1 //===- Reader.cpp - Code to read bytecode files ---------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This library implements the functionality defined in llvm/Bytecode/Reader.h
12 // Note that this library should be as fast as possible, reentrant, and
15 // TODO: Allow passing in an option to ignore the symbol table
17 //===----------------------------------------------------------------------===//
20 #include "llvm/Assembly/AutoUpgrade.h"
21 #include "llvm/Bytecode/BytecodeHandler.h"
22 #include "llvm/BasicBlock.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/SymbolTable.h"
27 #include "llvm/Bytecode/Format.h"
28 #include "llvm/Config/alloca.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/Compressor.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/ADT/StringExtras.h"
38 /// @brief A class for maintaining the slot number definition
39 /// as a placeholder for the actual definition for forward constants defs.
40 class ConstantPlaceHolder : public ConstantExpr {
41 ConstantPlaceHolder(); // DO NOT IMPLEMENT
42 void operator=(const ConstantPlaceHolder &); // DO NOT IMPLEMENT
45 ConstantPlaceHolder(const Type *Ty)
46 : ConstantExpr(Ty, Instruction::UserOp1, &Op, 1),
47 Op(UndefValue::get(Type::IntTy), this) {
52 // Provide some details on error
53 inline void BytecodeReader::error(std::string err) {
55 err += itostr(RevisionNum) ;
57 err += itostr(At-MemStart);
62 //===----------------------------------------------------------------------===//
63 // Bytecode Reading Methods
64 //===----------------------------------------------------------------------===//
66 /// Determine if the current block being read contains any more data.
67 inline bool BytecodeReader::moreInBlock() {
71 /// Throw an error if we've read past the end of the current block
72 inline void BytecodeReader::checkPastBlockEnd(const char * block_name) {
74 error(std::string("Attempt to read past the end of ") + block_name +
78 /// Align the buffer position to a 32 bit boundary
79 inline void BytecodeReader::align32() {
82 At = (const unsigned char *)((unsigned long)(At+3) & (~3UL));
84 if (Handler) Handler->handleAlignment(At - Save);
86 error("Ran out of data while aligning!");
90 /// Read a whole unsigned integer
91 inline unsigned BytecodeReader::read_uint() {
93 error("Ran out of data reading uint!");
95 return At[-4] | (At[-3] << 8) | (At[-2] << 16) | (At[-1] << 24);
98 /// Read a variable-bit-rate encoded unsigned integer
99 inline unsigned BytecodeReader::read_vbr_uint() {
106 error("Ran out of data reading vbr_uint!");
107 Result |= (unsigned)((*At++) & 0x7F) << Shift;
109 } while (At[-1] & 0x80);
110 if (Handler) Handler->handleVBR32(At-Save);
114 /// Read a variable-bit-rate encoded unsigned 64-bit integer.
115 inline uint64_t BytecodeReader::read_vbr_uint64() {
122 error("Ran out of data reading vbr_uint64!");
123 Result |= (uint64_t)((*At++) & 0x7F) << Shift;
125 } while (At[-1] & 0x80);
126 if (Handler) Handler->handleVBR64(At-Save);
130 /// Read a variable-bit-rate encoded signed 64-bit integer.
131 inline int64_t BytecodeReader::read_vbr_int64() {
132 uint64_t R = read_vbr_uint64();
135 return -(int64_t)(R >> 1);
136 else // There is no such thing as -0 with integers. "-0" really means
137 // 0x8000000000000000.
140 return (int64_t)(R >> 1);
143 /// Read a pascal-style string (length followed by text)
144 inline std::string BytecodeReader::read_str() {
145 unsigned Size = read_vbr_uint();
146 const unsigned char *OldAt = At;
148 if (At > BlockEnd) // Size invalid?
149 error("Ran out of data reading a string!");
150 return std::string((char*)OldAt, Size);
153 /// Read an arbitrary block of data
154 inline void BytecodeReader::read_data(void *Ptr, void *End) {
155 unsigned char *Start = (unsigned char *)Ptr;
156 unsigned Amount = (unsigned char *)End - Start;
157 if (At+Amount > BlockEnd)
158 error("Ran out of data!");
159 std::copy(At, At+Amount, Start);
163 /// Read a float value in little-endian order
164 inline void BytecodeReader::read_float(float& FloatVal) {
165 /// FIXME: This isn't optimal, it has size problems on some platforms
166 /// where FP is not IEEE.
167 FloatVal = BitsToFloat(At[0] | (At[1] << 8) | (At[2] << 16) | (At[3] << 24));
168 At+=sizeof(uint32_t);
171 /// Read a double value in little-endian order
172 inline void BytecodeReader::read_double(double& DoubleVal) {
173 /// FIXME: This isn't optimal, it has size problems on some platforms
174 /// where FP is not IEEE.
175 DoubleVal = BitsToDouble((uint64_t(At[0]) << 0) | (uint64_t(At[1]) << 8) |
176 (uint64_t(At[2]) << 16) | (uint64_t(At[3]) << 24) |
177 (uint64_t(At[4]) << 32) | (uint64_t(At[5]) << 40) |
178 (uint64_t(At[6]) << 48) | (uint64_t(At[7]) << 56));
179 At+=sizeof(uint64_t);
182 /// Read a block header and obtain its type and size
183 inline void BytecodeReader::read_block(unsigned &Type, unsigned &Size) {
184 if ( hasLongBlockHeaders ) {
188 case BytecodeFormat::Reserved_DoNotUse :
189 error("Reserved_DoNotUse used as Module Type?");
190 Type = BytecodeFormat::ModuleBlockID; break;
191 case BytecodeFormat::Module:
192 Type = BytecodeFormat::ModuleBlockID; break;
193 case BytecodeFormat::Function:
194 Type = BytecodeFormat::FunctionBlockID; break;
195 case BytecodeFormat::ConstantPool:
196 Type = BytecodeFormat::ConstantPoolBlockID; break;
197 case BytecodeFormat::SymbolTable:
198 Type = BytecodeFormat::SymbolTableBlockID; break;
199 case BytecodeFormat::ModuleGlobalInfo:
200 Type = BytecodeFormat::ModuleGlobalInfoBlockID; break;
201 case BytecodeFormat::GlobalTypePlane:
202 Type = BytecodeFormat::GlobalTypePlaneBlockID; break;
203 case BytecodeFormat::InstructionList:
204 Type = BytecodeFormat::InstructionListBlockID; break;
205 case BytecodeFormat::CompactionTable:
206 Type = BytecodeFormat::CompactionTableBlockID; break;
207 case BytecodeFormat::BasicBlock:
208 /// This block type isn't used after version 1.1. However, we have to
209 /// still allow the value in case this is an old bc format file.
210 /// We just let its value creep thru.
213 error("Invalid block id found: " + utostr(Type));
218 Type = Size & 0x1F; // mask low order five bits
219 Size >>= 5; // get rid of five low order bits, leaving high 27
222 if (At + Size > BlockEnd)
223 error("Attempt to size a block past end of memory");
224 BlockEnd = At + Size;
225 if (Handler) Handler->handleBlock(Type, BlockStart, Size);
229 /// In LLVM 1.2 and before, Types were derived from Value and so they were
230 /// written as part of the type planes along with any other Value. In LLVM
231 /// 1.3 this changed so that Type does not derive from Value. Consequently,
232 /// the BytecodeReader's containers for Values can't contain Types because
233 /// there's no inheritance relationship. This means that the "Type Type"
234 /// plane is defunct along with the Type::TypeTyID TypeID. In LLVM 1.3
235 /// whenever a bytecode construct must have both types and values together,
236 /// the types are always read/written first and then the Values. Furthermore
237 /// since Type::TypeTyID no longer exists, its value (12) now corresponds to
238 /// Type::LabelTyID. In order to overcome this we must "sanitize" all the
239 /// type TypeIDs we encounter. For LLVM 1.3 bytecode files, there's no change.
240 /// For LLVM 1.2 and before, this function will decrement the type id by
241 /// one to account for the missing Type::TypeTyID enumerator if the value is
242 /// larger than 12 (Type::LabelTyID). If the value is exactly 12, then this
243 /// function returns true, otherwise false. This helps detect situations
244 /// where the pre 1.3 bytecode is indicating that what follows is a type.
245 /// @returns true iff type id corresponds to pre 1.3 "type type"
246 inline bool BytecodeReader::sanitizeTypeId(unsigned &TypeId) {
247 if (hasTypeDerivedFromValue) { /// do nothing if 1.3 or later
248 if (TypeId == Type::LabelTyID) {
249 TypeId = Type::VoidTyID; // sanitize it
250 return true; // indicate we got TypeTyID in pre 1.3 bytecode
251 } else if (TypeId > Type::LabelTyID)
252 --TypeId; // shift all planes down because type type plane is missing
257 /// Reads a vbr uint to read in a type id and does the necessary
258 /// conversion on it by calling sanitizeTypeId.
259 /// @returns true iff \p TypeId read corresponds to a pre 1.3 "type type"
260 /// @see sanitizeTypeId
261 inline bool BytecodeReader::read_typeid(unsigned &TypeId) {
262 TypeId = read_vbr_uint();
263 if ( !has32BitTypes )
264 if ( TypeId == 0x00FFFFFF )
265 TypeId = read_vbr_uint();
266 return sanitizeTypeId(TypeId);
269 //===----------------------------------------------------------------------===//
271 //===----------------------------------------------------------------------===//
273 /// Determine if a type id has an implicit null value
274 inline bool BytecodeReader::hasImplicitNull(unsigned TyID) {
275 if (!hasExplicitPrimitiveZeros)
276 return TyID != Type::LabelTyID && TyID != Type::VoidTyID;
277 return TyID >= Type::FirstDerivedTyID;
280 /// Obtain a type given a typeid and account for things like compaction tables,
281 /// function level vs module level, and the offsetting for the primitive types.
282 const Type *BytecodeReader::getType(unsigned ID) {
283 if (ID < Type::FirstDerivedTyID)
284 if (const Type *T = Type::getPrimitiveType((Type::TypeID)ID))
285 return T; // Asked for a primitive type...
287 // Otherwise, derived types need offset...
288 ID -= Type::FirstDerivedTyID;
290 if (!CompactionTypes.empty()) {
291 if (ID >= CompactionTypes.size())
292 error("Type ID out of range for compaction table!");
293 return CompactionTypes[ID].first;
296 // Is it a module-level type?
297 if (ID < ModuleTypes.size())
298 return ModuleTypes[ID].get();
300 // Nope, is it a function-level type?
301 ID -= ModuleTypes.size();
302 if (ID < FunctionTypes.size())
303 return FunctionTypes[ID].get();
305 error("Illegal type reference!");
309 /// Get a sanitized type id. This just makes sure that the \p ID
310 /// is both sanitized and not the "type type" of pre-1.3 bytecode.
311 /// @see sanitizeTypeId
312 inline const Type* BytecodeReader::getSanitizedType(unsigned& ID) {
313 if (sanitizeTypeId(ID))
314 error("Invalid type id encountered");
318 /// This method just saves some coding. It uses read_typeid to read
319 /// in a sanitized type id, errors that its not the type type, and
320 /// then calls getType to return the type value.
321 inline const Type* BytecodeReader::readSanitizedType() {
324 error("Invalid type id encountered");
328 /// Get the slot number associated with a type accounting for primitive
329 /// types, compaction tables, and function level vs module level.
330 unsigned BytecodeReader::getTypeSlot(const Type *Ty) {
331 if (Ty->isPrimitiveType())
332 return Ty->getTypeID();
334 // Scan the compaction table for the type if needed.
335 if (!CompactionTypes.empty()) {
336 for (unsigned i = 0, e = CompactionTypes.size(); i != e; ++i)
337 if (CompactionTypes[i].first == Ty)
338 return Type::FirstDerivedTyID + i;
340 error("Couldn't find type specified in compaction table!");
343 // Check the function level types first...
344 TypeListTy::iterator I = std::find(FunctionTypes.begin(),
345 FunctionTypes.end(), Ty);
347 if (I != FunctionTypes.end())
348 return Type::FirstDerivedTyID + ModuleTypes.size() +
349 (&*I - &FunctionTypes[0]);
351 // If we don't have our cache yet, build it now.
352 if (ModuleTypeIDCache.empty()) {
354 ModuleTypeIDCache.reserve(ModuleTypes.size());
355 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end();
357 ModuleTypeIDCache.push_back(std::make_pair(*I, N));
359 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end());
362 // Binary search the cache for the entry.
363 std::vector<std::pair<const Type*, unsigned> >::iterator IT =
364 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(),
365 std::make_pair(Ty, 0U));
366 if (IT == ModuleTypeIDCache.end() || IT->first != Ty)
367 error("Didn't find type in ModuleTypes.");
369 return Type::FirstDerivedTyID + IT->second;
372 /// This is just like getType, but when a compaction table is in use, it is
373 /// ignored. It also ignores function level types.
375 const Type *BytecodeReader::getGlobalTableType(unsigned Slot) {
376 if (Slot < Type::FirstDerivedTyID) {
377 const Type *Ty = Type::getPrimitiveType((Type::TypeID)Slot);
379 error("Not a primitive type ID?");
382 Slot -= Type::FirstDerivedTyID;
383 if (Slot >= ModuleTypes.size())
384 error("Illegal compaction table type reference!");
385 return ModuleTypes[Slot];
388 /// This is just like getTypeSlot, but when a compaction table is in use, it
389 /// is ignored. It also ignores function level types.
390 unsigned BytecodeReader::getGlobalTableTypeSlot(const Type *Ty) {
391 if (Ty->isPrimitiveType())
392 return Ty->getTypeID();
394 // If we don't have our cache yet, build it now.
395 if (ModuleTypeIDCache.empty()) {
397 ModuleTypeIDCache.reserve(ModuleTypes.size());
398 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end();
400 ModuleTypeIDCache.push_back(std::make_pair(*I, N));
402 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end());
405 // Binary search the cache for the entry.
406 std::vector<std::pair<const Type*, unsigned> >::iterator IT =
407 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(),
408 std::make_pair(Ty, 0U));
409 if (IT == ModuleTypeIDCache.end() || IT->first != Ty)
410 error("Didn't find type in ModuleTypes.");
412 return Type::FirstDerivedTyID + IT->second;
415 /// Retrieve a value of a given type and slot number, possibly creating
416 /// it if it doesn't already exist.
417 Value * BytecodeReader::getValue(unsigned type, unsigned oNum, bool Create) {
418 assert(type != Type::LabelTyID && "getValue() cannot get blocks!");
421 // If there is a compaction table active, it defines the low-level numbers.
422 // If not, the module values define the low-level numbers.
423 if (CompactionValues.size() > type && !CompactionValues[type].empty()) {
424 if (Num < CompactionValues[type].size())
425 return CompactionValues[type][Num];
426 Num -= CompactionValues[type].size();
428 // By default, the global type id is the type id passed in
429 unsigned GlobalTyID = type;
431 // If the type plane was compactified, figure out the global type ID by
432 // adding the derived type ids and the distance.
433 if (!CompactionTypes.empty() && type >= Type::FirstDerivedTyID)
434 GlobalTyID = CompactionTypes[type-Type::FirstDerivedTyID].second;
436 if (hasImplicitNull(GlobalTyID)) {
437 const Type *Ty = getType(type);
438 if (!isa<OpaqueType>(Ty)) {
440 return Constant::getNullValue(Ty);
445 if (GlobalTyID < ModuleValues.size() && ModuleValues[GlobalTyID]) {
446 if (Num < ModuleValues[GlobalTyID]->size())
447 return ModuleValues[GlobalTyID]->getOperand(Num);
448 Num -= ModuleValues[GlobalTyID]->size();
452 if (FunctionValues.size() > type &&
453 FunctionValues[type] &&
454 Num < FunctionValues[type]->size())
455 return FunctionValues[type]->getOperand(Num);
457 if (!Create) return 0; // Do not create a placeholder?
459 // Did we already create a place holder?
460 std::pair<unsigned,unsigned> KeyValue(type, oNum);
461 ForwardReferenceMap::iterator I = ForwardReferences.lower_bound(KeyValue);
462 if (I != ForwardReferences.end() && I->first == KeyValue)
463 return I->second; // We have already created this placeholder
465 // If the type exists (it should)
466 if (const Type* Ty = getType(type)) {
467 // Create the place holder
468 Value *Val = new Argument(Ty);
469 ForwardReferences.insert(I, std::make_pair(KeyValue, Val));
472 throw "Can't create placeholder for value of type slot #" + utostr(type);
475 /// This is just like getValue, but when a compaction table is in use, it
476 /// is ignored. Also, no forward references or other fancy features are
478 Value* BytecodeReader::getGlobalTableValue(unsigned TyID, unsigned SlotNo) {
480 return Constant::getNullValue(getType(TyID));
482 if (!CompactionTypes.empty() && TyID >= Type::FirstDerivedTyID) {
483 TyID -= Type::FirstDerivedTyID;
484 if (TyID >= CompactionTypes.size())
485 error("Type ID out of range for compaction table!");
486 TyID = CompactionTypes[TyID].second;
491 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0 ||
492 SlotNo >= ModuleValues[TyID]->size()) {
493 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0)
494 error("Corrupt compaction table entry!"
495 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
496 + utostr(ModuleValues.size()));
498 error("Corrupt compaction table entry!"
499 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
500 + utostr(ModuleValues.size()) + ", "
501 + utohexstr(reinterpret_cast<uint64_t>(((void*)ModuleValues[TyID])))
503 + utostr(ModuleValues[TyID]->size()));
505 return ModuleValues[TyID]->getOperand(SlotNo);
508 /// Just like getValue, except that it returns a null pointer
509 /// only on error. It always returns a constant (meaning that if the value is
510 /// defined, but is not a constant, that is an error). If the specified
511 /// constant hasn't been parsed yet, a placeholder is defined and used.
512 /// Later, after the real value is parsed, the placeholder is eliminated.
513 Constant* BytecodeReader::getConstantValue(unsigned TypeSlot, unsigned Slot) {
514 if (Value *V = getValue(TypeSlot, Slot, false))
515 if (Constant *C = dyn_cast<Constant>(V))
516 return C; // If we already have the value parsed, just return it
518 error("Value for slot " + utostr(Slot) +
519 " is expected to be a constant!");
521 std::pair<unsigned, unsigned> Key(TypeSlot, Slot);
522 ConstantRefsType::iterator I = ConstantFwdRefs.lower_bound(Key);
524 if (I != ConstantFwdRefs.end() && I->first == Key) {
527 // Create a placeholder for the constant reference and
528 // keep track of the fact that we have a forward ref to recycle it
529 Constant *C = new ConstantPlaceHolder(getType(TypeSlot));
531 // Keep track of the fact that we have a forward ref to recycle it
532 ConstantFwdRefs.insert(I, std::make_pair(Key, C));
537 //===----------------------------------------------------------------------===//
538 // IR Construction Methods
539 //===----------------------------------------------------------------------===//
541 /// As values are created, they are inserted into the appropriate place
542 /// with this method. The ValueTable argument must be one of ModuleValues
543 /// or FunctionValues data members of this class.
544 unsigned BytecodeReader::insertValue(Value *Val, unsigned type,
545 ValueTable &ValueTab) {
546 assert((!isa<Constant>(Val) || !cast<Constant>(Val)->isNullValue()) ||
547 !hasImplicitNull(type) &&
548 "Cannot read null values from bytecode!");
550 if (ValueTab.size() <= type)
551 ValueTab.resize(type+1);
553 if (!ValueTab[type]) ValueTab[type] = new ValueList();
555 ValueTab[type]->push_back(Val);
557 bool HasOffset = hasImplicitNull(type) && !isa<OpaqueType>(Val->getType());
558 return ValueTab[type]->size()-1 + HasOffset;
561 /// Insert the arguments of a function as new values in the reader.
562 void BytecodeReader::insertArguments(Function* F) {
563 const FunctionType *FT = F->getFunctionType();
564 Function::arg_iterator AI = F->arg_begin();
565 for (FunctionType::param_iterator It = FT->param_begin();
566 It != FT->param_end(); ++It, ++AI)
567 insertValue(AI, getTypeSlot(AI->getType()), FunctionValues);
570 //===----------------------------------------------------------------------===//
571 // Bytecode Parsing Methods
572 //===----------------------------------------------------------------------===//
574 /// This method parses a single instruction. The instruction is
575 /// inserted at the end of the \p BB provided. The arguments of
576 /// the instruction are provided in the \p Oprnds vector.
577 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds,
581 // Clear instruction data
585 unsigned Op = read_uint();
587 // bits Instruction format: Common to all formats
588 // --------------------------
589 // 01-00: Opcode type, fixed to 1.
591 Opcode = (Op >> 2) & 63;
592 Oprnds.resize((Op >> 0) & 03);
594 // Extract the operands
595 switch (Oprnds.size()) {
597 // bits Instruction format:
598 // --------------------------
599 // 19-08: Resulting type plane
600 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
602 iType = (Op >> 8) & 4095;
603 Oprnds[0] = (Op >> 20) & 4095;
604 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands...
608 // bits Instruction format:
609 // --------------------------
610 // 15-08: Resulting type plane
614 iType = (Op >> 8) & 255;
615 Oprnds[0] = (Op >> 16) & 255;
616 Oprnds[1] = (Op >> 24) & 255;
619 // bits Instruction format:
620 // --------------------------
621 // 13-08: Resulting type plane
626 iType = (Op >> 8) & 63;
627 Oprnds[0] = (Op >> 14) & 63;
628 Oprnds[1] = (Op >> 20) & 63;
629 Oprnds[2] = (Op >> 26) & 63;
632 At -= 4; // Hrm, try this again...
633 Opcode = read_vbr_uint();
635 iType = read_vbr_uint();
637 unsigned NumOprnds = read_vbr_uint();
638 Oprnds.resize(NumOprnds);
641 error("Zero-argument instruction found; this is invalid.");
643 for (unsigned i = 0; i != NumOprnds; ++i)
644 Oprnds[i] = read_vbr_uint();
649 const Type *InstTy = getSanitizedType(iType);
651 // We have enough info to inform the handler now.
652 if (Handler) Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt);
654 // Declare the resulting instruction we'll build.
655 Instruction *Result = 0;
657 // If this is a bytecode format that did not include the unreachable
658 // instruction, bump up all opcodes numbers to make space.
659 if (hasNoUnreachableInst) {
660 if (Opcode >= Instruction::Unreachable &&
666 // Handle binary operators
667 if (Opcode >= Instruction::BinaryOpsBegin &&
668 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2)
669 Result = BinaryOperator::create((Instruction::BinaryOps)Opcode,
670 getValue(iType, Oprnds[0]),
671 getValue(iType, Oprnds[1]));
677 error("Illegal instruction read!");
679 case Instruction::VAArg:
680 Result = new VAArgInst(getValue(iType, Oprnds[0]),
681 getSanitizedType(Oprnds[1]));
683 case 32: { //VANext_old
684 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
685 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy,
689 //foo = alloca 1 of t
694 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vanext.fix");
695 BB->getInstList().push_back(foo);
696 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
697 BB->getInstList().push_back(bar);
698 BB->getInstList().push_back(new StoreInst(bar, foo));
699 Instruction* tmp = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
700 BB->getInstList().push_back(tmp);
701 Result = new LoadInst(foo);
704 case 33: { //VAArg_old
705 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
706 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy,
710 //foo = alloca 1 of t
714 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vaarg.fix");
715 BB->getInstList().push_back(foo);
716 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
717 BB->getInstList().push_back(bar);
718 BB->getInstList().push_back(new StoreInst(bar, foo));
719 Result = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
722 case Instruction::ExtractElement: {
723 if (Oprnds.size() != 2)
724 throw std::string("Invalid extractelement instruction!");
725 Result = new ExtractElementInst(getValue(iType, Oprnds[0]),
726 getValue(Type::UIntTyID, Oprnds[1]));
729 case Instruction::InsertElement: {
730 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy);
731 if (!PackedTy || Oprnds.size() != 3)
732 throw std::string("Invalid insertelement instruction!");
734 new InsertElementInst(getValue(iType, Oprnds[0]),
735 getValue(getTypeSlot(PackedTy->getElementType()),
737 getValue(Type::UIntTyID, Oprnds[2]));
740 case Instruction::Cast:
741 Result = new CastInst(getValue(iType, Oprnds[0]),
742 getSanitizedType(Oprnds[1]));
744 case Instruction::Select:
745 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]),
746 getValue(iType, Oprnds[1]),
747 getValue(iType, Oprnds[2]));
749 case Instruction::PHI: {
750 if (Oprnds.size() == 0 || (Oprnds.size() & 1))
751 error("Invalid phi node encountered!");
753 PHINode *PN = new PHINode(InstTy);
754 PN->reserveOperandSpace(Oprnds.size());
755 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2)
756 PN->addIncoming(getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1]));
761 case Instruction::Shl:
762 case Instruction::Shr:
763 Result = new ShiftInst((Instruction::OtherOps)Opcode,
764 getValue(iType, Oprnds[0]),
765 getValue(Type::UByteTyID, Oprnds[1]));
767 case Instruction::Ret:
768 if (Oprnds.size() == 0)
769 Result = new ReturnInst();
770 else if (Oprnds.size() == 1)
771 Result = new ReturnInst(getValue(iType, Oprnds[0]));
773 error("Unrecognized instruction!");
776 case Instruction::Br:
777 if (Oprnds.size() == 1)
778 Result = new BranchInst(getBasicBlock(Oprnds[0]));
779 else if (Oprnds.size() == 3)
780 Result = new BranchInst(getBasicBlock(Oprnds[0]),
781 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2]));
783 error("Invalid number of operands for a 'br' instruction!");
785 case Instruction::Switch: {
786 if (Oprnds.size() & 1)
787 error("Switch statement with odd number of arguments!");
789 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]),
790 getBasicBlock(Oprnds[1]),
792 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2)
793 I->addCase(cast<ConstantInt>(getValue(iType, Oprnds[i])),
794 getBasicBlock(Oprnds[i+1]));
799 case 58: // Call with extra operand for calling conv
800 case 59: // tail call, Fast CC
801 case 60: // normal call, Fast CC
802 case 61: // tail call, C Calling Conv
803 case Instruction::Call: { // Normal Call, C Calling Convention
804 if (Oprnds.size() == 0)
805 error("Invalid call instruction encountered!");
807 Value *F = getValue(iType, Oprnds[0]);
809 unsigned CallingConv = CallingConv::C;
810 bool isTailCall = false;
812 if (Opcode == 61 || Opcode == 59)
815 // Check to make sure we have a pointer to function type
816 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
817 if (PTy == 0) error("Call to non function pointer value!");
818 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
819 if (FTy == 0) error("Call to non function pointer value!");
821 std::vector<Value *> Params;
822 if (!FTy->isVarArg()) {
823 FunctionType::param_iterator It = FTy->param_begin();
826 isTailCall = Oprnds.back() & 1;
827 CallingConv = Oprnds.back() >> 1;
829 } else if (Opcode == 59 || Opcode == 60)
830 CallingConv = CallingConv::Fast;
832 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
833 if (It == FTy->param_end())
834 error("Invalid call instruction!");
835 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
837 if (It != FTy->param_end())
838 error("Invalid call instruction!");
840 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
842 unsigned FirstVariableOperand;
843 if (Oprnds.size() < FTy->getNumParams())
844 error("Call instruction missing operands!");
846 // Read all of the fixed arguments
847 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
848 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i]));
850 FirstVariableOperand = FTy->getNumParams();
852 if ((Oprnds.size()-FirstVariableOperand) & 1)
853 error("Invalid call instruction!"); // Must be pairs of type/value
855 for (unsigned i = FirstVariableOperand, e = Oprnds.size();
857 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
860 Result = new CallInst(F, Params);
861 if (isTailCall) cast<CallInst>(Result)->setTailCall();
862 if (CallingConv) cast<CallInst>(Result)->setCallingConv(CallingConv);
866 case 56: // Invoke with encoded CC
867 case 57: // Invoke Fast CC
868 case Instruction::Invoke: { // Invoke C CC
869 if (Oprnds.size() < 3)
870 error("Invalid invoke instruction!");
871 Value *F = getValue(iType, Oprnds[0]);
873 // Check to make sure we have a pointer to function type
874 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
876 error("Invoke to non function pointer value!");
877 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
879 error("Invoke to non function pointer value!");
881 std::vector<Value *> Params;
882 BasicBlock *Normal, *Except;
883 unsigned CallingConv = CallingConv::C;
886 CallingConv = CallingConv::Fast;
887 else if (Opcode == 56) {
888 CallingConv = Oprnds.back();
892 if (!FTy->isVarArg()) {
893 Normal = getBasicBlock(Oprnds[1]);
894 Except = getBasicBlock(Oprnds[2]);
896 FunctionType::param_iterator It = FTy->param_begin();
897 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) {
898 if (It == FTy->param_end())
899 error("Invalid invoke instruction!");
900 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
902 if (It != FTy->param_end())
903 error("Invalid invoke instruction!");
905 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
907 Normal = getBasicBlock(Oprnds[0]);
908 Except = getBasicBlock(Oprnds[1]);
910 unsigned FirstVariableArgument = FTy->getNumParams()+2;
911 for (unsigned i = 2; i != FirstVariableArgument; ++i)
912 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)),
915 if (Oprnds.size()-FirstVariableArgument & 1) // Must be type/value pairs
916 error("Invalid invoke instruction!");
918 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2)
919 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
922 Result = new InvokeInst(F, Normal, Except, Params);
923 if (CallingConv) cast<InvokeInst>(Result)->setCallingConv(CallingConv);
926 case Instruction::Malloc: {
928 if (Oprnds.size() == 2)
929 Align = (1 << Oprnds[1]) >> 1;
930 else if (Oprnds.size() > 2)
931 error("Invalid malloc instruction!");
932 if (!isa<PointerType>(InstTy))
933 error("Invalid malloc instruction!");
935 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(),
936 getValue(Type::UIntTyID, Oprnds[0]), Align);
940 case Instruction::Alloca: {
942 if (Oprnds.size() == 2)
943 Align = (1 << Oprnds[1]) >> 1;
944 else if (Oprnds.size() > 2)
945 error("Invalid alloca instruction!");
946 if (!isa<PointerType>(InstTy))
947 error("Invalid alloca instruction!");
949 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(),
950 getValue(Type::UIntTyID, Oprnds[0]), Align);
953 case Instruction::Free:
954 if (!isa<PointerType>(InstTy))
955 error("Invalid free instruction!");
956 Result = new FreeInst(getValue(iType, Oprnds[0]));
958 case Instruction::GetElementPtr: {
959 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy))
960 error("Invalid getelementptr instruction!");
962 std::vector<Value*> Idx;
964 const Type *NextTy = InstTy;
965 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
966 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy);
968 error("Invalid getelementptr instruction!");
970 unsigned ValIdx = Oprnds[i];
972 if (!hasRestrictedGEPTypes) {
973 // Struct indices are always uints, sequential type indices can be any
974 // of the 32 or 64-bit integer types. The actual choice of type is
975 // encoded in the low two bits of the slot number.
976 if (isa<StructType>(TopTy))
977 IdxTy = Type::UIntTyID;
979 switch (ValIdx & 3) {
981 case 0: IdxTy = Type::UIntTyID; break;
982 case 1: IdxTy = Type::IntTyID; break;
983 case 2: IdxTy = Type::ULongTyID; break;
984 case 3: IdxTy = Type::LongTyID; break;
989 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID;
992 Idx.push_back(getValue(IdxTy, ValIdx));
994 // Convert ubyte struct indices into uint struct indices.
995 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes)
996 if (ConstantUInt *C = dyn_cast<ConstantUInt>(Idx.back()))
997 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy);
999 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true);
1002 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx);
1006 case 62: // volatile load
1007 case Instruction::Load:
1008 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy))
1009 error("Invalid load instruction!");
1010 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62);
1013 case 63: // volatile store
1014 case Instruction::Store: {
1015 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2)
1016 error("Invalid store instruction!");
1018 Value *Ptr = getValue(iType, Oprnds[1]);
1019 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType();
1020 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr,
1024 case Instruction::Unwind:
1025 if (Oprnds.size() != 0) error("Invalid unwind instruction!");
1026 Result = new UnwindInst();
1028 case Instruction::Unreachable:
1029 if (Oprnds.size() != 0) error("Invalid unreachable instruction!");
1030 Result = new UnreachableInst();
1032 } // end switch(Opcode)
1034 BB->getInstList().push_back(Result);
1037 if (Result->getType() == InstTy)
1040 TypeSlot = getTypeSlot(Result->getType());
1042 insertValue(Result, TypeSlot, FunctionValues);
1045 /// Get a particular numbered basic block, which might be a forward reference.
1046 /// This works together with ParseBasicBlock to handle these forward references
1047 /// in a clean manner. This function is used when constructing phi, br, switch,
1048 /// and other instructions that reference basic blocks. Blocks are numbered
1049 /// sequentially as they appear in the function.
1050 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) {
1051 // Make sure there is room in the table...
1052 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1);
1054 // First check to see if this is a backwards reference, i.e., ParseBasicBlock
1055 // has already created this block, or if the forward reference has already
1057 if (ParsedBasicBlocks[ID])
1058 return ParsedBasicBlocks[ID];
1060 // Otherwise, the basic block has not yet been created. Do so and add it to
1061 // the ParsedBasicBlocks list.
1062 return ParsedBasicBlocks[ID] = new BasicBlock();
1065 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time.
1066 /// This method reads in one of the basicblock packets. This method is not used
1067 /// for bytecode files after LLVM 1.0
1068 /// @returns The basic block constructed.
1069 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) {
1070 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1074 if (ParsedBasicBlocks.size() == BlockNo)
1075 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1076 else if (ParsedBasicBlocks[BlockNo] == 0)
1077 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1079 BB = ParsedBasicBlocks[BlockNo];
1081 std::vector<unsigned> Operands;
1082 while (moreInBlock())
1083 ParseInstruction(Operands, BB);
1085 if (Handler) Handler->handleBasicBlockEnd(BlockNo);
1089 /// Parse all of the BasicBlock's & Instruction's in the body of a function.
1090 /// In post 1.0 bytecode files, we no longer emit basic block individually,
1091 /// in order to avoid per-basic-block overhead.
1092 /// @returns Rhe number of basic blocks encountered.
1093 unsigned BytecodeReader::ParseInstructionList(Function* F) {
1094 unsigned BlockNo = 0;
1095 std::vector<unsigned> Args;
1097 while (moreInBlock()) {
1098 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1100 if (ParsedBasicBlocks.size() == BlockNo)
1101 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1102 else if (ParsedBasicBlocks[BlockNo] == 0)
1103 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1105 BB = ParsedBasicBlocks[BlockNo];
1107 F->getBasicBlockList().push_back(BB);
1109 // Read instructions into this basic block until we get to a terminator
1110 while (moreInBlock() && !BB->getTerminator())
1111 ParseInstruction(Args, BB);
1113 if (!BB->getTerminator())
1114 error("Non-terminated basic block found!");
1116 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1);
1122 /// Parse a symbol table. This works for both module level and function
1123 /// level symbol tables. For function level symbol tables, the CurrentFunction
1124 /// parameter must be non-zero and the ST parameter must correspond to
1125 /// CurrentFunction's symbol table. For Module level symbol tables, the
1126 /// CurrentFunction argument must be zero.
1127 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction,
1129 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST);
1131 // Allow efficient basic block lookup by number.
1132 std::vector<BasicBlock*> BBMap;
1133 if (CurrentFunction)
1134 for (Function::iterator I = CurrentFunction->begin(),
1135 E = CurrentFunction->end(); I != E; ++I)
1138 /// In LLVM 1.3 we write types separately from values so
1139 /// The types are always first in the symbol table. This is
1140 /// because Type no longer derives from Value.
1141 if (!hasTypeDerivedFromValue) {
1142 // Symtab block header: [num entries]
1143 unsigned NumEntries = read_vbr_uint();
1144 for (unsigned i = 0; i < NumEntries; ++i) {
1145 // Symtab entry: [def slot #][name]
1146 unsigned slot = read_vbr_uint();
1147 std::string Name = read_str();
1148 const Type* T = getType(slot);
1149 ST->insert(Name, T);
1153 while (moreInBlock()) {
1154 // Symtab block header: [num entries][type id number]
1155 unsigned NumEntries = read_vbr_uint();
1157 bool isTypeType = read_typeid(Typ);
1158 const Type *Ty = getType(Typ);
1160 for (unsigned i = 0; i != NumEntries; ++i) {
1161 // Symtab entry: [def slot #][name]
1162 unsigned slot = read_vbr_uint();
1163 std::string Name = read_str();
1165 // if we're reading a pre 1.3 bytecode file and the type plane
1166 // is the "type type", handle it here
1168 const Type* T = getType(slot);
1170 error("Failed type look-up for name '" + Name + "'");
1171 ST->insert(Name, T);
1172 continue; // code below must be short circuited
1175 if (Typ == Type::LabelTyID) {
1176 if (slot < BBMap.size())
1179 V = getValue(Typ, slot, false); // Find mapping...
1182 error("Failed value look-up for name '" + Name + "'");
1187 checkPastBlockEnd("Symbol Table");
1188 if (Handler) Handler->handleSymbolTableEnd();
1191 /// Read in the types portion of a compaction table.
1192 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) {
1193 for (unsigned i = 0; i != NumEntries; ++i) {
1194 unsigned TypeSlot = 0;
1195 if (read_typeid(TypeSlot))
1196 error("Invalid type in compaction table: type type");
1197 const Type *Typ = getGlobalTableType(TypeSlot);
1198 CompactionTypes.push_back(std::make_pair(Typ, TypeSlot));
1199 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ);
1203 /// Parse a compaction table.
1204 void BytecodeReader::ParseCompactionTable() {
1206 // Notify handler that we're beginning a compaction table.
1207 if (Handler) Handler->handleCompactionTableBegin();
1209 // In LLVM 1.3 Type no longer derives from Value. So,
1210 // we always write them first in the compaction table
1211 // because they can't occupy a "type plane" where the
1213 if (! hasTypeDerivedFromValue) {
1214 unsigned NumEntries = read_vbr_uint();
1215 ParseCompactionTypes(NumEntries);
1218 // Compaction tables live in separate blocks so we have to loop
1219 // until we've read the whole thing.
1220 while (moreInBlock()) {
1221 // Read the number of Value* entries in the compaction table
1222 unsigned NumEntries = read_vbr_uint();
1224 unsigned isTypeType = false;
1226 // Decode the type from value read in. Most compaction table
1227 // planes will have one or two entries in them. If that's the
1228 // case then the length is encoded in the bottom two bits and
1229 // the higher bits encode the type. This saves another VBR value.
1230 if ((NumEntries & 3) == 3) {
1231 // In this case, both low-order bits are set (value 3). This
1232 // is a signal that the typeid follows.
1234 isTypeType = read_typeid(Ty);
1236 // In this case, the low-order bits specify the number of entries
1237 // and the high order bits specify the type.
1238 Ty = NumEntries >> 2;
1239 isTypeType = sanitizeTypeId(Ty);
1243 // if we're reading a pre 1.3 bytecode file and the type plane
1244 // is the "type type", handle it here
1246 ParseCompactionTypes(NumEntries);
1248 // Make sure we have enough room for the plane.
1249 if (Ty >= CompactionValues.size())
1250 CompactionValues.resize(Ty+1);
1252 // Make sure the plane is empty or we have some kind of error.
1253 if (!CompactionValues[Ty].empty())
1254 error("Compaction table plane contains multiple entries!");
1256 // Notify handler about the plane.
1257 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries);
1259 // Push the implicit zero.
1260 CompactionValues[Ty].push_back(Constant::getNullValue(getType(Ty)));
1262 // Read in each of the entries, put them in the compaction table
1263 // and notify the handler that we have a new compaction table value.
1264 for (unsigned i = 0; i != NumEntries; ++i) {
1265 unsigned ValSlot = read_vbr_uint();
1266 Value *V = getGlobalTableValue(Ty, ValSlot);
1267 CompactionValues[Ty].push_back(V);
1268 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot);
1272 // Notify handler that the compaction table is done.
1273 if (Handler) Handler->handleCompactionTableEnd();
1276 // Parse a single type. The typeid is read in first. If its a primitive type
1277 // then nothing else needs to be read, we know how to instantiate it. If its
1278 // a derived type, then additional data is read to fill out the type
1280 const Type *BytecodeReader::ParseType() {
1281 unsigned PrimType = 0;
1282 if (read_typeid(PrimType))
1283 error("Invalid type (type type) in type constants!");
1285 const Type *Result = 0;
1286 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType)))
1290 case Type::FunctionTyID: {
1291 const Type *RetType = readSanitizedType();
1293 unsigned NumParams = read_vbr_uint();
1295 std::vector<const Type*> Params;
1297 Params.push_back(readSanitizedType());
1299 bool isVarArg = Params.size() && Params.back() == Type::VoidTy;
1300 if (isVarArg) Params.pop_back();
1302 Result = FunctionType::get(RetType, Params, isVarArg);
1305 case Type::ArrayTyID: {
1306 const Type *ElementType = readSanitizedType();
1307 unsigned NumElements = read_vbr_uint();
1308 Result = ArrayType::get(ElementType, NumElements);
1311 case Type::PackedTyID: {
1312 const Type *ElementType = readSanitizedType();
1313 unsigned NumElements = read_vbr_uint();
1314 Result = PackedType::get(ElementType, NumElements);
1317 case Type::StructTyID: {
1318 std::vector<const Type*> Elements;
1320 if (read_typeid(Typ))
1321 error("Invalid element type (type type) for structure!");
1323 while (Typ) { // List is terminated by void/0 typeid
1324 Elements.push_back(getType(Typ));
1325 if (read_typeid(Typ))
1326 error("Invalid element type (type type) for structure!");
1329 Result = StructType::get(Elements);
1332 case Type::PointerTyID: {
1333 Result = PointerType::get(readSanitizedType());
1337 case Type::OpaqueTyID: {
1338 Result = OpaqueType::get();
1343 error("Don't know how to deserialize primitive type " + utostr(PrimType));
1346 if (Handler) Handler->handleType(Result);
1350 // ParseTypes - We have to use this weird code to handle recursive
1351 // types. We know that recursive types will only reference the current slab of
1352 // values in the type plane, but they can forward reference types before they
1353 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might
1354 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix
1355 // this ugly problem, we pessimistically insert an opaque type for each type we
1356 // are about to read. This means that forward references will resolve to
1357 // something and when we reread the type later, we can replace the opaque type
1358 // with a new resolved concrete type.
1360 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){
1361 assert(Tab.size() == 0 && "should not have read type constants in before!");
1363 // Insert a bunch of opaque types to be resolved later...
1364 Tab.reserve(NumEntries);
1365 for (unsigned i = 0; i != NumEntries; ++i)
1366 Tab.push_back(OpaqueType::get());
1369 Handler->handleTypeList(NumEntries);
1371 // If we are about to resolve types, make sure the type cache is clear.
1373 ModuleTypeIDCache.clear();
1375 // Loop through reading all of the types. Forward types will make use of the
1376 // opaque types just inserted.
1378 for (unsigned i = 0; i != NumEntries; ++i) {
1379 const Type* NewTy = ParseType();
1380 const Type* OldTy = Tab[i].get();
1382 error("Couldn't parse type!");
1384 // Don't directly push the new type on the Tab. Instead we want to replace
1385 // the opaque type we previously inserted with the new concrete value. This
1386 // approach helps with forward references to types. The refinement from the
1387 // abstract (opaque) type to the new type causes all uses of the abstract
1388 // type to use the concrete type (NewTy). This will also cause the opaque
1389 // type to be deleted.
1390 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy);
1392 // This should have replaced the old opaque type with the new type in the
1393 // value table... or with a preexisting type that was already in the system.
1394 // Let's just make sure it did.
1395 assert(Tab[i] != OldTy && "refineAbstractType didn't work!");
1399 /// Parse a single constant value
1400 Constant *BytecodeReader::ParseConstantValue(unsigned TypeID) {
1401 // We must check for a ConstantExpr before switching by type because
1402 // a ConstantExpr can be of any type, and has no explicit value.
1404 // 0 if not expr; numArgs if is expr
1405 unsigned isExprNumArgs = read_vbr_uint();
1407 if (isExprNumArgs) {
1408 // 'undef' is encoded with 'exprnumargs' == 1.
1409 if (!hasNoUndefValue)
1410 if (--isExprNumArgs == 0)
1411 return UndefValue::get(getType(TypeID));
1413 // FIXME: Encoding of constant exprs could be much more compact!
1414 std::vector<Constant*> ArgVec;
1415 ArgVec.reserve(isExprNumArgs);
1416 unsigned Opcode = read_vbr_uint();
1418 // Bytecode files before LLVM 1.4 need have a missing terminator inst.
1419 if (hasNoUnreachableInst) Opcode++;
1421 // Read the slot number and types of each of the arguments
1422 for (unsigned i = 0; i != isExprNumArgs; ++i) {
1423 unsigned ArgValSlot = read_vbr_uint();
1424 unsigned ArgTypeSlot = 0;
1425 if (read_typeid(ArgTypeSlot))
1426 error("Invalid argument type (type type) for constant value");
1428 // Get the arg value from its slot if it exists, otherwise a placeholder
1429 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot));
1432 // Construct a ConstantExpr of the appropriate kind
1433 if (isExprNumArgs == 1) { // All one-operand expressions
1434 if (Opcode != Instruction::Cast)
1435 error("Only cast instruction has one argument for ConstantExpr");
1437 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID));
1438 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1440 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr
1441 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end());
1443 if (hasRestrictedGEPTypes) {
1444 const Type *BaseTy = ArgVec[0]->getType();
1445 generic_gep_type_iterator<std::vector<Constant*>::iterator>
1446 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()),
1447 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end());
1448 for (unsigned i = 0; GTI != E; ++GTI, ++i)
1449 if (isa<StructType>(*GTI)) {
1450 if (IdxList[i]->getType() != Type::UByteTy)
1451 error("Invalid index for getelementptr!");
1452 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy);
1456 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList);
1457 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1459 } else if (Opcode == Instruction::Select) {
1460 if (ArgVec.size() != 3)
1461 error("Select instruction must have three arguments.");
1462 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1],
1464 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1466 } else if (Opcode == Instruction::ExtractElement) {
1467 if (ArgVec.size() != 2)
1468 error("ExtractElement instruction must have two arguments.");
1469 Constant* Result = ConstantExpr::getExtractElement(ArgVec[0], ArgVec[1]);
1470 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1472 } else if (Opcode == Instruction::InsertElement) {
1473 if (ArgVec.size() != 3)
1474 error("InsertElement instruction must have three arguments.");
1476 ConstantExpr::getInsertElement(ArgVec[0], ArgVec[1], ArgVec[2]);
1477 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1479 } else { // All other 2-operand expressions
1480 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]);
1481 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1486 // Ok, not an ConstantExpr. We now know how to read the given type...
1487 const Type *Ty = getType(TypeID);
1488 switch (Ty->getTypeID()) {
1489 case Type::BoolTyID: {
1490 unsigned Val = read_vbr_uint();
1491 if (Val != 0 && Val != 1)
1492 error("Invalid boolean value read.");
1493 Constant* Result = ConstantBool::get(Val == 1);
1494 if (Handler) Handler->handleConstantValue(Result);
1498 case Type::UByteTyID: // Unsigned integer types...
1499 case Type::UShortTyID:
1500 case Type::UIntTyID: {
1501 unsigned Val = read_vbr_uint();
1502 if (!ConstantUInt::isValueValidForType(Ty, Val))
1503 error("Invalid unsigned byte/short/int read.");
1504 Constant* Result = ConstantUInt::get(Ty, Val);
1505 if (Handler) Handler->handleConstantValue(Result);
1509 case Type::ULongTyID: {
1510 Constant* Result = ConstantUInt::get(Ty, read_vbr_uint64());
1511 if (Handler) Handler->handleConstantValue(Result);
1515 case Type::SByteTyID: // Signed integer types...
1516 case Type::ShortTyID:
1517 case Type::IntTyID: {
1518 case Type::LongTyID:
1519 int64_t Val = read_vbr_int64();
1520 if (!ConstantSInt::isValueValidForType(Ty, Val))
1521 error("Invalid signed byte/short/int/long read.");
1522 Constant* Result = ConstantSInt::get(Ty, Val);
1523 if (Handler) Handler->handleConstantValue(Result);
1527 case Type::FloatTyID: {
1530 Constant* Result = ConstantFP::get(Ty, Val);
1531 if (Handler) Handler->handleConstantValue(Result);
1535 case Type::DoubleTyID: {
1538 Constant* Result = ConstantFP::get(Ty, Val);
1539 if (Handler) Handler->handleConstantValue(Result);
1543 case Type::ArrayTyID: {
1544 const ArrayType *AT = cast<ArrayType>(Ty);
1545 unsigned NumElements = AT->getNumElements();
1546 unsigned TypeSlot = getTypeSlot(AT->getElementType());
1547 std::vector<Constant*> Elements;
1548 Elements.reserve(NumElements);
1549 while (NumElements--) // Read all of the elements of the constant.
1550 Elements.push_back(getConstantValue(TypeSlot,
1552 Constant* Result = ConstantArray::get(AT, Elements);
1553 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result);
1557 case Type::StructTyID: {
1558 const StructType *ST = cast<StructType>(Ty);
1560 std::vector<Constant *> Elements;
1561 Elements.reserve(ST->getNumElements());
1562 for (unsigned i = 0; i != ST->getNumElements(); ++i)
1563 Elements.push_back(getConstantValue(ST->getElementType(i),
1566 Constant* Result = ConstantStruct::get(ST, Elements);
1567 if (Handler) Handler->handleConstantStruct(ST, Elements, Result);
1571 case Type::PackedTyID: {
1572 const PackedType *PT = cast<PackedType>(Ty);
1573 unsigned NumElements = PT->getNumElements();
1574 unsigned TypeSlot = getTypeSlot(PT->getElementType());
1575 std::vector<Constant*> Elements;
1576 Elements.reserve(NumElements);
1577 while (NumElements--) // Read all of the elements of the constant.
1578 Elements.push_back(getConstantValue(TypeSlot,
1580 Constant* Result = ConstantPacked::get(PT, Elements);
1581 if (Handler) Handler->handleConstantPacked(PT, Elements, TypeSlot, Result);
1585 case Type::PointerTyID: { // ConstantPointerRef value (backwards compat).
1586 const PointerType *PT = cast<PointerType>(Ty);
1587 unsigned Slot = read_vbr_uint();
1589 // Check to see if we have already read this global variable...
1590 Value *Val = getValue(TypeID, Slot, false);
1592 GlobalValue *GV = dyn_cast<GlobalValue>(Val);
1593 if (!GV) error("GlobalValue not in ValueTable!");
1594 if (Handler) Handler->handleConstantPointer(PT, Slot, GV);
1597 error("Forward references are not allowed here.");
1602 error("Don't know how to deserialize constant value of type '" +
1603 Ty->getDescription());
1609 /// Resolve references for constants. This function resolves the forward
1610 /// referenced constants in the ConstantFwdRefs map. It uses the
1611 /// replaceAllUsesWith method of Value class to substitute the placeholder
1612 /// instance with the actual instance.
1613 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Typ,
1615 ConstantRefsType::iterator I =
1616 ConstantFwdRefs.find(std::make_pair(Typ, Slot));
1617 if (I == ConstantFwdRefs.end()) return; // Never forward referenced?
1619 Value *PH = I->second; // Get the placeholder...
1620 PH->replaceAllUsesWith(NewV);
1621 delete PH; // Delete the old placeholder
1622 ConstantFwdRefs.erase(I); // Remove the map entry for it
1625 /// Parse the constant strings section.
1626 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){
1627 for (; NumEntries; --NumEntries) {
1629 if (read_typeid(Typ))
1630 error("Invalid type (type type) for string constant");
1631 const Type *Ty = getType(Typ);
1632 if (!isa<ArrayType>(Ty))
1633 error("String constant data invalid!");
1635 const ArrayType *ATy = cast<ArrayType>(Ty);
1636 if (ATy->getElementType() != Type::SByteTy &&
1637 ATy->getElementType() != Type::UByteTy)
1638 error("String constant data invalid!");
1640 // Read character data. The type tells us how long the string is.
1641 char *Data = reinterpret_cast<char *>(alloca(ATy->getNumElements()));
1642 read_data(Data, Data+ATy->getNumElements());
1644 std::vector<Constant*> Elements(ATy->getNumElements());
1645 if (ATy->getElementType() == Type::SByteTy)
1646 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1647 Elements[i] = ConstantSInt::get(Type::SByteTy, (signed char)Data[i]);
1649 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1650 Elements[i] = ConstantUInt::get(Type::UByteTy, (unsigned char)Data[i]);
1652 // Create the constant, inserting it as needed.
1653 Constant *C = ConstantArray::get(ATy, Elements);
1654 unsigned Slot = insertValue(C, Typ, Tab);
1655 ResolveReferencesToConstant(C, Typ, Slot);
1656 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C));
1660 /// Parse the constant pool.
1661 void BytecodeReader::ParseConstantPool(ValueTable &Tab,
1662 TypeListTy &TypeTab,
1664 if (Handler) Handler->handleGlobalConstantsBegin();
1666 /// In LLVM 1.3 Type does not derive from Value so the types
1667 /// do not occupy a plane. Consequently, we read the types
1668 /// first in the constant pool.
1669 if (isFunction && !hasTypeDerivedFromValue) {
1670 unsigned NumEntries = read_vbr_uint();
1671 ParseTypes(TypeTab, NumEntries);
1674 while (moreInBlock()) {
1675 unsigned NumEntries = read_vbr_uint();
1677 bool isTypeType = read_typeid(Typ);
1679 /// In LLVM 1.2 and before, Types were written to the
1680 /// bytecode file in the "Type Type" plane (#12).
1681 /// In 1.3 plane 12 is now the label plane. Handle this here.
1683 ParseTypes(TypeTab, NumEntries);
1684 } else if (Typ == Type::VoidTyID) {
1685 /// Use of Type::VoidTyID is a misnomer. It actually means
1686 /// that the following plane is constant strings
1687 assert(&Tab == &ModuleValues && "Cannot read strings in functions!");
1688 ParseStringConstants(NumEntries, Tab);
1690 for (unsigned i = 0; i < NumEntries; ++i) {
1691 Constant *C = ParseConstantValue(Typ);
1692 assert(C && "ParseConstantValue returned NULL!");
1693 unsigned Slot = insertValue(C, Typ, Tab);
1695 // If we are reading a function constant table, make sure that we adjust
1696 // the slot number to be the real global constant number.
1698 if (&Tab != &ModuleValues && Typ < ModuleValues.size() &&
1700 Slot += ModuleValues[Typ]->size();
1701 ResolveReferencesToConstant(C, Typ, Slot);
1706 // After we have finished parsing the constant pool, we had better not have
1707 // any dangling references left.
1708 if (!ConstantFwdRefs.empty()) {
1709 ConstantRefsType::const_iterator I = ConstantFwdRefs.begin();
1710 Constant* missingConst = I->second;
1711 error(utostr(ConstantFwdRefs.size()) +
1712 " unresolved constant reference exist. First one is '" +
1713 missingConst->getName() + "' of type '" +
1714 missingConst->getType()->getDescription() + "'.");
1717 checkPastBlockEnd("Constant Pool");
1718 if (Handler) Handler->handleGlobalConstantsEnd();
1721 /// Parse the contents of a function. Note that this function can be
1722 /// called lazily by materializeFunction
1723 /// @see materializeFunction
1724 void BytecodeReader::ParseFunctionBody(Function* F) {
1726 unsigned FuncSize = BlockEnd - At;
1727 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage;
1729 unsigned LinkageType = read_vbr_uint();
1730 switch (LinkageType) {
1731 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1732 case 1: Linkage = GlobalValue::WeakLinkage; break;
1733 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1734 case 3: Linkage = GlobalValue::InternalLinkage; break;
1735 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1737 error("Invalid linkage type for Function.");
1738 Linkage = GlobalValue::InternalLinkage;
1742 F->setLinkage(Linkage);
1743 if (Handler) Handler->handleFunctionBegin(F,FuncSize);
1745 // Keep track of how many basic blocks we have read in...
1746 unsigned BlockNum = 0;
1747 bool InsertedArguments = false;
1749 BufPtr MyEnd = BlockEnd;
1750 while (At < MyEnd) {
1751 unsigned Type, Size;
1753 read_block(Type, Size);
1756 case BytecodeFormat::ConstantPoolBlockID:
1757 if (!InsertedArguments) {
1758 // Insert arguments into the value table before we parse the first basic
1759 // block in the function, but after we potentially read in the
1760 // compaction table.
1762 InsertedArguments = true;
1765 ParseConstantPool(FunctionValues, FunctionTypes, true);
1768 case BytecodeFormat::CompactionTableBlockID:
1769 ParseCompactionTable();
1772 case BytecodeFormat::BasicBlock: {
1773 if (!InsertedArguments) {
1774 // Insert arguments into the value table before we parse the first basic
1775 // block in the function, but after we potentially read in the
1776 // compaction table.
1778 InsertedArguments = true;
1781 BasicBlock *BB = ParseBasicBlock(BlockNum++);
1782 F->getBasicBlockList().push_back(BB);
1786 case BytecodeFormat::InstructionListBlockID: {
1787 // Insert arguments into the value table before we parse the instruction
1788 // list for the function, but after we potentially read in the compaction
1790 if (!InsertedArguments) {
1792 InsertedArguments = true;
1796 error("Already parsed basic blocks!");
1797 BlockNum = ParseInstructionList(F);
1801 case BytecodeFormat::SymbolTableBlockID:
1802 ParseSymbolTable(F, &F->getSymbolTable());
1808 error("Wrapped around reading bytecode.");
1813 // Malformed bc file if read past end of block.
1817 // Make sure there were no references to non-existant basic blocks.
1818 if (BlockNum != ParsedBasicBlocks.size())
1819 error("Illegal basic block operand reference");
1821 ParsedBasicBlocks.clear();
1823 // Resolve forward references. Replace any uses of a forward reference value
1824 // with the real value.
1825 while (!ForwardReferences.empty()) {
1826 std::map<std::pair<unsigned,unsigned>, Value*>::iterator
1827 I = ForwardReferences.begin();
1828 Value *V = getValue(I->first.first, I->first.second, false);
1829 Value *PlaceHolder = I->second;
1830 PlaceHolder->replaceAllUsesWith(V);
1831 ForwardReferences.erase(I);
1835 // Clear out function-level types...
1836 FunctionTypes.clear();
1837 CompactionTypes.clear();
1838 CompactionValues.clear();
1839 freeTable(FunctionValues);
1841 if (Handler) Handler->handleFunctionEnd(F);
1844 /// This function parses LLVM functions lazily. It obtains the type of the
1845 /// function and records where the body of the function is in the bytecode
1846 /// buffer. The caller can then use the ParseNextFunction and
1847 /// ParseAllFunctionBodies to get handler events for the functions.
1848 void BytecodeReader::ParseFunctionLazily() {
1849 if (FunctionSignatureList.empty())
1850 error("FunctionSignatureList empty!");
1852 Function *Func = FunctionSignatureList.back();
1853 FunctionSignatureList.pop_back();
1855 // Save the information for future reading of the function
1856 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd);
1858 // This function has a body but it's not loaded so it appears `External'.
1859 // Mark it as a `Ghost' instead to notify the users that it has a body.
1860 Func->setLinkage(GlobalValue::GhostLinkage);
1862 // Pretend we've `parsed' this function
1866 /// The ParserFunction method lazily parses one function. Use this method to
1867 /// casue the parser to parse a specific function in the module. Note that
1868 /// this will remove the function from what is to be included by
1869 /// ParseAllFunctionBodies.
1870 /// @see ParseAllFunctionBodies
1871 /// @see ParseBytecode
1872 void BytecodeReader::ParseFunction(Function* Func) {
1873 // Find {start, end} pointers and slot in the map. If not there, we're done.
1874 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func);
1876 // Make sure we found it
1877 if (Fi == LazyFunctionLoadMap.end()) {
1878 error("Unrecognized function of type " + Func->getType()->getDescription());
1882 BlockStart = At = Fi->second.Buf;
1883 BlockEnd = Fi->second.EndBuf;
1884 assert(Fi->first == Func && "Found wrong function?");
1886 LazyFunctionLoadMap.erase(Fi);
1888 this->ParseFunctionBody(Func);
1891 /// The ParseAllFunctionBodies method parses through all the previously
1892 /// unparsed functions in the bytecode file. If you want to completely parse
1893 /// a bytecode file, this method should be called after Parsebytecode because
1894 /// Parsebytecode only records the locations in the bytecode file of where
1895 /// the function definitions are located. This function uses that information
1896 /// to materialize the functions.
1897 /// @see ParseBytecode
1898 void BytecodeReader::ParseAllFunctionBodies() {
1899 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin();
1900 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end();
1903 Function* Func = Fi->first;
1904 BlockStart = At = Fi->second.Buf;
1905 BlockEnd = Fi->second.EndBuf;
1906 ParseFunctionBody(Func);
1909 LazyFunctionLoadMap.clear();
1912 /// Parse the global type list
1913 void BytecodeReader::ParseGlobalTypes() {
1914 // Read the number of types
1915 unsigned NumEntries = read_vbr_uint();
1917 // Ignore the type plane identifier for types if the bc file is pre 1.3
1918 if (hasTypeDerivedFromValue)
1921 ParseTypes(ModuleTypes, NumEntries);
1924 /// Parse the Global info (types, global vars, constants)
1925 void BytecodeReader::ParseModuleGlobalInfo() {
1927 if (Handler) Handler->handleModuleGlobalsBegin();
1929 // SectionID - If a global has an explicit section specified, this map
1930 // remembers the ID until we can translate it into a string.
1931 std::map<GlobalValue*, unsigned> SectionID;
1933 // Read global variables...
1934 unsigned VarType = read_vbr_uint();
1935 while (VarType != Type::VoidTyID) { // List is terminated by Void
1936 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 =
1937 // Linkage, bit4+ = slot#
1938 unsigned SlotNo = VarType >> 5;
1939 if (sanitizeTypeId(SlotNo))
1940 error("Invalid type (type type) for global var!");
1941 unsigned LinkageID = (VarType >> 2) & 7;
1942 bool isConstant = VarType & 1;
1943 bool hasInitializer = (VarType & 2) != 0;
1944 unsigned Alignment = 0;
1945 unsigned GlobalSectionID = 0;
1947 // An extension word is present when linkage = 3 (internal) and hasinit = 0.
1948 if (LinkageID == 3 && !hasInitializer) {
1949 unsigned ExtWord = read_vbr_uint();
1950 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
1951 // linkage, bit 4-8 = alignment (log2), bits 10+ = future use.
1952 hasInitializer = ExtWord & 1;
1953 LinkageID = (ExtWord >> 1) & 7;
1954 Alignment = (1 << ((ExtWord >> 4) & 31)) >> 1;
1956 if (ExtWord & (1 << 9)) // Has a section ID.
1957 GlobalSectionID = read_vbr_uint();
1960 GlobalValue::LinkageTypes Linkage;
1961 switch (LinkageID) {
1962 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1963 case 1: Linkage = GlobalValue::WeakLinkage; break;
1964 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1965 case 3: Linkage = GlobalValue::InternalLinkage; break;
1966 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1968 error("Unknown linkage type: " + utostr(LinkageID));
1969 Linkage = GlobalValue::InternalLinkage;
1973 const Type *Ty = getType(SlotNo);
1975 error("Global has no type! SlotNo=" + utostr(SlotNo));
1977 if (!isa<PointerType>(Ty))
1978 error("Global not a pointer type! Ty= " + Ty->getDescription());
1980 const Type *ElTy = cast<PointerType>(Ty)->getElementType();
1982 // Create the global variable...
1983 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage,
1985 GV->setAlignment(Alignment);
1986 insertValue(GV, SlotNo, ModuleValues);
1988 if (GlobalSectionID != 0)
1989 SectionID[GV] = GlobalSectionID;
1991 unsigned initSlot = 0;
1992 if (hasInitializer) {
1993 initSlot = read_vbr_uint();
1994 GlobalInits.push_back(std::make_pair(GV, initSlot));
1997 // Notify handler about the global value.
1999 Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo,initSlot);
2002 VarType = read_vbr_uint();
2005 // Read the function objects for all of the functions that are coming
2006 unsigned FnSignature = read_vbr_uint();
2008 if (hasNoFlagsForFunctions)
2009 FnSignature = (FnSignature << 5) + 1;
2011 // List is terminated by VoidTy.
2012 while (((FnSignature & (~0U >> 1)) >> 5) != Type::VoidTyID) {
2013 const Type *Ty = getType((FnSignature & (~0U >> 1)) >> 5);
2014 if (!isa<PointerType>(Ty) ||
2015 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) {
2016 error("Function not a pointer to function type! Ty = " +
2017 Ty->getDescription());
2020 // We create functions by passing the underlying FunctionType to create...
2021 const FunctionType* FTy =
2022 cast<FunctionType>(cast<PointerType>(Ty)->getElementType());
2024 // Insert the place holder.
2025 Function *Func = new Function(FTy, GlobalValue::ExternalLinkage,
2028 insertValue(Func, (FnSignature & (~0U >> 1)) >> 5, ModuleValues);
2030 // Flags are not used yet.
2031 unsigned Flags = FnSignature & 31;
2033 // Save this for later so we know type of lazily instantiated functions.
2034 // Note that known-external functions do not have FunctionInfo blocks, so we
2035 // do not add them to the FunctionSignatureList.
2036 if ((Flags & (1 << 4)) == 0)
2037 FunctionSignatureList.push_back(Func);
2039 // Get the calling convention from the low bits.
2040 unsigned CC = Flags & 15;
2041 unsigned Alignment = 0;
2042 if (FnSignature & (1 << 31)) { // Has extension word?
2043 unsigned ExtWord = read_vbr_uint();
2044 Alignment = (1 << (ExtWord & 31)) >> 1;
2045 CC |= ((ExtWord >> 5) & 15) << 4;
2047 if (ExtWord & (1 << 10)) // Has a section ID.
2048 SectionID[Func] = read_vbr_uint();
2051 Func->setCallingConv(CC-1);
2052 Func->setAlignment(Alignment);
2054 if (Handler) Handler->handleFunctionDeclaration(Func);
2056 // Get the next function signature.
2057 FnSignature = read_vbr_uint();
2058 if (hasNoFlagsForFunctions)
2059 FnSignature = (FnSignature << 5) + 1;
2062 // Now that the function signature list is set up, reverse it so that we can
2063 // remove elements efficiently from the back of the vector.
2064 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end());
2066 /// SectionNames - This contains the list of section names encoded in the
2067 /// moduleinfoblock. Functions and globals with an explicit section index
2068 /// into this to get their section name.
2069 std::vector<std::string> SectionNames;
2071 if (hasInconsistentModuleGlobalInfo) {
2073 } else if (!hasNoDependentLibraries) {
2074 // If this bytecode format has dependent library information in it, read in
2075 // the number of dependent library items that follow.
2076 unsigned num_dep_libs = read_vbr_uint();
2077 std::string dep_lib;
2078 while (num_dep_libs--) {
2079 dep_lib = read_str();
2080 TheModule->addLibrary(dep_lib);
2082 Handler->handleDependentLibrary(dep_lib);
2085 // Read target triple and place into the module.
2086 std::string triple = read_str();
2087 TheModule->setTargetTriple(triple);
2089 Handler->handleTargetTriple(triple);
2091 if (At != BlockEnd && !hasAlignment) {
2092 // If the file has section info in it, read the section names now.
2093 unsigned NumSections = read_vbr_uint();
2094 while (NumSections--)
2095 SectionNames.push_back(read_str());
2099 // If any globals are in specified sections, assign them now.
2100 for (std::map<GlobalValue*, unsigned>::iterator I = SectionID.begin(), E =
2101 SectionID.end(); I != E; ++I)
2103 if (I->second > SectionID.size())
2104 error("SectionID out of range for global!");
2105 I->first->setSection(SectionNames[I->second-1]);
2108 // This is for future proofing... in the future extra fields may be added that
2109 // we don't understand, so we transparently ignore them.
2113 if (Handler) Handler->handleModuleGlobalsEnd();
2116 /// Parse the version information and decode it by setting flags on the
2117 /// Reader that enable backward compatibility of the reader.
2118 void BytecodeReader::ParseVersionInfo() {
2119 unsigned Version = read_vbr_uint();
2121 // Unpack version number: low four bits are for flags, top bits = version
2122 Module::Endianness Endianness;
2123 Module::PointerSize PointerSize;
2124 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian;
2125 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32;
2127 bool hasNoEndianness = Version & 4;
2128 bool hasNoPointerSize = Version & 8;
2130 RevisionNum = Version >> 4;
2132 // Default values for the current bytecode version
2133 hasInconsistentModuleGlobalInfo = false;
2134 hasExplicitPrimitiveZeros = false;
2135 hasRestrictedGEPTypes = false;
2136 hasTypeDerivedFromValue = false;
2137 hasLongBlockHeaders = false;
2138 has32BitTypes = false;
2139 hasNoDependentLibraries = false;
2140 hasAlignment = false;
2141 hasNoUndefValue = false;
2142 hasNoFlagsForFunctions = false;
2143 hasNoUnreachableInst = false;
2145 switch (RevisionNum) {
2146 case 0: // LLVM 1.0, 1.1 (Released)
2147 // Base LLVM 1.0 bytecode format.
2148 hasInconsistentModuleGlobalInfo = true;
2149 hasExplicitPrimitiveZeros = true;
2153 case 1: // LLVM 1.2 (Released)
2154 // LLVM 1.2 added explicit support for emitting strings efficiently.
2156 // Also, it fixed the problem where the size of the ModuleGlobalInfo block
2157 // included the size for the alignment at the end, where the rest of the
2160 // LLVM 1.2 and before required that GEP indices be ubyte constants for
2161 // structures and longs for sequential types.
2162 hasRestrictedGEPTypes = true;
2164 // LLVM 1.2 and before had the Type class derive from Value class. This
2165 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are
2166 // written differently because Types can no longer be part of the
2167 // type planes for Values.
2168 hasTypeDerivedFromValue = true;
2172 case 2: // 1.2.5 (Not Released)
2174 // LLVM 1.2 and earlier had two-word block headers. This is a bit wasteful,
2175 // especially for small files where the 8 bytes per block is a large
2176 // fraction of the total block size. In LLVM 1.3, the block type and length
2177 // are compressed into a single 32-bit unsigned integer. 27 bits for length,
2178 // 5 bits for block type.
2179 hasLongBlockHeaders = true;
2181 // LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
2182 // this has been reduced to vbr_uint24. It shouldn't make much difference
2183 // since we haven't run into a module with > 24 million types, but for
2184 // safety the 24-bit restriction has been enforced in 1.3 to free some bits
2185 // in various places and to ensure consistency.
2186 has32BitTypes = true;
2188 // LLVM 1.2 and earlier did not provide a target triple nor a list of
2189 // libraries on which the bytecode is dependent. LLVM 1.3 provides these
2190 // features, for use in future versions of LLVM.
2191 hasNoDependentLibraries = true;
2195 case 3: // LLVM 1.3 (Released)
2196 // LLVM 1.3 and earlier caused alignment bytes to be written on some block
2197 // boundaries and at the end of some strings. In extreme cases (e.g. lots
2198 // of GEP references to a constant array), this can increase the file size
2199 // by 30% or more. In version 1.4 alignment is done away with completely.
2200 hasAlignment = true;
2204 case 4: // 1.3.1 (Not Released)
2205 // In version 4, we did not support the 'undef' constant.
2206 hasNoUndefValue = true;
2208 // In version 4 and above, we did not include space for flags for functions
2209 // in the module info block.
2210 hasNoFlagsForFunctions = true;
2212 // In version 4 and above, we did not include the 'unreachable' instruction
2213 // in the opcode numbering in the bytecode file.
2214 hasNoUnreachableInst = true;
2219 case 5: // 1.4 (Released)
2223 error("Unknown bytecode version number: " + itostr(RevisionNum));
2226 if (hasNoEndianness) Endianness = Module::AnyEndianness;
2227 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize;
2229 TheModule->setEndianness(Endianness);
2230 TheModule->setPointerSize(PointerSize);
2232 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize);
2235 /// Parse a whole module.
2236 void BytecodeReader::ParseModule() {
2237 unsigned Type, Size;
2239 FunctionSignatureList.clear(); // Just in case...
2241 // Read into instance variables...
2245 bool SeenModuleGlobalInfo = false;
2246 bool SeenGlobalTypePlane = false;
2247 BufPtr MyEnd = BlockEnd;
2248 while (At < MyEnd) {
2250 read_block(Type, Size);
2254 case BytecodeFormat::GlobalTypePlaneBlockID:
2255 if (SeenGlobalTypePlane)
2256 error("Two GlobalTypePlane Blocks Encountered!");
2260 SeenGlobalTypePlane = true;
2263 case BytecodeFormat::ModuleGlobalInfoBlockID:
2264 if (SeenModuleGlobalInfo)
2265 error("Two ModuleGlobalInfo Blocks Encountered!");
2266 ParseModuleGlobalInfo();
2267 SeenModuleGlobalInfo = true;
2270 case BytecodeFormat::ConstantPoolBlockID:
2271 ParseConstantPool(ModuleValues, ModuleTypes,false);
2274 case BytecodeFormat::FunctionBlockID:
2275 ParseFunctionLazily();
2278 case BytecodeFormat::SymbolTableBlockID:
2279 ParseSymbolTable(0, &TheModule->getSymbolTable());
2285 error("Unexpected Block of Type #" + utostr(Type) + " encountered!");
2293 // After the module constant pool has been read, we can safely initialize
2294 // global variables...
2295 while (!GlobalInits.empty()) {
2296 GlobalVariable *GV = GlobalInits.back().first;
2297 unsigned Slot = GlobalInits.back().second;
2298 GlobalInits.pop_back();
2300 // Look up the initializer value...
2301 // FIXME: Preserve this type ID!
2303 const llvm::PointerType* GVType = GV->getType();
2304 unsigned TypeSlot = getTypeSlot(GVType->getElementType());
2305 if (Constant *CV = getConstantValue(TypeSlot, Slot)) {
2306 if (GV->hasInitializer())
2307 error("Global *already* has an initializer?!");
2308 if (Handler) Handler->handleGlobalInitializer(GV,CV);
2309 GV->setInitializer(CV);
2311 error("Cannot find initializer value.");
2314 if (!ConstantFwdRefs.empty())
2315 error("Use of undefined constants in a module");
2317 /// Make sure we pulled them all out. If we didn't then there's a declaration
2318 /// but a missing body. That's not allowed.
2319 if (!FunctionSignatureList.empty())
2320 error("Function declared, but bytecode stream ended before definition");
2323 /// This function completely parses a bytecode buffer given by the \p Buf
2324 /// and \p Length parameters.
2325 void BytecodeReader::ParseBytecode(BufPtr Buf, unsigned Length,
2326 const std::string &ModuleID) {
2330 At = MemStart = BlockStart = Buf;
2331 MemEnd = BlockEnd = Buf + Length;
2333 // Create the module
2334 TheModule = new Module(ModuleID);
2336 if (Handler) Handler->handleStart(TheModule, Length);
2338 // Read the four bytes of the signature.
2339 unsigned Sig = read_uint();
2341 // If this is a compressed file
2342 if (Sig == ('l' | ('l' << 8) | ('v' << 16) | ('c' << 24))) {
2344 // Invoke the decompression of the bytecode. Note that we have to skip the
2345 // file's magic number which is not part of the compressed block. Hence,
2346 // the Buf+4 and Length-4. The result goes into decompressedBlock, a data
2347 // member for retention until BytecodeReader is destructed.
2348 unsigned decompressedLength = Compressor::decompressToNewBuffer(
2349 (char*)Buf+4,Length-4,decompressedBlock);
2351 // We must adjust the buffer pointers used by the bytecode reader to point
2352 // into the new decompressed block. After decompression, the
2353 // decompressedBlock will point to a contiguous memory area that has
2354 // the decompressed data.
2355 At = MemStart = BlockStart = Buf = (BufPtr) decompressedBlock;
2356 MemEnd = BlockEnd = Buf + decompressedLength;
2358 // else if this isn't a regular (uncompressed) bytecode file, then its
2359 // and error, generate that now.
2360 } else if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) {
2361 error("Invalid bytecode signature: " + utohexstr(Sig));
2364 // Tell the handler we're starting a module
2365 if (Handler) Handler->handleModuleBegin(ModuleID);
2367 // Get the module block and size and verify. This is handled specially
2368 // because the module block/size is always written in long format. Other
2369 // blocks are written in short format so the read_block method is used.
2370 unsigned Type, Size;
2373 if (Type != BytecodeFormat::ModuleBlockID) {
2374 error("Expected Module Block! Type:" + utostr(Type) + ", Size:"
2378 // It looks like the darwin ranlib program is broken, and adds trailing
2379 // garbage to the end of some bytecode files. This hack allows the bc
2380 // reader to ignore trailing garbage on bytecode files.
2381 if (At + Size < MemEnd)
2382 MemEnd = BlockEnd = At+Size;
2384 if (At + Size != MemEnd)
2385 error("Invalid Top Level Block Length! Type:" + utostr(Type)
2386 + ", Size:" + utostr(Size));
2388 // Parse the module contents
2389 this->ParseModule();
2391 // Look for intrinsic functions and CallInst that need to be upgraded
2392 for (Module::iterator FI = TheModule->begin(), FE = TheModule->end();
2394 UpgradeCallsToIntrinsic(FI);
2396 // Check for missing functions
2398 error("Function expected, but bytecode stream ended!");
2400 // Tell the handler we're done with the module
2402 Handler->handleModuleEnd(ModuleID);
2404 // Tell the handler we're finished the parse
2405 if (Handler) Handler->handleFinish();
2407 } catch (std::string& errstr) {
2408 if (Handler) Handler->handleError(errstr);
2412 if (decompressedBlock != 0 ) {
2413 ::free(decompressedBlock);
2414 decompressedBlock = 0;
2418 std::string msg("Unknown Exception Occurred");
2419 if (Handler) Handler->handleError(msg);
2423 if (decompressedBlock != 0) {
2424 ::free(decompressedBlock);
2425 decompressedBlock = 0;
2431 //===----------------------------------------------------------------------===//
2432 //=== Default Implementations of Handler Methods
2433 //===----------------------------------------------------------------------===//
2435 BytecodeHandler::~BytecodeHandler() {}