" + Name + "'"); 01202 V->setName(Name); 01203 } 01204 } 01205 } 01206 checkPastBlockEnd("Symbol Table"); 01207 if (Handler) Handler->handleSymbolTableEnd(); 01208 } 01209 01210 /// Read in the types portion of a compaction table. 01211 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) { 01212 for (unsigned i = 0; i != NumEntries; ++i) { 01213 unsigned TypeSlot = 0; 01214 if (read_typeid(TypeSlot)) 01215 error("Invalid type in compaction table: type type"); 01216 const Type *Typ = getGlobalTableType(TypeSlot); 01217 CompactionTypes.push_back(std::make_pair(Typ, TypeSlot)); 01218 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ); 01219 } 01220 } 01221 01222 /// Parse a compaction table. 01223 void BytecodeReader::ParseCompactionTable() { 01224 01225 // Notify handler that we're beginning a compaction table. 01226 if (Handler) Handler->handleCompactionTableBegin(); 01227 01228 // In LLVM 1.3 Type no longer derives from Value. So, 01229 // we always write them first in the compaction table 01230 // because they can't occupy a "type plane" where the 01231 // Values reside. 01232 if (! hasTypeDerivedFromValue) { 01233 unsigned NumEntries = read_vbr_uint(); 01234 ParseCompactionTypes(NumEntries); 01235 } 01236 01237 // Compaction tables live in separate blocks so we have to loop 01238 // until we've read the whole thing. 01239 while (moreInBlock()) { 01240 // Read the number of Value* entries in the compaction table 01241 unsigned NumEntries = read_vbr_uint(); 01242 unsigned Ty = 0; 01243 unsigned isTypeType = false; 01244 01245 // Decode the type from value read in. Most compaction table 01246 // planes will have one or two entries in them. If that's the 01247 // case then the length is encoded in the bottom two bits and 01248 // the higher bits encode the type. This saves another VBR value. 01249 if ((NumEntries & 3) == 3) { 01250 // In this case, both low-order bits are set (value 3). This 01251 // is a signal that the typeid follows. 01252 NumEntries >>= 2; 01253 isTypeType = read_typeid(Ty); 01254 } else { 01255 // In this case, the low-order bits specify the number of entries 01256 // and the high order bits specify the type. 01257 Ty = NumEntries >> 2; 01258 isTypeType = sanitizeTypeId(Ty); 01259 NumEntries &= 3; 01260 } 01261 01262 // if we're reading a pre 1.3 bytecode file and the type plane 01263 // is the "type type", handle it here 01264 if (isTypeType) { 01265 ParseCompactionTypes(NumEntries); 01266 } else { 01267 // Make sure we have enough room for the plane. 01268 if (Ty >= CompactionValues.size()) 01269 CompactionValues.resize(Ty+1); 01270 01271 // Make sure the plane is empty or we have some kind of error. 01272 if (!CompactionValues[Ty].empty()) 01273 error("Compaction table plane contains multiple entries!"); 01274 01275 // Notify handler about the plane. 01276 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries); 01277 01278 // Push the implicit zero. 01279 CompactionValues[Ty].push_back(Constant::getNullValue(getType(Ty))); 01280 01281 // Read in each of the entries, put them in the compaction table 01282 // and notify the handler that we have a new compaction table value. 01283 for (unsigned i = 0; i != NumEntries; ++i) { 01284 unsigned ValSlot = read_vbr_uint(); 01285 Value *V = getGlobalTableValue(Ty, ValSlot); 01286 CompactionValues[Ty].push_back(V); 01287 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot); 01288 } 01289 } 01290 } 01291 // Notify handler that the compaction table is done. 01292 if (Handler) Handler->handleCompactionTableEnd(); 01293 } 01294 01295 // Parse a single type. The typeid is read in first. If its a primitive type 01296 // then nothing else needs to be read, we know how to instantiate it. If its 01297 // a derived type, then additional data is read to fill out the type 01298 // definition. 01299 const Type *BytecodeReader::ParseType() { 01300 unsigned PrimType = 0; 01301 if (read_typeid(PrimType)) 01302 error("Invalid type (type type) in type constants!"); 01303 01304 const Type *Result = 0; 01305 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType))) 01306 return Result; 01307 01308 switch (PrimType) { 01309 case Type::FunctionTyID: { 01310 const Type *RetType = readSanitizedType(); 01311 01312 unsigned NumParams = read_vbr_uint(); 01313 01314 std::vector<const Type*> Params; 01315 while (NumParams--) 01316 Params.push_back(readSanitizedType()); 01317 01318 bool isVarArg = Params.size() && Params.back() == Type::VoidTy; 01319 if (isVarArg) Params.pop_back(); 01320 01321 Result = FunctionType::get(RetType, Params, isVarArg); 01322 break; 01323 } 01324 case Type::ArrayTyID: { 01325 const Type *ElementType = readSanitizedType(); 01326 unsigned NumElements = read_vbr_uint(); 01327 Result = ArrayType::get(ElementType, NumElements); 01328 break; 01329 } 01330 case Type::PackedTyID: { 01331 const Type *ElementType = readSanitizedType(); 01332 unsigned NumElements = read_vbr_uint(); 01333 Result = PackedType::get(ElementType, NumElements); 01334 break; 01335 } 01336 case Type::StructTyID: { 01337 std::vector<const Type*> Elements; 01338 unsigned Typ = 0; 01339 if (read_typeid(Typ)) 01340 error("Invalid element type (type type) for structure!"); 01341 01342 while (Typ) { // List is terminated by void/0 typeid 01343 Elements.push_back(getType(Typ)); 01344 if (read_typeid(Typ)) 01345 error("Invalid element type (type type) for structure!"); 01346 } 01347 01348 Result = StructType::get(Elements); 01349 break; 01350 } 01351 case Type::PointerTyID: { 01352 Result = PointerType::get(readSanitizedType()); 01353 break; 01354 } 01355 01356 case Type::OpaqueTyID: { 01357 Result = OpaqueType::get(); 01358 break; 01359 } 01360 01361 default: 01362 error("Don't know how to deserialize primitive type " + utostr(PrimType)); 01363 break; 01364 } 01365 if (Handler) Handler->handleType(Result); 01366 return Result; 01367 } 01368 01369 // ParseTypes - We have to use this weird code to handle recursive 01370 // types. We know that recursive types will only reference the current slab of 01371 // values in the type plane, but they can forward reference types before they 01372 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might 01373 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix 01374 // this ugly problem, we pessimistically insert an opaque type for each type we 01375 // are about to read. This means that forward references will resolve to 01376 // something and when we reread the type later, we can replace the opaque type 01377 // with a new resolved concrete type. 01378 // 01379 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){ 01380 assert(Tab.size() == 0 && "should not have read type constants in before!"); 01381 01382 // Insert a bunch of opaque types to be resolved later... 01383 Tab.reserve(NumEntries); 01384 for (unsigned i = 0; i != NumEntries; ++i) 01385 Tab.push_back(OpaqueType::get()); 01386 01387 if (Handler) 01388 Handler->handleTypeList(NumEntries); 01389 01390 // If we are about to resolve types, make sure the type cache is clear. 01391 if (NumEntries) 01392 ModuleTypeIDCache.clear(); 01393 01394 // Loop through reading all of the types. Forward types will make use of the 01395 // opaque types just inserted. 01396 // 01397 for (unsigned i = 0; i != NumEntries; ++i) { 01398 const Type* NewTy = ParseType(); 01399 const Type* OldTy = Tab[i].get(); 01400 if (NewTy == 0) 01401 error("Couldn't parse type!"); 01402 01403 // Don't directly push the new type on the Tab. Instead we want to replace 01404 // the opaque type we previously inserted with the new concrete value. This 01405 // approach helps with forward references to types. The refinement from the 01406 // abstract (opaque) type to the new type causes all uses of the abstract 01407 // type to use the concrete type (NewTy). This will also cause the opaque 01408 // type to be deleted. 01409 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy); 01410 01411 // This should have replaced the old opaque type with the new type in the 01412 // value table... or with a preexisting type that was already in the system. 01413 // Let's just make sure it did. 01414 assert(Tab[i] != OldTy && "refineAbstractType didn't work!"); 01415 } 01416 } 01417 01418 /// Parse a single constant value 01419 Value *BytecodeReader::ParseConstantPoolValue(unsigned TypeID) { 01420 // We must check for a ConstantExpr before switching by type because 01421 // a ConstantExpr can be of any type, and has no explicit value. 01422 // 01423 // 0 if not expr; numArgs if is expr 01424 unsigned isExprNumArgs = read_vbr_uint(); 01425 01426 if (isExprNumArgs) { 01427 if (!hasNoUndefValue) { 01428 // 'undef' is encoded with 'exprnumargs' == 1. 01429 if (isExprNumArgs == 1) 01430 return UndefValue::get(getType(TypeID)); 01431 01432 // Inline asm is encoded with exprnumargs == ~0U. 01433 if (isExprNumArgs == ~0U) { 01434 std::string AsmStr = read_str(); 01435 std::string ConstraintStr = read_str(); 01436 unsigned Flags = read_vbr_uint(); 01437 01438 const PointerType *PTy = dyn_cast<PointerType>(getType(TypeID)); 01439 const FunctionType *FTy = 01440 PTy ? dyn_cast<FunctionType>(PTy->getElementType()) : 0; 01441 01442 if (!FTy || !InlineAsm::Verify(FTy, ConstraintStr)) 01443 error("Invalid constraints for inline asm"); 01444 if (Flags & ~1U) 01445 error("Invalid flags for inline asm"); 01446 bool HasSideEffects = Flags & 1; 01447 return InlineAsm::get(FTy, AsmStr, ConstraintStr, HasSideEffects); 01448 } 01449 01450 --isExprNumArgs; 01451 } 01452 01453 // FIXME: Encoding of constant exprs could be much more compact! 01454 std::vector<Constant*> ArgVec; 01455 ArgVec.reserve(isExprNumArgs); 01456 unsigned Opcode = read_vbr_uint(); 01457 01458 // Bytecode files before LLVM 1.4 need have a missing terminator inst. 01459 if (hasNoUnreachableInst) Opcode++; 01460 01461 // Read the slot number and types of each of the arguments 01462 for (unsigned i = 0; i != isExprNumArgs; ++i) { 01463 unsigned ArgValSlot = read_vbr_uint(); 01464 unsigned ArgTypeSlot = 0; 01465 if (read_typeid(ArgTypeSlot)) 01466 error("Invalid argument type (type type) for constant value"); 01467 01468 // Get the arg value from its slot if it exists, otherwise a placeholder 01469 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot)); 01470 } 01471 01472 // Construct a ConstantExpr of the appropriate kind 01473 if (isExprNumArgs == 1) { // All one-operand expressions 01474 if (Opcode != Instruction::Cast) 01475 error("Only cast instruction has one argument for ConstantExpr"); 01476 01477 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID)); 01478 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01479 return Result; 01480 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr 01481 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end()); 01482 01483 if (hasRestrictedGEPTypes) { 01484 const Type *BaseTy = ArgVec[0]->getType(); 01485 generic_gep_type_iterator<std::vector<Constant*>::iterator> 01486 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()), 01487 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end()); 01488 for (unsigned i = 0; GTI != E; ++GTI, ++i) 01489 if (isa<StructType>(*GTI)) { 01490 if (IdxList[i]->getType() != Type::UByteTy) 01491 error("Invalid index for getelementptr!"); 01492 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy); 01493 } 01494 } 01495 01496 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList); 01497 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01498 return Result; 01499 } else if (Opcode == Instruction::Select) { 01500 if (ArgVec.size() != 3) 01501 error("Select instruction must have three arguments."); 01502 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1], 01503 ArgVec[2]); 01504 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01505 return Result; 01506 } else if (Opcode == Instruction::ExtractElement) { 01507 if (ArgVec.size() != 2 || 01508 !ExtractElementInst::isValidOperands(ArgVec[0], ArgVec[1])) 01509 error("Invalid extractelement constand expr arguments"); 01510 Constant* Result = ConstantExpr::getExtractElement(ArgVec[0], ArgVec[1]); 01511 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01512 return Result; 01513 } else if (Opcode == Instruction::InsertElement) { 01514 if (ArgVec.size() != 3 || 01515 !InsertElementInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2])) 01516 error("Invalid insertelement constand expr arguments"); 01517 01518 Constant *Result = 01519 ConstantExpr::getInsertElement(ArgVec[0], ArgVec[1], ArgVec[2]); 01520 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01521 return Result; 01522 } else if (Opcode == Instruction::ShuffleVector) { 01523 if (ArgVec.size() != 3 || 01524 !ShuffleVectorInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2])) 01525 error("Invalid shufflevector constant expr arguments."); 01526 Constant *Result = 01527 ConstantExpr::getShuffleVector(ArgVec[0], ArgVec[1], ArgVec[2]); 01528 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01529 return Result; 01530 } else { // All other 2-operand expressions 01531 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]); 01532 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result); 01533 return Result; 01534 } 01535 } 01536 01537 // Ok, not an ConstantExpr. We now know how to read the given type... 01538 const Type *Ty = getType(TypeID); 01539 Constant *Result = 0; 01540 switch (Ty->getTypeID()) { 01541 case Type::BoolTyID: { 01542 unsigned Val = read_vbr_uint(); 01543 if (Val != 0 && Val != 1) 01544 error("Invalid boolean value read."); 01545 Result = ConstantBool::get(Val == 1); 01546 if (Handler) Handler->handleConstantValue(Result); 01547 break; 01548 } 01549 01550 case Type::UByteTyID: // Unsigned integer types... 01551 case Type::UShortTyID: 01552 case Type::UIntTyID: { 01553 unsigned Val = read_vbr_uint(); 01554 if (!ConstantUInt::isValueValidForType(Ty, Val)) 01555 error("Invalid unsigned byte/short/int read."); 01556 Result = ConstantUInt::get(Ty, Val); 01557 if (Handler) Handler->handleConstantValue(Result); 01558 break; 01559 } 01560 01561 case Type::ULongTyID: 01562 Result = ConstantUInt::get(Ty, read_vbr_uint64()); 01563 if (Handler) Handler->handleConstantValue(Result); 01564 break; 01565 01566 case Type::SByteTyID: // Signed integer types... 01567 case Type::ShortTyID: 01568 case Type::IntTyID: 01569 case Type::LongTyID: { 01570 int64_t Val = read_vbr_int64(); 01571 if (!ConstantSInt::isValueValidForType(Ty, Val)) 01572 error("Invalid signed byte/short/int/long read."); 01573 Result = ConstantSInt::get(Ty, Val); 01574 if (Handler) Handler->handleConstantValue(Result); 01575 break; 01576 } 01577 01578 case Type::FloatTyID: { 01579 float Val; 01580 read_float(Val); 01581 Result = ConstantFP::get(Ty, Val); 01582 if (Handler) Handler->handleConstantValue(Result); 01583 break; 01584 } 01585 01586 case Type::DoubleTyID: { 01587 double Val; 01588 read_double(Val); 01589 Result = ConstantFP::get(Ty, Val); 01590 if (Handler) Handler->handleConstantValue(Result); 01591 break; 01592 } 01593 01594 case Type::ArrayTyID: { 01595 const ArrayType *AT = cast<ArrayType>(Ty); 01596 unsigned NumElements = AT->getNumElements(); 01597 unsigned TypeSlot = getTypeSlot(AT->getElementType()); 01598 std::vector<Constant*> Elements; 01599 Elements.reserve(NumElements); 01600 while (NumElements--) // Read all of the elements of the constant. 01601 Elements.push_back(getConstantValue(TypeSlot, 01602 read_vbr_uint())); 01603 Result = ConstantArray::get(AT, Elements); 01604 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result); 01605 break; 01606 } 01607 01608 case Type::StructTyID: { 01609 const StructType *ST = cast<StructType>(Ty); 01610 01611 std::vector<Constant *> Elements; 01612 Elements.reserve(ST->getNumElements()); 01613 for (unsigned i = 0; i != ST->getNumElements(); ++i) 01614 Elements.push_back(getConstantValue(ST->getElementType(i), 01615 read_vbr_uint())); 01616 01617 Result = ConstantStruct::get(ST, Elements); 01618 if (Handler) Handler->handleConstantStruct(ST, Elements, Result); 01619 break; 01620 } 01621 01622 case Type::PackedTyID: { 01623 const PackedType *PT = cast<PackedType>(Ty); 01624 unsigned NumElements = PT->getNumElements(); 01625 unsigned TypeSlot = getTypeSlot(PT->getElementType()); 01626 std::vector<Constant*> Elements; 01627 Elements.reserve(NumElements); 01628 while (NumElements--) // Read all of the elements of the constant. 01629 Elements.push_back(getConstantValue(TypeSlot, 01630 read_vbr_uint())); 01631 Result = ConstantPacked::get(PT, Elements); 01632 if (Handler) Handler->handleConstantPacked(PT, Elements, TypeSlot, Result); 01633 break; 01634 } 01635 01636 case Type::PointerTyID: { // ConstantPointerRef value (backwards compat). 01637 const PointerType *PT = cast<PointerType>(Ty); 01638 unsigned Slot = read_vbr_uint(); 01639 01640 // Check to see if we have already read this global variable... 01641 Value *Val = getValue(TypeID, Slot, false); 01642 if (Val) { 01643 GlobalValue *GV = dyn_cast<GlobalValue>(Val); 01644 if (!GV) error("GlobalValue not in ValueTable!"); 01645 if (Handler) Handler->handleConstantPointer(PT, Slot, GV); 01646 return GV; 01647 } else { 01648 error("Forward references are not allowed here."); 01649 } 01650 } 01651 01652 default: 01653 error("Don't know how to deserialize constant value of type '" + 01654 Ty->getDescription()); 01655 break; 01656 } 01657 01658 // Check that we didn't read a null constant if they are implicit for this 01659 // type plane. Do not do this check for constantexprs, as they may be folded 01660 // to a null value in a way that isn't predicted when a .bc file is initially 01661 // produced. 01662 assert((!isa<Constant>(Result) || !cast<Constant>(Result)->isNullValue()) || 01663 !hasImplicitNull(TypeID) && 01664 "Cannot read null values from bytecode!"); 01665 return Result; 01666 } 01667 01668 /// Resolve references for constants. This function resolves the forward 01669 /// referenced constants in the ConstantFwdRefs map. It uses the 01670 /// replaceAllUsesWith method of Value class to substitute the placeholder 01671 /// instance with the actual instance. 01672 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Typ, 01673 unsigned Slot) { 01674 ConstantRefsType::iterator I = 01675 ConstantFwdRefs.find(std::make_pair(Typ, Slot)); 01676 if (I == ConstantFwdRefs.end()) return; // Never forward referenced? 01677 01678 Value *PH = I->second; // Get the placeholder... 01679 PH->replaceAllUsesWith(NewV); 01680 delete PH; // Delete the old placeholder 01681 ConstantFwdRefs.erase(I); // Remove the map entry for it 01682 } 01683 01684 /// Parse the constant strings section. 01685 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){ 01686 for (; NumEntries; --NumEntries) { 01687 unsigned Typ = 0; 01688 if (read_typeid(Typ)) 01689 error("Invalid type (type type) for string constant"); 01690 const Type *Ty = getType(Typ); 01691 if (!isa<ArrayType>(Ty)) 01692 error("String constant data invalid!"); 01693 01694 const ArrayType *ATy = cast<ArrayType>(Ty); 01695 if (ATy->getElementType() != Type::SByteTy && 01696 ATy->getElementType() != Type::UByteTy) 01697 error("String constant data invalid!"); 01698 01699 // Read character data. The type tells us how long the string is. 01700 char *Data = reinterpret_cast<char *>(alloca(ATy->getNumElements())); 01701 read_data(Data, Data+ATy->getNumElements()); 01702 01703 std::vector<Constant*> Elements(ATy->getNumElements()); 01704 if (ATy->getElementType() == Type::SByteTy) 01705 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) 01706 Elements[i] = ConstantSInt::get(Type::SByteTy, (signed char)Data[i]); 01707 else 01708 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) 01709 Elements[i] = ConstantUInt::get(Type::UByteTy, (unsigned char)Data[i]); 01710 01711 // Create the constant, inserting it as needed. 01712 Constant *C = ConstantArray::get(ATy, Elements); 01713 unsigned Slot = insertValue(C, Typ, Tab); 01714 ResolveReferencesToConstant(C, Typ, Slot); 01715 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C)); 01716 } 01717 } 01718 01719 /// Parse the constant pool. 01720 void BytecodeReader::ParseConstantPool(ValueTable &Tab, 01721 TypeListTy &TypeTab, 01722 bool isFunction) { 01723 if (Handler) Handler->handleGlobalConstantsBegin(); 01724 01725 /// In LLVM 1.3 Type does not derive from Value so the types 01726 /// do not occupy a plane. Consequently, we read the types 01727 /// first in the constant pool. 01728 if (isFunction && !hasTypeDerivedFromValue) { 01729 unsigned NumEntries = read_vbr_uint(); 01730 ParseTypes(TypeTab, NumEntries); 01731 } 01732 01733 while (moreInBlock()) { 01734 unsigned NumEntries = read_vbr_uint(); 01735 unsigned Typ = 0; 01736 bool isTypeType = read_typeid(Typ); 01737 01738 /// In LLVM 1.2 and before, Types were written to the 01739 /// bytecode file in the "Type Type" plane (#12). 01740 /// In 1.3 plane 12 is now the label plane. Handle this here. 01741 if (isTypeType) { 01742 ParseTypes(TypeTab, NumEntries); 01743 } else if (Typ == Type::VoidTyID) { 01744 /// Use of Type::VoidTyID is a misnomer. It actually means 01745 /// that the following plane is constant strings 01746 assert(&Tab == &ModuleValues && "Cannot read strings in functions!"); 01747 ParseStringConstants(NumEntries, Tab); 01748 } else { 01749 for (unsigned i = 0; i < NumEntries; ++i) { 01750 Value *V = ParseConstantPoolValue(Typ); 01751 assert(V && "ParseConstantPoolValue returned NULL!"); 01752 unsigned Slot = insertValue(V, Typ, Tab); 01753 01754 // If we are reading a function constant table, make sure that we adjust 01755 // the slot number to be the real global constant number. 01756 // 01757 if (&Tab != &ModuleValues && Typ < ModuleValues.size() && 01758 ModuleValues[Typ]) 01759 Slot += ModuleValues[Typ]->size(); 01760 if (Constant *C = dyn_cast<Constant>(V)) 01761 ResolveReferencesToConstant(C, Typ, Slot); 01762 } 01763 } 01764 } 01765 01766 // After we have finished parsing the constant pool, we had better not have 01767 // any dangling references left. 01768 if (!ConstantFwdRefs.empty()) { 01769 ConstantRefsType::const_iterator I = ConstantFwdRefs.begin(); 01770 Constant* missingConst = I->second; 01771 error(utostr(ConstantFwdRefs.size()) + 01772 " unresolved constant reference exist. First one is '" + 01773 missingConst->getName() + "' of type '" + 01774 missingConst->getType()->getDescription() + "'."); 01775 } 01776 01777 checkPastBlockEnd("Constant Pool"); 01778 if (Handler) Handler->handleGlobalConstantsEnd(); 01779 } 01780 01781 /// Parse the contents of a function. Note that this function can be 01782 /// called lazily by materializeFunction 01783 /// @see materializeFunction 01784 void BytecodeReader::ParseFunctionBody(Function* F) { 01785 01786 unsigned FuncSize = BlockEnd - At; 01787 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage; 01788 01789 unsigned LinkageType = read_vbr_uint(); 01790 switch (LinkageType) { 01791 case 0: Linkage = GlobalValue::ExternalLinkage; break; 01792 case 1: Linkage = GlobalValue::WeakLinkage; break; 01793 case 2: Linkage = GlobalValue::AppendingLinkage; break; 01794 case 3: Linkage = GlobalValue::InternalLinkage; break; 01795 case 4: Linkage = GlobalValue::LinkOnceLinkage; break; 01796 default: 01797 error("Invalid linkage type for Function."); 01798 Linkage = GlobalValue::InternalLinkage; 01799 break; 01800 } 01801 01802 F->setLinkage(Linkage); 01803 if (Handler) Handler->handleFunctionBegin(F,FuncSize); 01804 01805 // Keep track of how many basic blocks we have read in... 01806 unsigned BlockNum = 0; 01807 bool InsertedArguments = false; 01808 01809 BufPtr MyEnd = BlockEnd; 01810 while (At < MyEnd) { 01811 unsigned Type, Size; 01812 BufPtr OldAt = At; 01813 read_block(Type, Size); 01814 01815 switch (Type) { 01816 case BytecodeFormat::ConstantPoolBlockID: 01817 if (!InsertedArguments) { 01818 // Insert arguments into the value table before we parse the first basic 01819 // block in the function, but after we potentially read in the 01820 // compaction table. 01821 insertArguments(F); 01822 InsertedArguments = true; 01823 } 01824 01825 ParseConstantPool(FunctionValues, FunctionTypes, true); 01826 break; 01827 01828 case BytecodeFormat::CompactionTableBlockID: 01829 ParseCompactionTable(); 01830 break; 01831 01832 case BytecodeFormat::BasicBlock: { 01833 if (!InsertedArguments) { 01834 // Insert arguments into the value table before we parse the first basic 01835 // block in the function, but after we potentially read in the 01836 // compaction table. 01837 insertArguments(F); 01838 InsertedArguments = true; 01839 } 01840 01841 BasicBlock *BB = ParseBasicBlock(BlockNum++); 01842 F->getBasicBlockList().push_back(BB); 01843 break; 01844 } 01845 01846 case BytecodeFormat::InstructionListBlockID: { 01847 // Insert arguments into the value table before we parse the instruction 01848 // list for the function, but after we potentially read in the compaction 01849 // table. 01850 if (!InsertedArguments) { 01851 insertArguments(F); 01852 InsertedArguments = true; 01853 } 01854 01855 if (BlockNum) 01856 error("Already parsed basic blocks!"); 01857 BlockNum = ParseInstructionList(F); 01858 break; 01859 } 01860 01861 case BytecodeFormat::SymbolTableBlockID: 01862 ParseSymbolTable(F, &F->getSymbolTable()); 01863 break; 01864 01865 default: 01866 At += Size; 01867 if (OldAt > At) 01868 error("Wrapped around reading bytecode."); 01869 break; 01870 } 01871 BlockEnd = MyEnd; 01872 01873 // Malformed bc file if read past end of block. 01874 align32(); 01875 } 01876 01877 // Make sure there were no references to non-existant basic blocks. 01878 if (BlockNum != ParsedBasicBlocks.size()) 01879 error("Illegal basic block operand reference"); 01880 01881 ParsedBasicBlocks.clear(); 01882 01883 // Resolve forward references. Replace any uses of a forward reference value 01884 // with the real value. 01885 while (!ForwardReferences.empty()) { 01886 std::map<std::pair<unsigned,unsigned>, Value*>::iterator 01887 I = ForwardReferences.begin(); 01888 Value *V = getValue(I->first.first, I->first.second, false); 01889 Value *PlaceHolder = I->second; 01890 PlaceHolder->replaceAllUsesWith(V); 01891 ForwardReferences.erase(I); 01892 delete PlaceHolder; 01893 } 01894 01895 // If upgraded intrinsic functions were detected during reading of the 01896 // module information, then we need to look for instructions that need to 01897 // be upgraded. This can't be done while the instructions are read in because 01898 // additional instructions inserted mess up the slot numbering. 01899 if (!upgradedFunctions.empty()) { 01900 for (Function::iterator BI = F->begin(), BE = F->end(); BI != BE; ++BI) 01901 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); 01902 II != IE;) 01903 if (CallInst* CI = dyn_cast<CallInst>(II++)) { 01904 std::map<Function*,Function*>::iterator FI = 01905 upgradedFunctions.find(CI->getCalledFunction()); 01906 if (FI != upgradedFunctions.end()) 01907 UpgradeIntrinsicCall(CI, FI->second); 01908 } 01909 } 01910 01911 // Clear out function-level types... 01912 FunctionTypes.clear(); 01913 CompactionTypes.clear(); 01914 CompactionValues.clear(); 01915 freeTable(FunctionValues); 01916 01917 if (Handler) Handler->handleFunctionEnd(F); 01918 } 01919 01920 /// This function parses LLVM functions lazily. It obtains the type of the 01921 /// function and records where the body of the function is in the bytecode 01922 /// buffer. The caller can then use the ParseNextFunction and 01923 /// ParseAllFunctionBodies to get handler events for the functions. 01924 void BytecodeReader::ParseFunctionLazily() { 01925 if (FunctionSignatureList.empty()) 01926 error("FunctionSignatureList empty!"); 01927 01928 Function *Func = FunctionSignatureList.back(); 01929 FunctionSignatureList.pop_back(); 01930 01931 // Save the information for future reading of the function 01932 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd); 01933 01934 // This function has a body but it's not loaded so it appears `External'. 01935 // Mark it as a `Ghost' instead to notify the users that it has a body. 01936 Func->setLinkage(GlobalValue::GhostLinkage); 01937 01938 // Pretend we've `parsed' this function 01939 At = BlockEnd; 01940 } 01941 01942 /// The ParserFunction method lazily parses one function. Use this method to 01943 /// casue the parser to parse a specific function in the module. Note that 01944 /// this will remove the function from what is to be included by 01945 /// ParseAllFunctionBodies. 01946 /// @see ParseAllFunctionBodies 01947 /// @see ParseBytecode 01948 void BytecodeReader::ParseFunction(Function* Func) { 01949 // Find {start, end} pointers and slot in the map. If not there, we're done. 01950 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func); 01951 01952 // Make sure we found it 01953 if (Fi == LazyFunctionLoadMap.end()) { 01954 error("Unrecognized function of type " + Func->getType()->getDescription()); 01955 return; 01956 } 01957 01958 BlockStart = At = Fi->second.Buf; 01959 BlockEnd = Fi->second.EndBuf; 01960 assert(Fi->first == Func && "Found wrong function?"); 01961 01962 LazyFunctionLoadMap.erase(Fi); 01963 01964 this->ParseFunctionBody(Func); 01965 } 01966 01967 /// The ParseAllFunctionBodies method parses through all the previously 01968 /// unparsed functions in the bytecode file. If you want to completely parse 01969 /// a bytecode file, this method should be called after Parsebytecode because 01970 /// Parsebytecode only records the locations in the bytecode file of where 01971 /// the function definitions are located. This function uses that information 01972 /// to materialize the functions. 01973 /// @see ParseBytecode 01974 void BytecodeReader::ParseAllFunctionBodies() { 01975 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin(); 01976 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end(); 01977 01978 while (Fi != Fe) { 01979 Function* Func = Fi->first; 01980 BlockStart = At = Fi->second.Buf; 01981 BlockEnd = Fi->second.EndBuf; 01982 ParseFunctionBody(Func); 01983 ++Fi; 01984 } 01985 LazyFunctionLoadMap.clear(); 01986 01987 } 01988 01989 /// Parse the global type list 01990 void BytecodeReader::ParseGlobalTypes() { 01991 // Read the number of types 01992 unsigned NumEntries = read_vbr_uint(); 01993 01994 // Ignore the type plane identifier for types if the bc file is pre 1.3 01995 if (hasTypeDerivedFromValue) 01996 read_vbr_uint(); 01997 01998 ParseTypes(ModuleTypes, NumEntries); 01999 } 02000 02001 /// Parse the Global info (types, global vars, constants) 02002 void BytecodeReader::ParseModuleGlobalInfo() { 02003 02004 if (Handler) Handler->handleModuleGlobalsBegin(); 02005 02006 // SectionID - If a global has an explicit section specified, this map 02007 // remembers the ID until we can translate it into a string. 02008 std::map<GlobalValue*, unsigned> SectionID; 02009 02010 // Read global variables... 02011 unsigned VarType = read_vbr_uint(); 02012 while (VarType != Type::VoidTyID) { // List is terminated by Void 02013 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 = 02014 // Linkage, bit4+ = slot# 02015 unsigned SlotNo = VarType >> 5; 02016 if (sanitizeTypeId(SlotNo)) 02017 error("Invalid type (type type) for global var!"); 02018 unsigned LinkageID = (VarType >> 2) & 7; 02019 bool isConstant = VarType & 1; 02020 bool hasInitializer = (VarType & 2) != 0; 02021 unsigned Alignment = 0; 02022 unsigned GlobalSectionID = 0; 02023 02024 // An extension word is present when linkage = 3 (internal) and hasinit = 0. 02025 if (LinkageID == 3 && !hasInitializer) { 02026 unsigned ExtWord = read_vbr_uint(); 02027 // The extension word has this format: bit 0 = has initializer, bit 1-3 = 02028 // linkage, bit 4-8 = alignment (log2), bits 10+ = future use. 02029 hasInitializer = ExtWord & 1; 02030 LinkageID = (ExtWord >> 1) & 7; 02031 Alignment = (1 << ((ExtWord >> 4) & 31)) >> 1; 02032 02033 if (ExtWord & (1 << 9)) // Has a section ID. 02034 GlobalSectionID = read_vbr_uint(); 02035 } 02036 02037 GlobalValue::LinkageTypes Linkage; 02038 switch (LinkageID) { 02039 case 0: Linkage = GlobalValue::ExternalLinkage; break; 02040 case 1: Linkage = GlobalValue::WeakLinkage; break; 02041 case 2: Linkage = GlobalValue::AppendingLinkage; break; 02042 case 3: Linkage = GlobalValue::InternalLinkage; break; 02043 case 4: Linkage = GlobalValue::LinkOnceLinkage; break; 02044 default: 02045 error("Unknown linkage type: " + utostr(LinkageID)); 02046 Linkage = GlobalValue::InternalLinkage; 02047 break; 02048 } 02049 02050 const Type *Ty = getType(SlotNo); 02051 if (!Ty) 02052 error("Global has no type! SlotNo=" + utostr(SlotNo)); 02053 02054 if (!isa<PointerType>(Ty)) 02055 error("Global not a pointer type! Ty= " + Ty->getDescription()); 02056 02057 const Type *ElTy = cast<PointerType>(Ty)->getElementType(); 02058 02059 // Create the global variable... 02060 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage, 02061 0, "", TheModule); 02062 GV->setAlignment(Alignment); 02063 insertValue(GV, SlotNo, ModuleValues); 02064 02065 if (GlobalSectionID != 0) 02066 SectionID[GV] = GlobalSectionID; 02067 02068 unsigned initSlot = 0; 02069 if (hasInitializer) { 02070 initSlot = read_vbr_uint(); 02071 GlobalInits.push_back(std::make_pair(GV, initSlot)); 02072 } 02073 02074 // Notify handler about the global value. 02075 if (Handler) 02076 Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo,initSlot); 02077 02078 // Get next item 02079 VarType = read_vbr_uint(); 02080 } 02081 02082 // Read the function objects for all of the functions that are coming 02083 unsigned FnSignature = read_vbr_uint(); 02084 02085 if (hasNoFlagsForFunctions) 02086 FnSignature = (FnSignature << 5) + 1; 02087 02088 // List is terminated by VoidTy. 02089 while (((FnSignature & (~0U >> 1)) >> 5) != Type::VoidTyID) { 02090 const Type *Ty = getType((FnSignature & (~0U >> 1)) >> 5); 02091 if (!isa<PointerType>(Ty) || 02092 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) { 02093 error("Function not a pointer to function type! Ty = " + 02094 Ty->getDescription()); 02095 } 02096 02097 // We create functions by passing the underlying FunctionType to create... 02098 const FunctionType* FTy = 02099 cast<FunctionType>(cast<PointerType>(Ty)->getElementType()); 02100 02101 // Insert the place holder. 02102 Function *Func = new Function(FTy, GlobalValue::ExternalLinkage, 02103 "", TheModule); 02104 02105 insertValue(Func, (FnSignature & (~0U >> 1)) >> 5, ModuleValues); 02106 02107 // Flags are not used yet. 02108 unsigned Flags = FnSignature & 31; 02109 02110 // Save this for later so we know type of lazily instantiated functions. 02111 // Note that known-external functions do not have FunctionInfo blocks, so we 02112 // do not add them to the FunctionSignatureList. 02113 if ((Flags & (1 << 4)) == 0) 02114 FunctionSignatureList.push_back(Func); 02115 02116 // Get the calling convention from the low bits. 02117 unsigned CC = Flags & 15; 02118 unsigned Alignment = 0; 02119 if (FnSignature & (1 << 31)) { // Has extension word? 02120 unsigned ExtWord = read_vbr_uint(); 02121 Alignment = (1 << (ExtWord & 31)) >> 1; 02122 CC |= ((ExtWord >> 5) & 15) << 4; 02123 02124 if (ExtWord & (1 << 10)) // Has a section ID. 02125 SectionID[Func] = read_vbr_uint(); 02126 } 02127 02128 Func->setCallingConv(CC-1); 02129 Func->setAlignment(Alignment); 02130 02131 if (Handler) Handler->handleFunctionDeclaration(Func); 02132 02133 // Get the next function signature. 02134 FnSignature = read_vbr_uint(); 02135 if (hasNoFlagsForFunctions) 02136 FnSignature = (FnSignature << 5) + 1; 02137 } 02138 02139 // Now that the function signature list is set up, reverse it so that we can 02140 // remove elements efficiently from the back of the vector. 02141 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end()); 02142 02143 /// SectionNames - This contains the list of section names encoded in the 02144 /// moduleinfoblock. Functions and globals with an explicit section index 02145 /// into this to get their section name. 02146 std::vector<std::string> SectionNames; 02147 02148 if (hasInconsistentModuleGlobalInfo) { 02149 align32(); 02150 } else if (!hasNoDependentLibraries) { 02151 // If this bytecode format has dependent library information in it, read in 02152 // the number of dependent library items that follow. 02153 unsigned num_dep_libs = read_vbr_uint(); 02154 std::string dep_lib; 02155 while (num_dep_libs--) { 02156 dep_lib = read_str(); 02157 TheModule->addLibrary(dep_lib); 02158 if (Handler) 02159 Handler->handleDependentLibrary(dep_lib); 02160 } 02161 02162 // Read target triple and place into the module. 02163 std::string triple = read_str(); 02164 TheModule->setTargetTriple(triple); 02165 if (Handler) 02166 Handler->handleTargetTriple(triple); 02167 02168 if (!hasAlignment && At != BlockEnd) { 02169 // If the file has section info in it, read the section names now. 02170 unsigned NumSections = read_vbr_uint(); 02171 while (NumSections--) 02172 SectionNames.push_back(read_str()); 02173 } 02174 02175 // If the file has module-level inline asm, read it now. 02176 if (!hasAlignment && At != BlockEnd) 02177 TheModule->setModuleInlineAsm(read_str()); 02178 } 02179 02180 // If any globals are in specified sections, assign them now. 02181 for (std::map<GlobalValue*, unsigned>::iterator I = SectionID.begin(), E = 02182 SectionID.end(); I != E; ++I) 02183 if (I->second) { 02184 if (I->second > SectionID.size()) 02185 error("SectionID out of range for global!"); 02186 I->first->setSection(SectionNames[I->second-1]); 02187 } 02188 02189 // This is for future proofing... in the future extra fields may be added that 02190 // we don't understand, so we transparently ignore them. 02191 // 02192 At = BlockEnd; 02193 02194 if (Handler) Handler->handleModuleGlobalsEnd(); 02195 } 02196 02197 /// Parse the version information and decode it by setting flags on the 02198 /// Reader that enable backward compatibility of the reader. 02199 void BytecodeReader::ParseVersionInfo() { 02200 unsigned Version = read_vbr_uint(); 02201 02202 // Unpack version number: low four bits are for flags, top bits = version 02203 Module::Endianness Endianness; 02204 Module::PointerSize PointerSize; 02205 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian; 02206 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32; 02207 02208 bool hasNoEndianness = Version & 4; 02209 bool hasNoPointerSize = Version & 8; 02210 02211 RevisionNum = Version >> 4; 02212 02213 // Default values for the current bytecode version 02214 hasInconsistentModuleGlobalInfo = false; 02215 hasExplicitPrimitiveZeros = false; 02216 hasRestrictedGEPTypes = false; 02217 hasTypeDerivedFromValue = false; 02218 hasLongBlockHeaders = false; 02219 has32BitTypes = false; 02220 hasNoDependentLibraries = false; 02221 hasAlignment = false; 02222 hasNoUndefValue = false; 02223 hasNoFlagsForFunctions = false; 02224 hasNoUnreachableInst = false; 02225 02226 switch (RevisionNum) { 02227 case 0: // LLVM 1.0, 1.1 (Released) 02228 // Base LLVM 1.0 bytecode format. 02229 hasInconsistentModuleGlobalInfo = true; 02230 hasExplicitPrimitiveZeros = true; 02231 02232 // FALL THROUGH 02233 02234 case 1: // LLVM 1.2 (Released) 02235 // LLVM 1.2 added explicit support for emitting strings efficiently. 02236 02237 // Also, it fixed the problem where the size of the ModuleGlobalInfo block 02238 // included the size for the alignment at the end, where the rest of the 02239 // blocks did not. 02240 02241 // LLVM 1.2 and before required that GEP indices be ubyte constants for 02242 // structures and longs for sequential types. 02243 hasRestrictedGEPTypes = true; 02244 02245 // LLVM 1.2 and before had the Type class derive from Value class. This 02246 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are 02247 // written differently because Types can no longer be part of the 02248 // type planes for Values. 02249 hasTypeDerivedFromValue = true; 02250 02251 // FALL THROUGH 02252 02253 case 2: // 1.2.5 (Not Released) 02254 02255 // LLVM 1.2 and earlier had two-word block headers. This is a bit wasteful, 02256 // especially for small files where the 8 bytes per block is a large 02257 // fraction of the total block size. In LLVM 1.3, the block type and length 02258 // are compressed into a single 32-bit unsigned integer. 27 bits for length, 02259 // 5 bits for block type. 02260 hasLongBlockHeaders = true; 02261 02262 // LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3 02263 // this has been reduced to vbr_uint24. It shouldn't make much difference 02264 // since we haven't run into a module with > 24 million types, but for 02265 // safety the 24-bit restriction has been enforced in 1.3 to free some bits 02266 // in various places and to ensure consistency. 02267 has32BitTypes = true; 02268 02269 // LLVM 1.2 and earlier did not provide a target triple nor a list of 02270 // libraries on which the bytecode is dependent. LLVM 1.3 provides these 02271 // features, for use in future versions of LLVM. 02272 hasNoDependentLibraries = true; 02273 02274 // FALL THROUGH 02275 02276 case 3: // LLVM 1.3 (Released) 02277 // LLVM 1.3 and earlier caused alignment bytes to be written on some block 02278 // boundaries and at the end of some strings. In extreme cases (e.g. lots 02279 // of GEP references to a constant array), this can increase the file size 02280 // by 30% or more. In version 1.4 alignment is done away with completely. 02281 hasAlignment = true; 02282 02283 // FALL THROUGH 02284 02285 case 4: // 1.3.1 (Not Released) 02286 // In version 4, we did not support the 'undef' constant. 02287 hasNoUndefValue = true; 02288 02289 // In version 4 and above, we did not include space for flags for functions 02290 // in the module info block. 02291 hasNoFlagsForFunctions = true; 02292 02293 // In version 4 and above, we did not include the 'unreachable' instruction 02294 // in the opcode numbering in the bytecode file. 02295 hasNoUnreachableInst = true; 02296 break; 02297 02298 // FALL THROUGH 02299 02300 case 5: // 1.4 (Released) 02301 break; 02302 02303 default: 02304 error("Unknown bytecode version number: " + itostr(RevisionNum)); 02305 } 02306 02307 if (hasNoEndianness) Endianness = Module::AnyEndianness; 02308 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize; 02309 02310 TheModule->setEndianness(Endianness); 02311 TheModule->setPointerSize(PointerSize); 02312 02313 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize); 02314 } 02315 02316 /// Parse a whole module. 02317 void BytecodeReader::ParseModule() { 02318 unsigned Type, Size; 02319 02320 FunctionSignatureList.clear(); // Just in case... 02321 02322 // Read into instance variables... 02323 ParseVersionInfo(); 02324 align32(); 02325 02326 bool SeenModuleGlobalInfo = false; 02327 bool SeenGlobalTypePlane = false; 02328 BufPtr MyEnd = BlockEnd; 02329 while (At < MyEnd) { 02330 BufPtr OldAt = At; 02331 read_block(Type, Size); 02332 02333 switch (Type) { 02334 02335 case BytecodeFormat::GlobalTypePlaneBlockID: 02336 if (SeenGlobalTypePlane) 02337 error("Two GlobalTypePlane Blocks Encountered!"); 02338 02339 if (Size > 0) 02340 ParseGlobalTypes(); 02341 SeenGlobalTypePlane = true; 02342 break; 02343 02344 case BytecodeFormat::ModuleGlobalInfoBlockID: 02345 if (SeenModuleGlobalInfo) 02346 error("Two ModuleGlobalInfo Blocks Encountered!"); 02347 ParseModuleGlobalInfo(); 02348 SeenModuleGlobalInfo = true; 02349 break; 02350 02351 case BytecodeFormat::ConstantPoolBlockID: 02352 ParseConstantPool(ModuleValues, ModuleTypes,false); 02353 break; 02354 02355 case BytecodeFormat::FunctionBlockID: 02356 ParseFunctionLazily(); 02357 break; 02358 02359 case BytecodeFormat::SymbolTableBlockID: 02360 ParseSymbolTable(0, &TheModule->getSymbolTable()); 02361 break; 02362 02363 default: 02364 At += Size; 02365 if (OldAt > At) { 02366 error("Unexpected Block of Type #" + utostr(Type) + " encountered!"); 02367 } 02368 break; 02369 } 02370 BlockEnd = MyEnd; 02371 align32(); 02372 } 02373 02374 // After the module constant pool has been read, we can safely initialize 02375 // global variables... 02376 while (!GlobalInits.empty()) { 02377 GlobalVariable *GV = GlobalInits.back().first; 02378 unsigned Slot = GlobalInits.back().second; 02379 GlobalInits.pop_back(); 02380 02381 // Look up the initializer value... 02382 // FIXME: Preserve this type ID! 02383 02384 const llvm::PointerType* GVType = GV->getType(); 02385 unsigned TypeSlot = getTypeSlot(GVType->getElementType()); 02386 if (Constant *CV = getConstantValue(TypeSlot, Slot)) { 02387 if (GV->hasInitializer()) 02388 error("Global *already* has an initializer?!"); 02389 if (Handler) Handler->handleGlobalInitializer(GV,CV); 02390 GV->setInitializer(CV); 02391 } else 02392 error("Cannot find initializer value."); 02393 } 02394 02395 if (!ConstantFwdRefs.empty()) 02396 error("Use of undefined constants in a module"); 02397 02398 /// Make sure we pulled them all out. If we didn't then there's a declaration 02399 /// but a missing body. That's not allowed. 02400 if (!FunctionSignatureList.empty()) 02401 error("Function declared, but bytecode stream ended before definition"); 02402 } 02403 02404 /// This function completely parses a bytecode buffer given by the \p Buf 02405 /// and \p Length parameters. 02406 void BytecodeReader::ParseBytecode(BufPtr Buf, unsigned Length, 02407 const std::string &ModuleID) { 02408 02409 try { 02410 RevisionNum = 0; 02411 At = MemStart = BlockStart = Buf; 02412 MemEnd = BlockEnd = Buf + Length; 02413 02414 // Create the module 02415 TheModule = new Module(ModuleID); 02416 02417 if (Handler) Handler->handleStart(TheModule, Length); 02418 02419 // Read the four bytes of the signature. 02420 unsigned Sig = read_uint(); 02421 02422 // If this is a compressed file 02423 if (Sig == ('l' | ('l' << 8) | ('v' << 16) | ('c' << 24))) { 02424 02425 // Invoke the decompression of the bytecode. Note that we have to skip the 02426 // file's magic number which is not part of the compressed block. Hence, 02427 // the Buf+4 and Length-4. The result goes into decompressedBlock, a data 02428 // member for retention until BytecodeReader is destructed. 02429 unsigned decompressedLength = Compressor::decompressToNewBuffer( 02430 (char*)Buf+4,Length-4,decompressedBlock); 02431 02432 // We must adjust the buffer pointers used by the bytecode reader to point 02433 // into the new decompressed block. After decompression, the 02434 // decompressedBlock will point to a contiguous memory area that has 02435 // the decompressed data. 02436 At = MemStart = BlockStart = Buf = (BufPtr) decompressedBlock; 02437 MemEnd = BlockEnd = Buf + decompressedLength; 02438 02439 // else if this isn't a regular (uncompressed) bytecode file, then its 02440 // and error, generate that now. 02441 } else if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) { 02442 error("Invalid bytecode signature: " + utohexstr(Sig)); 02443 } 02444 02445 // Tell the handler we're starting a module 02446 if (Handler) Handler->handleModuleBegin(ModuleID); 02447 02448 // Get the module block and size and verify. This is handled specially 02449 // because the module block/size is always written in long format. Other 02450 // blocks are written in short format so the read_block method is used. 02451 unsigned Type, Size; 02452 Type = read_uint(); 02453 Size = read_uint(); 02454 if (Type != BytecodeFormat::ModuleBlockID) { 02455 error("Expected Module Block! Type:" + utostr(Type) + ", Size:" 02456 + utostr(Size)); 02457 } 02458 02459 // It looks like the darwin ranlib program is broken, and adds trailing 02460 // garbage to the end of some bytecode files. This hack allows the bc 02461 // reader to ignore trailing garbage on bytecode files. 02462 if (At + Size < MemEnd) 02463 MemEnd = BlockEnd = At+Size; 02464 02465 if (At + Size != MemEnd) 02466 error("Invalid Top Level Block Length! Type:" + utostr(Type) 02467 + ", Size:" + utostr(Size)); 02468 02469 // Parse the module contents 02470 this->ParseModule(); 02471 02472 // Check for missing functions 02473 if (hasFunctions()) 02474 error("Function expected, but bytecode stream ended!"); 02475 02476 // Look for intrinsic functions to upgrade, upgrade them, and save the 02477 // mapping from old function to new for use later when instructions are 02478 // converted. 02479 for (Module::iterator FI = TheModule->begin(), FE = TheModule->end(); 02480 FI != FE; ++FI) 02481 if (Function* newF = UpgradeIntrinsicFunction(FI)) { 02482 upgradedFunctions.insert(std::make_pair(FI, newF)); 02483 FI->setName(""); 02484 } 02485 02486 // Tell the handler we're done with the module 02487 if (Handler) 02488 Handler->handleModuleEnd(ModuleID); 02489 02490 // Tell the handler we're finished the parse 02491 if (Handler) Handler->handleFinish(); 02492 02493 } catch (std::string& errstr) { 02494 if (Handler) Handler->handleError(errstr); 02495 freeState(); 02496 delete TheModule; 02497 TheModule = 0; 02498 if (decompressedBlock != 0 ) { 02499 ::free(decompressedBlock); 02500 decompressedBlock = 0; 02501 } 02502 throw; 02503 } catch (...) { 02504 std::string msg("Unknown Exception Occurred"); 02505 if (Handler) Handler->handleError(msg); 02506 freeState(); 02507 delete TheModule; 02508 TheModule = 0; 02509 if (decompressedBlock != 0) { 02510 ::free(decompressedBlock); 02511 decompressedBlock = 0; 02512 } 02513 throw msg; 02514 } 02515 } 02516 02517 //===----------------------------------------------------------------------===// 02518 //=== Default Implementations of Handler Methods 02519 //===----------------------------------------------------------------------===// 02520 02521 BytecodeHandler::~BytecodeHandler() {} 02522