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/Bytecode/BytecodeHandler.h"
21 #include "llvm/BasicBlock.h"
22 #include "llvm/Constants.h"
23 #include "llvm/Instructions.h"
24 #include "llvm/SymbolTable.h"
25 #include "llvm/Bytecode/Format.h"
26 #include "llvm/Support/GetElementPtrTypeIterator.h"
27 #include "llvm/Support/Compressor.h"
28 #include "llvm/ADT/StringExtras.h"
35 /// @brief A class for maintaining the slot number definition
36 /// as a placeholder for the actual definition for forward constants defs.
37 class ConstantPlaceHolder : public ConstantExpr {
39 ConstantPlaceHolder(); // DO NOT IMPLEMENT
40 void operator=(const ConstantPlaceHolder &); // DO NOT IMPLEMENT
42 ConstantPlaceHolder(const Type *Ty, unsigned id)
43 : ConstantExpr(Instruction::UserOp1, Constant::getNullValue(Ty), Ty),
45 unsigned getID() { return ID; }
50 // Provide some details on error
51 inline void BytecodeReader::error(std::string err) {
53 err += itostr(RevisionNum) ;
55 err += itostr(At-MemStart);
60 //===----------------------------------------------------------------------===//
61 // Bytecode Reading Methods
62 //===----------------------------------------------------------------------===//
64 /// Determine if the current block being read contains any more data.
65 inline bool BytecodeReader::moreInBlock() {
69 /// Throw an error if we've read past the end of the current block
70 inline void BytecodeReader::checkPastBlockEnd(const char * block_name) {
72 error(std::string("Attempt to read past the end of ") + block_name +
76 /// Align the buffer position to a 32 bit boundary
77 inline void BytecodeReader::align32() {
80 At = (const unsigned char *)((unsigned long)(At+3) & (~3UL));
82 if (Handler) Handler->handleAlignment(At - Save);
84 error("Ran out of data while aligning!");
88 /// Read a whole unsigned integer
89 inline unsigned BytecodeReader::read_uint() {
91 error("Ran out of data reading uint!");
93 return At[-4] | (At[-3] << 8) | (At[-2] << 16) | (At[-1] << 24);
96 /// Read a variable-bit-rate encoded unsigned integer
97 inline unsigned BytecodeReader::read_vbr_uint() {
104 error("Ran out of data reading vbr_uint!");
105 Result |= (unsigned)((*At++) & 0x7F) << Shift;
107 } while (At[-1] & 0x80);
108 if (Handler) Handler->handleVBR32(At-Save);
112 /// Read a variable-bit-rate encoded unsigned 64-bit integer.
113 inline uint64_t BytecodeReader::read_vbr_uint64() {
120 error("Ran out of data reading vbr_uint64!");
121 Result |= (uint64_t)((*At++) & 0x7F) << Shift;
123 } while (At[-1] & 0x80);
124 if (Handler) Handler->handleVBR64(At-Save);
128 /// Read a variable-bit-rate encoded signed 64-bit integer.
129 inline int64_t BytecodeReader::read_vbr_int64() {
130 uint64_t R = read_vbr_uint64();
133 return -(int64_t)(R >> 1);
134 else // There is no such thing as -0 with integers. "-0" really means
135 // 0x8000000000000000.
138 return (int64_t)(R >> 1);
141 /// Read a pascal-style string (length followed by text)
142 inline std::string BytecodeReader::read_str() {
143 unsigned Size = read_vbr_uint();
144 const unsigned char *OldAt = At;
146 if (At > BlockEnd) // Size invalid?
147 error("Ran out of data reading a string!");
148 return std::string((char*)OldAt, Size);
151 /// Read an arbitrary block of data
152 inline void BytecodeReader::read_data(void *Ptr, void *End) {
153 unsigned char *Start = (unsigned char *)Ptr;
154 unsigned Amount = (unsigned char *)End - Start;
155 if (At+Amount > BlockEnd)
156 error("Ran out of data!");
157 std::copy(At, At+Amount, Start);
161 /// Read a float value in little-endian order
162 inline void BytecodeReader::read_float(float& FloatVal) {
163 /// FIXME: This isn't optimal, it has size problems on some platforms
164 /// where FP is not IEEE.
169 FloatUnion.i = At[0] | (At[1] << 8) | (At[2] << 16) | (At[3] << 24);
170 At+=sizeof(uint32_t);
171 FloatVal = FloatUnion.f;
174 /// Read a double value in little-endian order
175 inline void BytecodeReader::read_double(double& DoubleVal) {
176 /// FIXME: This isn't optimal, it has size problems on some platforms
177 /// where FP is not IEEE.
182 DoubleUnion.i = (uint64_t(At[0]) << 0) | (uint64_t(At[1]) << 8) |
183 (uint64_t(At[2]) << 16) | (uint64_t(At[3]) << 24) |
184 (uint64_t(At[4]) << 32) | (uint64_t(At[5]) << 40) |
185 (uint64_t(At[6]) << 48) | (uint64_t(At[7]) << 56);
186 At+=sizeof(uint64_t);
187 DoubleVal = DoubleUnion.d;
190 /// Read a block header and obtain its type and size
191 inline void BytecodeReader::read_block(unsigned &Type, unsigned &Size) {
192 if ( hasLongBlockHeaders ) {
196 case BytecodeFormat::Reserved_DoNotUse :
197 error("Reserved_DoNotUse used as Module Type?");
198 Type = BytecodeFormat::ModuleBlockID; break;
199 case BytecodeFormat::Module:
200 Type = BytecodeFormat::ModuleBlockID; break;
201 case BytecodeFormat::Function:
202 Type = BytecodeFormat::FunctionBlockID; break;
203 case BytecodeFormat::ConstantPool:
204 Type = BytecodeFormat::ConstantPoolBlockID; break;
205 case BytecodeFormat::SymbolTable:
206 Type = BytecodeFormat::SymbolTableBlockID; break;
207 case BytecodeFormat::ModuleGlobalInfo:
208 Type = BytecodeFormat::ModuleGlobalInfoBlockID; break;
209 case BytecodeFormat::GlobalTypePlane:
210 Type = BytecodeFormat::GlobalTypePlaneBlockID; break;
211 case BytecodeFormat::InstructionList:
212 Type = BytecodeFormat::InstructionListBlockID; break;
213 case BytecodeFormat::CompactionTable:
214 Type = BytecodeFormat::CompactionTableBlockID; break;
215 case BytecodeFormat::BasicBlock:
216 /// This block type isn't used after version 1.1. However, we have to
217 /// still allow the value in case this is an old bc format file.
218 /// We just let its value creep thru.
221 error("Invalid block id found: " + utostr(Type));
226 Type = Size & 0x1F; // mask low order five bits
227 Size >>= 5; // get rid of five low order bits, leaving high 27
230 if (At + Size > BlockEnd)
231 error("Attempt to size a block past end of memory");
232 BlockEnd = At + Size;
233 if (Handler) Handler->handleBlock(Type, BlockStart, Size);
237 /// In LLVM 1.2 and before, Types were derived from Value and so they were
238 /// written as part of the type planes along with any other Value. In LLVM
239 /// 1.3 this changed so that Type does not derive from Value. Consequently,
240 /// the BytecodeReader's containers for Values can't contain Types because
241 /// there's no inheritance relationship. This means that the "Type Type"
242 /// plane is defunct along with the Type::TypeTyID TypeID. In LLVM 1.3
243 /// whenever a bytecode construct must have both types and values together,
244 /// the types are always read/written first and then the Values. Furthermore
245 /// since Type::TypeTyID no longer exists, its value (12) now corresponds to
246 /// Type::LabelTyID. In order to overcome this we must "sanitize" all the
247 /// type TypeIDs we encounter. For LLVM 1.3 bytecode files, there's no change.
248 /// For LLVM 1.2 and before, this function will decrement the type id by
249 /// one to account for the missing Type::TypeTyID enumerator if the value is
250 /// larger than 12 (Type::LabelTyID). If the value is exactly 12, then this
251 /// function returns true, otherwise false. This helps detect situations
252 /// where the pre 1.3 bytecode is indicating that what follows is a type.
253 /// @returns true iff type id corresponds to pre 1.3 "type type"
254 inline bool BytecodeReader::sanitizeTypeId(unsigned &TypeId) {
255 if (hasTypeDerivedFromValue) { /// do nothing if 1.3 or later
256 if (TypeId == Type::LabelTyID) {
257 TypeId = Type::VoidTyID; // sanitize it
258 return true; // indicate we got TypeTyID in pre 1.3 bytecode
259 } else if (TypeId > Type::LabelTyID)
260 --TypeId; // shift all planes down because type type plane is missing
265 /// Reads a vbr uint to read in a type id and does the necessary
266 /// conversion on it by calling sanitizeTypeId.
267 /// @returns true iff \p TypeId read corresponds to a pre 1.3 "type type"
268 /// @see sanitizeTypeId
269 inline bool BytecodeReader::read_typeid(unsigned &TypeId) {
270 TypeId = read_vbr_uint();
271 if ( !has32BitTypes )
272 if ( TypeId == 0x00FFFFFF )
273 TypeId = read_vbr_uint();
274 return sanitizeTypeId(TypeId);
277 //===----------------------------------------------------------------------===//
279 //===----------------------------------------------------------------------===//
281 /// Determine if a type id has an implicit null value
282 inline bool BytecodeReader::hasImplicitNull(unsigned TyID) {
283 if (!hasExplicitPrimitiveZeros)
284 return TyID != Type::LabelTyID && TyID != Type::VoidTyID;
285 return TyID >= Type::FirstDerivedTyID;
288 /// Obtain a type given a typeid and account for things like compaction tables,
289 /// function level vs module level, and the offsetting for the primitive types.
290 const Type *BytecodeReader::getType(unsigned ID) {
291 if (ID < Type::FirstDerivedTyID)
292 if (const Type *T = Type::getPrimitiveType((Type::TypeID)ID))
293 return T; // Asked for a primitive type...
295 // Otherwise, derived types need offset...
296 ID -= Type::FirstDerivedTyID;
298 if (!CompactionTypes.empty()) {
299 if (ID >= CompactionTypes.size())
300 error("Type ID out of range for compaction table!");
301 return CompactionTypes[ID].first;
304 // Is it a module-level type?
305 if (ID < ModuleTypes.size())
306 return ModuleTypes[ID].get();
308 // Nope, is it a function-level type?
309 ID -= ModuleTypes.size();
310 if (ID < FunctionTypes.size())
311 return FunctionTypes[ID].get();
313 error("Illegal type reference!");
317 /// Get a sanitized type id. This just makes sure that the \p ID
318 /// is both sanitized and not the "type type" of pre-1.3 bytecode.
319 /// @see sanitizeTypeId
320 inline const Type* BytecodeReader::getSanitizedType(unsigned& ID) {
321 if (sanitizeTypeId(ID))
322 error("Invalid type id encountered");
326 /// This method just saves some coding. It uses read_typeid to read
327 /// in a sanitized type id, errors that its not the type type, and
328 /// then calls getType to return the type value.
329 inline const Type* BytecodeReader::readSanitizedType() {
332 error("Invalid type id encountered");
336 /// Get the slot number associated with a type accounting for primitive
337 /// types, compaction tables, and function level vs module level.
338 unsigned BytecodeReader::getTypeSlot(const Type *Ty) {
339 if (Ty->isPrimitiveType())
340 return Ty->getTypeID();
342 // Scan the compaction table for the type if needed.
343 if (!CompactionTypes.empty()) {
344 for (unsigned i = 0, e = CompactionTypes.size(); i != e; ++i)
345 if (CompactionTypes[i].first == Ty)
346 return Type::FirstDerivedTyID + i;
348 error("Couldn't find type specified in compaction table!");
351 // Check the function level types first...
352 TypeListTy::iterator I = std::find(FunctionTypes.begin(),
353 FunctionTypes.end(), Ty);
355 if (I != FunctionTypes.end())
356 return Type::FirstDerivedTyID + ModuleTypes.size() +
357 (&*I - &FunctionTypes[0]);
359 // Check the module level types now...
360 I = std::find(ModuleTypes.begin(), ModuleTypes.end(), Ty);
361 if (I == ModuleTypes.end())
362 error("Didn't find type in ModuleTypes.");
363 return Type::FirstDerivedTyID + (&*I - &ModuleTypes[0]);
366 /// This is just like getType, but when a compaction table is in use, it is
367 /// ignored. It also ignores function level types.
369 const Type *BytecodeReader::getGlobalTableType(unsigned Slot) {
370 if (Slot < Type::FirstDerivedTyID) {
371 const Type *Ty = Type::getPrimitiveType((Type::TypeID)Slot);
373 error("Not a primitive type ID?");
376 Slot -= Type::FirstDerivedTyID;
377 if (Slot >= ModuleTypes.size())
378 error("Illegal compaction table type reference!");
379 return ModuleTypes[Slot];
382 /// This is just like getTypeSlot, but when a compaction table is in use, it
383 /// is ignored. It also ignores function level types.
384 unsigned BytecodeReader::getGlobalTableTypeSlot(const Type *Ty) {
385 if (Ty->isPrimitiveType())
386 return Ty->getTypeID();
387 TypeListTy::iterator I = std::find(ModuleTypes.begin(),
388 ModuleTypes.end(), Ty);
389 if (I == ModuleTypes.end())
390 error("Didn't find type in ModuleTypes.");
391 return Type::FirstDerivedTyID + (&*I - &ModuleTypes[0]);
394 /// Retrieve a value of a given type and slot number, possibly creating
395 /// it if it doesn't already exist.
396 Value * BytecodeReader::getValue(unsigned type, unsigned oNum, bool Create) {
397 assert(type != Type::LabelTyID && "getValue() cannot get blocks!");
400 // If there is a compaction table active, it defines the low-level numbers.
401 // If not, the module values define the low-level numbers.
402 if (CompactionValues.size() > type && !CompactionValues[type].empty()) {
403 if (Num < CompactionValues[type].size())
404 return CompactionValues[type][Num];
405 Num -= CompactionValues[type].size();
407 // By default, the global type id is the type id passed in
408 unsigned GlobalTyID = type;
410 // If the type plane was compactified, figure out the global type ID by
411 // adding the derived type ids and the distance.
412 if (!CompactionTypes.empty() && type >= Type::FirstDerivedTyID)
413 GlobalTyID = CompactionTypes[type-Type::FirstDerivedTyID].second;
415 if (hasImplicitNull(GlobalTyID)) {
417 return Constant::getNullValue(getType(type));
421 if (GlobalTyID < ModuleValues.size() && ModuleValues[GlobalTyID]) {
422 if (Num < ModuleValues[GlobalTyID]->size())
423 return ModuleValues[GlobalTyID]->getOperand(Num);
424 Num -= ModuleValues[GlobalTyID]->size();
428 if (FunctionValues.size() > type &&
429 FunctionValues[type] &&
430 Num < FunctionValues[type]->size())
431 return FunctionValues[type]->getOperand(Num);
433 if (!Create) return 0; // Do not create a placeholder?
435 // Did we already create a place holder?
436 std::pair<unsigned,unsigned> KeyValue(type, oNum);
437 ForwardReferenceMap::iterator I = ForwardReferences.lower_bound(KeyValue);
438 if (I != ForwardReferences.end() && I->first == KeyValue)
439 return I->second; // We have already created this placeholder
441 // If the type exists (it should)
442 if (const Type* Ty = getType(type)) {
443 // Create the place holder
444 Value *Val = new Argument(Ty);
445 ForwardReferences.insert(I, std::make_pair(KeyValue, Val));
448 throw "Can't create placeholder for value of type slot #" + utostr(type);
451 /// This is just like getValue, but when a compaction table is in use, it
452 /// is ignored. Also, no forward references or other fancy features are
454 Value* BytecodeReader::getGlobalTableValue(unsigned TyID, unsigned SlotNo) {
456 return Constant::getNullValue(getType(TyID));
458 if (!CompactionTypes.empty() && TyID >= Type::FirstDerivedTyID) {
459 TyID -= Type::FirstDerivedTyID;
460 if (TyID >= CompactionTypes.size())
461 error("Type ID out of range for compaction table!");
462 TyID = CompactionTypes[TyID].second;
467 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0 ||
468 SlotNo >= ModuleValues[TyID]->size()) {
469 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0)
470 error("Corrupt compaction table entry!"
471 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
472 + utostr(ModuleValues.size()));
474 error("Corrupt compaction table entry!"
475 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
476 + utostr(ModuleValues.size()) + ", "
477 + utohexstr(reinterpret_cast<uint64_t>(((void*)ModuleValues[TyID])))
479 + utostr(ModuleValues[TyID]->size()));
481 return ModuleValues[TyID]->getOperand(SlotNo);
484 /// Just like getValue, except that it returns a null pointer
485 /// only on error. It always returns a constant (meaning that if the value is
486 /// defined, but is not a constant, that is an error). If the specified
487 /// constant hasn't been parsed yet, a placeholder is defined and used.
488 /// Later, after the real value is parsed, the placeholder is eliminated.
489 Constant* BytecodeReader::getConstantValue(unsigned TypeSlot, unsigned Slot) {
490 if (Value *V = getValue(TypeSlot, Slot, false))
491 if (Constant *C = dyn_cast<Constant>(V))
492 return C; // If we already have the value parsed, just return it
494 error("Value for slot " + utostr(Slot) +
495 " is expected to be a constant!");
497 const Type *Ty = getType(TypeSlot);
498 std::pair<const Type*, unsigned> Key(Ty, Slot);
499 ConstantRefsType::iterator I = ConstantFwdRefs.lower_bound(Key);
501 if (I != ConstantFwdRefs.end() && I->first == Key) {
504 // Create a placeholder for the constant reference and
505 // keep track of the fact that we have a forward ref to recycle it
506 Constant *C = new ConstantPlaceHolder(Ty, Slot);
508 // Keep track of the fact that we have a forward ref to recycle it
509 ConstantFwdRefs.insert(I, std::make_pair(Key, C));
514 //===----------------------------------------------------------------------===//
515 // IR Construction Methods
516 //===----------------------------------------------------------------------===//
518 /// As values are created, they are inserted into the appropriate place
519 /// with this method. The ValueTable argument must be one of ModuleValues
520 /// or FunctionValues data members of this class.
521 unsigned BytecodeReader::insertValue(Value *Val, unsigned type,
522 ValueTable &ValueTab) {
523 assert((!isa<Constant>(Val) || !cast<Constant>(Val)->isNullValue()) ||
524 !hasImplicitNull(type) &&
525 "Cannot read null values from bytecode!");
527 if (ValueTab.size() <= type)
528 ValueTab.resize(type+1);
530 if (!ValueTab[type]) ValueTab[type] = new ValueList();
532 ValueTab[type]->push_back(Val);
534 bool HasOffset = hasImplicitNull(type);
535 return ValueTab[type]->size()-1 + HasOffset;
538 /// Insert the arguments of a function as new values in the reader.
539 void BytecodeReader::insertArguments(Function* F) {
540 const FunctionType *FT = F->getFunctionType();
541 Function::aiterator AI = F->abegin();
542 for (FunctionType::param_iterator It = FT->param_begin();
543 It != FT->param_end(); ++It, ++AI)
544 insertValue(AI, getTypeSlot(AI->getType()), FunctionValues);
547 //===----------------------------------------------------------------------===//
548 // Bytecode Parsing Methods
549 //===----------------------------------------------------------------------===//
551 /// This method parses a single instruction. The instruction is
552 /// inserted at the end of the \p BB provided. The arguments of
553 /// the instruction are provided in the \p Oprnds vector.
554 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds,
558 // Clear instruction data
562 unsigned Op = read_uint();
564 // bits Instruction format: Common to all formats
565 // --------------------------
566 // 01-00: Opcode type, fixed to 1.
568 Opcode = (Op >> 2) & 63;
569 Oprnds.resize((Op >> 0) & 03);
571 // Extract the operands
572 switch (Oprnds.size()) {
574 // bits Instruction format:
575 // --------------------------
576 // 19-08: Resulting type plane
577 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
579 iType = (Op >> 8) & 4095;
580 Oprnds[0] = (Op >> 20) & 4095;
581 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands...
585 // bits Instruction format:
586 // --------------------------
587 // 15-08: Resulting type plane
591 iType = (Op >> 8) & 255;
592 Oprnds[0] = (Op >> 16) & 255;
593 Oprnds[1] = (Op >> 24) & 255;
596 // bits Instruction format:
597 // --------------------------
598 // 13-08: Resulting type plane
603 iType = (Op >> 8) & 63;
604 Oprnds[0] = (Op >> 14) & 63;
605 Oprnds[1] = (Op >> 20) & 63;
606 Oprnds[2] = (Op >> 26) & 63;
609 At -= 4; // Hrm, try this again...
610 Opcode = read_vbr_uint();
612 iType = read_vbr_uint();
614 unsigned NumOprnds = read_vbr_uint();
615 Oprnds.resize(NumOprnds);
618 error("Zero-argument instruction found; this is invalid.");
620 for (unsigned i = 0; i != NumOprnds; ++i)
621 Oprnds[i] = read_vbr_uint();
626 const Type *InstTy = getSanitizedType(iType);
628 // We have enough info to inform the handler now.
629 if (Handler) Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt);
631 // Declare the resulting instruction we'll build.
632 Instruction *Result = 0;
634 // If this is a bytecode format that did not include the unreachable
635 // instruction, bump up all opcodes numbers to make space.
636 if (hasNoUnreachableInst) {
637 if (Opcode >= Instruction::Unreachable &&
643 // Handle binary operators
644 if (Opcode >= Instruction::BinaryOpsBegin &&
645 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2)
646 Result = BinaryOperator::create((Instruction::BinaryOps)Opcode,
647 getValue(iType, Oprnds[0]),
648 getValue(iType, Oprnds[1]));
653 error("Illegal instruction read!");
655 case Instruction::VAArg:
656 Result = new VAArgInst(getValue(iType, Oprnds[0]),
657 getSanitizedType(Oprnds[1]));
659 case Instruction::VANext:
660 Result = new VANextInst(getValue(iType, Oprnds[0]),
661 getSanitizedType(Oprnds[1]));
663 case Instruction::Cast:
664 Result = new CastInst(getValue(iType, Oprnds[0]),
665 getSanitizedType(Oprnds[1]));
667 case Instruction::Select:
668 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]),
669 getValue(iType, Oprnds[1]),
670 getValue(iType, Oprnds[2]));
672 case Instruction::PHI: {
673 if (Oprnds.size() == 0 || (Oprnds.size() & 1))
674 error("Invalid phi node encountered!");
676 PHINode *PN = new PHINode(InstTy);
677 PN->op_reserve(Oprnds.size());
678 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2)
679 PN->addIncoming(getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1]));
684 case Instruction::Shl:
685 case Instruction::Shr:
686 Result = new ShiftInst((Instruction::OtherOps)Opcode,
687 getValue(iType, Oprnds[0]),
688 getValue(Type::UByteTyID, Oprnds[1]));
690 case Instruction::Ret:
691 if (Oprnds.size() == 0)
692 Result = new ReturnInst();
693 else if (Oprnds.size() == 1)
694 Result = new ReturnInst(getValue(iType, Oprnds[0]));
696 error("Unrecognized instruction!");
699 case Instruction::Br:
700 if (Oprnds.size() == 1)
701 Result = new BranchInst(getBasicBlock(Oprnds[0]));
702 else if (Oprnds.size() == 3)
703 Result = new BranchInst(getBasicBlock(Oprnds[0]),
704 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2]));
706 error("Invalid number of operands for a 'br' instruction!");
708 case Instruction::Switch: {
709 if (Oprnds.size() & 1)
710 error("Switch statement with odd number of arguments!");
712 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]),
713 getBasicBlock(Oprnds[1]));
714 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2)
715 I->addCase(cast<Constant>(getValue(iType, Oprnds[i])),
716 getBasicBlock(Oprnds[i+1]));
721 case Instruction::Call: {
722 if (Oprnds.size() == 0)
723 error("Invalid call instruction encountered!");
725 Value *F = getValue(iType, Oprnds[0]);
727 // Check to make sure we have a pointer to function type
728 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
729 if (PTy == 0) error("Call to non function pointer value!");
730 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
731 if (FTy == 0) error("Call to non function pointer value!");
733 std::vector<Value *> Params;
734 if (!FTy->isVarArg()) {
735 FunctionType::param_iterator It = FTy->param_begin();
737 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
738 if (It == FTy->param_end())
739 error("Invalid call instruction!");
740 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
742 if (It != FTy->param_end())
743 error("Invalid call instruction!");
745 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
747 unsigned FirstVariableOperand;
748 if (Oprnds.size() < FTy->getNumParams())
749 error("Call instruction missing operands!");
751 // Read all of the fixed arguments
752 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
753 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i]));
755 FirstVariableOperand = FTy->getNumParams();
757 if ((Oprnds.size()-FirstVariableOperand) & 1)
758 error("Invalid call instruction!"); // Must be pairs of type/value
760 for (unsigned i = FirstVariableOperand, e = Oprnds.size();
762 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
765 Result = new CallInst(F, Params);
768 case Instruction::Invoke: {
769 if (Oprnds.size() < 3)
770 error("Invalid invoke instruction!");
771 Value *F = getValue(iType, Oprnds[0]);
773 // Check to make sure we have a pointer to function type
774 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
776 error("Invoke to non function pointer value!");
777 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
779 error("Invoke to non function pointer value!");
781 std::vector<Value *> Params;
782 BasicBlock *Normal, *Except;
784 if (!FTy->isVarArg()) {
785 Normal = getBasicBlock(Oprnds[1]);
786 Except = getBasicBlock(Oprnds[2]);
788 FunctionType::param_iterator It = FTy->param_begin();
789 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) {
790 if (It == FTy->param_end())
791 error("Invalid invoke instruction!");
792 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
794 if (It != FTy->param_end())
795 error("Invalid invoke instruction!");
797 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
799 Normal = getBasicBlock(Oprnds[0]);
800 Except = getBasicBlock(Oprnds[1]);
802 unsigned FirstVariableArgument = FTy->getNumParams()+2;
803 for (unsigned i = 2; i != FirstVariableArgument; ++i)
804 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)),
807 if (Oprnds.size()-FirstVariableArgument & 1) // Must be type/value pairs
808 error("Invalid invoke instruction!");
810 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2)
811 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
814 Result = new InvokeInst(F, Normal, Except, Params);
817 case Instruction::Malloc:
818 if (Oprnds.size() > 2)
819 error("Invalid malloc instruction!");
820 if (!isa<PointerType>(InstTy))
821 error("Invalid malloc instruction!");
823 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(),
824 Oprnds.size() ? getValue(Type::UIntTyID,
828 case Instruction::Alloca:
829 if (Oprnds.size() > 2)
830 error("Invalid alloca instruction!");
831 if (!isa<PointerType>(InstTy))
832 error("Invalid alloca instruction!");
834 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(),
835 Oprnds.size() ? getValue(Type::UIntTyID,
838 case Instruction::Free:
839 if (!isa<PointerType>(InstTy))
840 error("Invalid free instruction!");
841 Result = new FreeInst(getValue(iType, Oprnds[0]));
843 case Instruction::GetElementPtr: {
844 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy))
845 error("Invalid getelementptr instruction!");
847 std::vector<Value*> Idx;
849 const Type *NextTy = InstTy;
850 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
851 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy);
853 error("Invalid getelementptr instruction!");
855 unsigned ValIdx = Oprnds[i];
857 if (!hasRestrictedGEPTypes) {
858 // Struct indices are always uints, sequential type indices can be any
859 // of the 32 or 64-bit integer types. The actual choice of type is
860 // encoded in the low two bits of the slot number.
861 if (isa<StructType>(TopTy))
862 IdxTy = Type::UIntTyID;
864 switch (ValIdx & 3) {
866 case 0: IdxTy = Type::UIntTyID; break;
867 case 1: IdxTy = Type::IntTyID; break;
868 case 2: IdxTy = Type::ULongTyID; break;
869 case 3: IdxTy = Type::LongTyID; break;
874 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID;
877 Idx.push_back(getValue(IdxTy, ValIdx));
879 // Convert ubyte struct indices into uint struct indices.
880 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes)
881 if (ConstantUInt *C = dyn_cast<ConstantUInt>(Idx.back()))
882 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy);
884 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true);
887 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx);
891 case 62: // volatile load
892 case Instruction::Load:
893 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy))
894 error("Invalid load instruction!");
895 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62);
898 case 63: // volatile store
899 case Instruction::Store: {
900 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2)
901 error("Invalid store instruction!");
903 Value *Ptr = getValue(iType, Oprnds[1]);
904 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType();
905 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr,
909 case Instruction::Unwind:
910 if (Oprnds.size() != 0) error("Invalid unwind instruction!");
911 Result = new UnwindInst();
913 case Instruction::Unreachable:
914 if (Oprnds.size() != 0) error("Invalid unreachable instruction!");
915 Result = new UnreachableInst();
917 } // end switch(Opcode)
920 if (Result->getType() == InstTy)
923 TypeSlot = getTypeSlot(Result->getType());
925 insertValue(Result, TypeSlot, FunctionValues);
926 BB->getInstList().push_back(Result);
929 /// Get a particular numbered basic block, which might be a forward reference.
930 /// This works together with ParseBasicBlock to handle these forward references
931 /// in a clean manner. This function is used when constructing phi, br, switch,
932 /// and other instructions that reference basic blocks. Blocks are numbered
933 /// sequentially as they appear in the function.
934 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) {
935 // Make sure there is room in the table...
936 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1);
938 // First check to see if this is a backwards reference, i.e., ParseBasicBlock
939 // has already created this block, or if the forward reference has already
941 if (ParsedBasicBlocks[ID])
942 return ParsedBasicBlocks[ID];
944 // Otherwise, the basic block has not yet been created. Do so and add it to
945 // the ParsedBasicBlocks list.
946 return ParsedBasicBlocks[ID] = new BasicBlock();
949 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time.
950 /// This method reads in one of the basicblock packets. This method is not used
951 /// for bytecode files after LLVM 1.0
952 /// @returns The basic block constructed.
953 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) {
954 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
958 if (ParsedBasicBlocks.size() == BlockNo)
959 ParsedBasicBlocks.push_back(BB = new BasicBlock());
960 else if (ParsedBasicBlocks[BlockNo] == 0)
961 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
963 BB = ParsedBasicBlocks[BlockNo];
965 std::vector<unsigned> Operands;
966 while (moreInBlock())
967 ParseInstruction(Operands, BB);
969 if (Handler) Handler->handleBasicBlockEnd(BlockNo);
973 /// Parse all of the BasicBlock's & Instruction's in the body of a function.
974 /// In post 1.0 bytecode files, we no longer emit basic block individually,
975 /// in order to avoid per-basic-block overhead.
976 /// @returns Rhe number of basic blocks encountered.
977 unsigned BytecodeReader::ParseInstructionList(Function* F) {
978 unsigned BlockNo = 0;
979 std::vector<unsigned> Args;
981 while (moreInBlock()) {
982 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
984 if (ParsedBasicBlocks.size() == BlockNo)
985 ParsedBasicBlocks.push_back(BB = new BasicBlock());
986 else if (ParsedBasicBlocks[BlockNo] == 0)
987 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
989 BB = ParsedBasicBlocks[BlockNo];
991 F->getBasicBlockList().push_back(BB);
993 // Read instructions into this basic block until we get to a terminator
994 while (moreInBlock() && !BB->getTerminator())
995 ParseInstruction(Args, BB);
997 if (!BB->getTerminator())
998 error("Non-terminated basic block found!");
1000 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1);
1006 /// Parse a symbol table. This works for both module level and function
1007 /// level symbol tables. For function level symbol tables, the CurrentFunction
1008 /// parameter must be non-zero and the ST parameter must correspond to
1009 /// CurrentFunction's symbol table. For Module level symbol tables, the
1010 /// CurrentFunction argument must be zero.
1011 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction,
1013 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST);
1015 // Allow efficient basic block lookup by number.
1016 std::vector<BasicBlock*> BBMap;
1017 if (CurrentFunction)
1018 for (Function::iterator I = CurrentFunction->begin(),
1019 E = CurrentFunction->end(); I != E; ++I)
1022 /// In LLVM 1.3 we write types separately from values so
1023 /// The types are always first in the symbol table. This is
1024 /// because Type no longer derives from Value.
1025 if (!hasTypeDerivedFromValue) {
1026 // Symtab block header: [num entries]
1027 unsigned NumEntries = read_vbr_uint();
1028 for (unsigned i = 0; i < NumEntries; ++i) {
1029 // Symtab entry: [def slot #][name]
1030 unsigned slot = read_vbr_uint();
1031 std::string Name = read_str();
1032 const Type* T = getType(slot);
1033 ST->insert(Name, T);
1037 while (moreInBlock()) {
1038 // Symtab block header: [num entries][type id number]
1039 unsigned NumEntries = read_vbr_uint();
1041 bool isTypeType = read_typeid(Typ);
1042 const Type *Ty = getType(Typ);
1044 for (unsigned i = 0; i != NumEntries; ++i) {
1045 // Symtab entry: [def slot #][name]
1046 unsigned slot = read_vbr_uint();
1047 std::string Name = read_str();
1049 // if we're reading a pre 1.3 bytecode file and the type plane
1050 // is the "type type", handle it here
1052 const Type* T = getType(slot);
1054 error("Failed type look-up for name '" + Name + "'");
1055 ST->insert(Name, T);
1056 continue; // code below must be short circuited
1059 if (Typ == Type::LabelTyID) {
1060 if (slot < BBMap.size())
1063 V = getValue(Typ, slot, false); // Find mapping...
1066 error("Failed value look-up for name '" + Name + "'");
1067 V->setName(Name, ST);
1071 checkPastBlockEnd("Symbol Table");
1072 if (Handler) Handler->handleSymbolTableEnd();
1075 /// Read in the types portion of a compaction table.
1076 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) {
1077 for (unsigned i = 0; i != NumEntries; ++i) {
1078 unsigned TypeSlot = 0;
1079 if (read_typeid(TypeSlot))
1080 error("Invalid type in compaction table: type type");
1081 const Type *Typ = getGlobalTableType(TypeSlot);
1082 CompactionTypes.push_back(std::make_pair(Typ, TypeSlot));
1083 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ);
1087 /// Parse a compaction table.
1088 void BytecodeReader::ParseCompactionTable() {
1090 // Notify handler that we're beginning a compaction table.
1091 if (Handler) Handler->handleCompactionTableBegin();
1093 // In LLVM 1.3 Type no longer derives from Value. So,
1094 // we always write them first in the compaction table
1095 // because they can't occupy a "type plane" where the
1097 if (! hasTypeDerivedFromValue) {
1098 unsigned NumEntries = read_vbr_uint();
1099 ParseCompactionTypes(NumEntries);
1102 // Compaction tables live in separate blocks so we have to loop
1103 // until we've read the whole thing.
1104 while (moreInBlock()) {
1105 // Read the number of Value* entries in the compaction table
1106 unsigned NumEntries = read_vbr_uint();
1108 unsigned isTypeType = false;
1110 // Decode the type from value read in. Most compaction table
1111 // planes will have one or two entries in them. If that's the
1112 // case then the length is encoded in the bottom two bits and
1113 // the higher bits encode the type. This saves another VBR value.
1114 if ((NumEntries & 3) == 3) {
1115 // In this case, both low-order bits are set (value 3). This
1116 // is a signal that the typeid follows.
1118 isTypeType = read_typeid(Ty);
1120 // In this case, the low-order bits specify the number of entries
1121 // and the high order bits specify the type.
1122 Ty = NumEntries >> 2;
1123 isTypeType = sanitizeTypeId(Ty);
1127 // if we're reading a pre 1.3 bytecode file and the type plane
1128 // is the "type type", handle it here
1130 ParseCompactionTypes(NumEntries);
1132 // Make sure we have enough room for the plane.
1133 if (Ty >= CompactionValues.size())
1134 CompactionValues.resize(Ty+1);
1136 // Make sure the plane is empty or we have some kind of error.
1137 if (!CompactionValues[Ty].empty())
1138 error("Compaction table plane contains multiple entries!");
1140 // Notify handler about the plane.
1141 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries);
1143 // Push the implicit zero.
1144 CompactionValues[Ty].push_back(Constant::getNullValue(getType(Ty)));
1146 // Read in each of the entries, put them in the compaction table
1147 // and notify the handler that we have a new compaction table value.
1148 for (unsigned i = 0; i != NumEntries; ++i) {
1149 unsigned ValSlot = read_vbr_uint();
1150 Value *V = getGlobalTableValue(Ty, ValSlot);
1151 CompactionValues[Ty].push_back(V);
1152 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot);
1156 // Notify handler that the compaction table is done.
1157 if (Handler) Handler->handleCompactionTableEnd();
1160 // Parse a single type. The typeid is read in first. If its a primitive type
1161 // then nothing else needs to be read, we know how to instantiate it. If its
1162 // a derived type, then additional data is read to fill out the type
1164 const Type *BytecodeReader::ParseType() {
1165 unsigned PrimType = 0;
1166 if (read_typeid(PrimType))
1167 error("Invalid type (type type) in type constants!");
1169 const Type *Result = 0;
1170 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType)))
1174 case Type::FunctionTyID: {
1175 const Type *RetType = readSanitizedType();
1177 unsigned NumParams = read_vbr_uint();
1179 std::vector<const Type*> Params;
1181 Params.push_back(readSanitizedType());
1183 bool isVarArg = Params.size() && Params.back() == Type::VoidTy;
1184 if (isVarArg) Params.pop_back();
1186 Result = FunctionType::get(RetType, Params, isVarArg);
1189 case Type::ArrayTyID: {
1190 const Type *ElementType = readSanitizedType();
1191 unsigned NumElements = read_vbr_uint();
1192 Result = ArrayType::get(ElementType, NumElements);
1195 case Type::PackedTyID: {
1196 const Type *ElementType = readSanitizedType();
1197 unsigned NumElements = read_vbr_uint();
1198 Result = PackedType::get(ElementType, NumElements);
1201 case Type::StructTyID: {
1202 std::vector<const Type*> Elements;
1204 if (read_typeid(Typ))
1205 error("Invalid element type (type type) for structure!");
1207 while (Typ) { // List is terminated by void/0 typeid
1208 Elements.push_back(getType(Typ));
1209 if (read_typeid(Typ))
1210 error("Invalid element type (type type) for structure!");
1213 Result = StructType::get(Elements);
1216 case Type::PointerTyID: {
1217 Result = PointerType::get(readSanitizedType());
1221 case Type::OpaqueTyID: {
1222 Result = OpaqueType::get();
1227 error("Don't know how to deserialize primitive type " + utostr(PrimType));
1230 if (Handler) Handler->handleType(Result);
1234 // ParseTypes - We have to use this weird code to handle recursive
1235 // types. We know that recursive types will only reference the current slab of
1236 // values in the type plane, but they can forward reference types before they
1237 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might
1238 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix
1239 // this ugly problem, we pessimistically insert an opaque type for each type we
1240 // are about to read. This means that forward references will resolve to
1241 // something and when we reread the type later, we can replace the opaque type
1242 // with a new resolved concrete type.
1244 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){
1245 assert(Tab.size() == 0 && "should not have read type constants in before!");
1247 // Insert a bunch of opaque types to be resolved later...
1248 Tab.reserve(NumEntries);
1249 for (unsigned i = 0; i != NumEntries; ++i)
1250 Tab.push_back(OpaqueType::get());
1253 Handler->handleTypeList(NumEntries);
1255 // Loop through reading all of the types. Forward types will make use of the
1256 // opaque types just inserted.
1258 for (unsigned i = 0; i != NumEntries; ++i) {
1259 const Type* NewTy = ParseType();
1260 const Type* OldTy = Tab[i].get();
1262 error("Couldn't parse type!");
1264 // Don't directly push the new type on the Tab. Instead we want to replace
1265 // the opaque type we previously inserted with the new concrete value. This
1266 // approach helps with forward references to types. The refinement from the
1267 // abstract (opaque) type to the new type causes all uses of the abstract
1268 // type to use the concrete type (NewTy). This will also cause the opaque
1269 // type to be deleted.
1270 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy);
1272 // This should have replaced the old opaque type with the new type in the
1273 // value table... or with a preexisting type that was already in the system.
1274 // Let's just make sure it did.
1275 assert(Tab[i] != OldTy && "refineAbstractType didn't work!");
1279 /// Parse a single constant value
1280 Constant *BytecodeReader::ParseConstantValue(unsigned TypeID) {
1281 // We must check for a ConstantExpr before switching by type because
1282 // a ConstantExpr can be of any type, and has no explicit value.
1284 // 0 if not expr; numArgs if is expr
1285 unsigned isExprNumArgs = read_vbr_uint();
1287 if (isExprNumArgs) {
1288 // 'undef' is encoded with 'exprnumargs' == 1.
1289 if (!hasNoUndefValue)
1290 if (--isExprNumArgs == 0)
1291 return UndefValue::get(getType(TypeID));
1293 // FIXME: Encoding of constant exprs could be much more compact!
1294 std::vector<Constant*> ArgVec;
1295 ArgVec.reserve(isExprNumArgs);
1296 unsigned Opcode = read_vbr_uint();
1298 // Bytecode files before LLVM 1.4 need have a missing terminator inst.
1299 if (hasNoUnreachableInst) Opcode++;
1301 // Read the slot number and types of each of the arguments
1302 for (unsigned i = 0; i != isExprNumArgs; ++i) {
1303 unsigned ArgValSlot = read_vbr_uint();
1304 unsigned ArgTypeSlot = 0;
1305 if (read_typeid(ArgTypeSlot))
1306 error("Invalid argument type (type type) for constant value");
1308 // Get the arg value from its slot if it exists, otherwise a placeholder
1309 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot));
1312 // Construct a ConstantExpr of the appropriate kind
1313 if (isExprNumArgs == 1) { // All one-operand expressions
1314 if (Opcode != Instruction::Cast)
1315 error("Only cast instruction has one argument for ConstantExpr");
1317 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID));
1318 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1320 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr
1321 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end());
1323 if (hasRestrictedGEPTypes) {
1324 const Type *BaseTy = ArgVec[0]->getType();
1325 generic_gep_type_iterator<std::vector<Constant*>::iterator>
1326 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()),
1327 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end());
1328 for (unsigned i = 0; GTI != E; ++GTI, ++i)
1329 if (isa<StructType>(*GTI)) {
1330 if (IdxList[i]->getType() != Type::UByteTy)
1331 error("Invalid index for getelementptr!");
1332 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy);
1336 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList);
1337 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1339 } else if (Opcode == Instruction::Select) {
1340 if (ArgVec.size() != 3)
1341 error("Select instruction must have three arguments.");
1342 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1],
1344 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1346 } else { // All other 2-operand expressions
1347 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]);
1348 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1353 // Ok, not an ConstantExpr. We now know how to read the given type...
1354 const Type *Ty = getType(TypeID);
1355 switch (Ty->getTypeID()) {
1356 case Type::BoolTyID: {
1357 unsigned Val = read_vbr_uint();
1358 if (Val != 0 && Val != 1)
1359 error("Invalid boolean value read.");
1360 Constant* Result = ConstantBool::get(Val == 1);
1361 if (Handler) Handler->handleConstantValue(Result);
1365 case Type::UByteTyID: // Unsigned integer types...
1366 case Type::UShortTyID:
1367 case Type::UIntTyID: {
1368 unsigned Val = read_vbr_uint();
1369 if (!ConstantUInt::isValueValidForType(Ty, Val))
1370 error("Invalid unsigned byte/short/int read.");
1371 Constant* Result = ConstantUInt::get(Ty, Val);
1372 if (Handler) Handler->handleConstantValue(Result);
1376 case Type::ULongTyID: {
1377 Constant* Result = ConstantUInt::get(Ty, read_vbr_uint64());
1378 if (Handler) Handler->handleConstantValue(Result);
1382 case Type::SByteTyID: // Signed integer types...
1383 case Type::ShortTyID:
1384 case Type::IntTyID: {
1385 case Type::LongTyID:
1386 int64_t Val = read_vbr_int64();
1387 if (!ConstantSInt::isValueValidForType(Ty, Val))
1388 error("Invalid signed byte/short/int/long read.");
1389 Constant* Result = ConstantSInt::get(Ty, Val);
1390 if (Handler) Handler->handleConstantValue(Result);
1394 case Type::FloatTyID: {
1397 Constant* Result = ConstantFP::get(Ty, Val);
1398 if (Handler) Handler->handleConstantValue(Result);
1402 case Type::DoubleTyID: {
1405 Constant* Result = ConstantFP::get(Ty, Val);
1406 if (Handler) Handler->handleConstantValue(Result);
1410 case Type::ArrayTyID: {
1411 const ArrayType *AT = cast<ArrayType>(Ty);
1412 unsigned NumElements = AT->getNumElements();
1413 unsigned TypeSlot = getTypeSlot(AT->getElementType());
1414 std::vector<Constant*> Elements;
1415 Elements.reserve(NumElements);
1416 while (NumElements--) // Read all of the elements of the constant.
1417 Elements.push_back(getConstantValue(TypeSlot,
1419 Constant* Result = ConstantArray::get(AT, Elements);
1420 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result);
1424 case Type::StructTyID: {
1425 const StructType *ST = cast<StructType>(Ty);
1427 std::vector<Constant *> Elements;
1428 Elements.reserve(ST->getNumElements());
1429 for (unsigned i = 0; i != ST->getNumElements(); ++i)
1430 Elements.push_back(getConstantValue(ST->getElementType(i),
1433 Constant* Result = ConstantStruct::get(ST, Elements);
1434 if (Handler) Handler->handleConstantStruct(ST, Elements, Result);
1438 case Type::PackedTyID: {
1439 const PackedType *PT = cast<PackedType>(Ty);
1440 unsigned NumElements = PT->getNumElements();
1441 unsigned TypeSlot = getTypeSlot(PT->getElementType());
1442 std::vector<Constant*> Elements;
1443 Elements.reserve(NumElements);
1444 while (NumElements--) // Read all of the elements of the constant.
1445 Elements.push_back(getConstantValue(TypeSlot,
1447 Constant* Result = ConstantPacked::get(PT, Elements);
1448 if (Handler) Handler->handleConstantPacked(PT, Elements, TypeSlot, Result);
1452 case Type::PointerTyID: { // ConstantPointerRef value (backwards compat).
1453 const PointerType *PT = cast<PointerType>(Ty);
1454 unsigned Slot = read_vbr_uint();
1456 // Check to see if we have already read this global variable...
1457 Value *Val = getValue(TypeID, Slot, false);
1459 GlobalValue *GV = dyn_cast<GlobalValue>(Val);
1460 if (!GV) error("GlobalValue not in ValueTable!");
1461 if (Handler) Handler->handleConstantPointer(PT, Slot, GV);
1464 error("Forward references are not allowed here.");
1469 error("Don't know how to deserialize constant value of type '" +
1470 Ty->getDescription());
1476 /// Resolve references for constants. This function resolves the forward
1477 /// referenced constants in the ConstantFwdRefs map. It uses the
1478 /// replaceAllUsesWith method of Value class to substitute the placeholder
1479 /// instance with the actual instance.
1480 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Slot){
1481 ConstantRefsType::iterator I =
1482 ConstantFwdRefs.find(std::make_pair(NewV->getType(), Slot));
1483 if (I == ConstantFwdRefs.end()) return; // Never forward referenced?
1485 Value *PH = I->second; // Get the placeholder...
1486 PH->replaceAllUsesWith(NewV);
1487 delete PH; // Delete the old placeholder
1488 ConstantFwdRefs.erase(I); // Remove the map entry for it
1491 /// Parse the constant strings section.
1492 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){
1493 for (; NumEntries; --NumEntries) {
1495 if (read_typeid(Typ))
1496 error("Invalid type (type type) for string constant");
1497 const Type *Ty = getType(Typ);
1498 if (!isa<ArrayType>(Ty))
1499 error("String constant data invalid!");
1501 const ArrayType *ATy = cast<ArrayType>(Ty);
1502 if (ATy->getElementType() != Type::SByteTy &&
1503 ATy->getElementType() != Type::UByteTy)
1504 error("String constant data invalid!");
1506 // Read character data. The type tells us how long the string is.
1507 char Data[ATy->getNumElements()];
1508 read_data(Data, Data+ATy->getNumElements());
1510 std::vector<Constant*> Elements(ATy->getNumElements());
1511 if (ATy->getElementType() == Type::SByteTy)
1512 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1513 Elements[i] = ConstantSInt::get(Type::SByteTy, (signed char)Data[i]);
1515 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1516 Elements[i] = ConstantUInt::get(Type::UByteTy, (unsigned char)Data[i]);
1518 // Create the constant, inserting it as needed.
1519 Constant *C = ConstantArray::get(ATy, Elements);
1520 unsigned Slot = insertValue(C, Typ, Tab);
1521 ResolveReferencesToConstant(C, Slot);
1522 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C));
1526 /// Parse the constant pool.
1527 void BytecodeReader::ParseConstantPool(ValueTable &Tab,
1528 TypeListTy &TypeTab,
1530 if (Handler) Handler->handleGlobalConstantsBegin();
1532 /// In LLVM 1.3 Type does not derive from Value so the types
1533 /// do not occupy a plane. Consequently, we read the types
1534 /// first in the constant pool.
1535 if (isFunction && !hasTypeDerivedFromValue) {
1536 unsigned NumEntries = read_vbr_uint();
1537 ParseTypes(TypeTab, NumEntries);
1540 while (moreInBlock()) {
1541 unsigned NumEntries = read_vbr_uint();
1543 bool isTypeType = read_typeid(Typ);
1545 /// In LLVM 1.2 and before, Types were written to the
1546 /// bytecode file in the "Type Type" plane (#12).
1547 /// In 1.3 plane 12 is now the label plane. Handle this here.
1549 ParseTypes(TypeTab, NumEntries);
1550 } else if (Typ == Type::VoidTyID) {
1551 /// Use of Type::VoidTyID is a misnomer. It actually means
1552 /// that the following plane is constant strings
1553 assert(&Tab == &ModuleValues && "Cannot read strings in functions!");
1554 ParseStringConstants(NumEntries, Tab);
1556 for (unsigned i = 0; i < NumEntries; ++i) {
1557 Constant *C = ParseConstantValue(Typ);
1558 assert(C && "ParseConstantValue returned NULL!");
1559 unsigned Slot = insertValue(C, Typ, Tab);
1561 // If we are reading a function constant table, make sure that we adjust
1562 // the slot number to be the real global constant number.
1564 if (&Tab != &ModuleValues && Typ < ModuleValues.size() &&
1566 Slot += ModuleValues[Typ]->size();
1567 ResolveReferencesToConstant(C, Slot);
1572 // After we have finished parsing the constant pool, we had better not have
1573 // any dangling references left.
1574 if (!ConstantFwdRefs.empty()) {
1575 typedef std::map<std::pair<const Type*,unsigned>, Constant*> ConstantRefsType;
1576 ConstantRefsType::const_iterator I = ConstantFwdRefs.begin();
1577 const Type* missingType = I->first.first;
1578 Constant* missingConst = I->second;
1579 error(utostr(ConstantFwdRefs.size()) +
1580 " unresolved constant reference exist. First one is '" +
1581 missingConst->getName() + "' of type '" +
1582 missingType->getDescription() + "'.");
1585 checkPastBlockEnd("Constant Pool");
1586 if (Handler) Handler->handleGlobalConstantsEnd();
1589 /// Parse the contents of a function. Note that this function can be
1590 /// called lazily by materializeFunction
1591 /// @see materializeFunction
1592 void BytecodeReader::ParseFunctionBody(Function* F) {
1594 unsigned FuncSize = BlockEnd - At;
1595 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage;
1597 unsigned LinkageType = read_vbr_uint();
1598 switch (LinkageType) {
1599 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1600 case 1: Linkage = GlobalValue::WeakLinkage; break;
1601 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1602 case 3: Linkage = GlobalValue::InternalLinkage; break;
1603 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1605 error("Invalid linkage type for Function.");
1606 Linkage = GlobalValue::InternalLinkage;
1610 F->setLinkage(Linkage);
1611 if (Handler) Handler->handleFunctionBegin(F,FuncSize);
1613 // Keep track of how many basic blocks we have read in...
1614 unsigned BlockNum = 0;
1615 bool InsertedArguments = false;
1617 BufPtr MyEnd = BlockEnd;
1618 while (At < MyEnd) {
1619 unsigned Type, Size;
1621 read_block(Type, Size);
1624 case BytecodeFormat::ConstantPoolBlockID:
1625 if (!InsertedArguments) {
1626 // Insert arguments into the value table before we parse the first basic
1627 // block in the function, but after we potentially read in the
1628 // compaction table.
1630 InsertedArguments = true;
1633 ParseConstantPool(FunctionValues, FunctionTypes, true);
1636 case BytecodeFormat::CompactionTableBlockID:
1637 ParseCompactionTable();
1640 case BytecodeFormat::BasicBlock: {
1641 if (!InsertedArguments) {
1642 // Insert arguments into the value table before we parse the first basic
1643 // block in the function, but after we potentially read in the
1644 // compaction table.
1646 InsertedArguments = true;
1649 BasicBlock *BB = ParseBasicBlock(BlockNum++);
1650 F->getBasicBlockList().push_back(BB);
1654 case BytecodeFormat::InstructionListBlockID: {
1655 // Insert arguments into the value table before we parse the instruction
1656 // list for the function, but after we potentially read in the compaction
1658 if (!InsertedArguments) {
1660 InsertedArguments = true;
1664 error("Already parsed basic blocks!");
1665 BlockNum = ParseInstructionList(F);
1669 case BytecodeFormat::SymbolTableBlockID:
1670 ParseSymbolTable(F, &F->getSymbolTable());
1676 error("Wrapped around reading bytecode.");
1681 // Malformed bc file if read past end of block.
1685 // Make sure there were no references to non-existant basic blocks.
1686 if (BlockNum != ParsedBasicBlocks.size())
1687 error("Illegal basic block operand reference");
1689 ParsedBasicBlocks.clear();
1691 // Resolve forward references. Replace any uses of a forward reference value
1692 // with the real value.
1694 // replaceAllUsesWith is very inefficient for instructions which have a LARGE
1695 // number of operands. PHI nodes often have forward references, and can also
1696 // often have a very large number of operands.
1698 // FIXME: REEVALUATE. replaceAllUsesWith is _much_ faster now, and this code
1699 // should be simplified back to using it!
1701 std::map<Value*, Value*> ForwardRefMapping;
1702 for (std::map<std::pair<unsigned,unsigned>, Value*>::iterator
1703 I = ForwardReferences.begin(), E = ForwardReferences.end();
1705 ForwardRefMapping[I->second] = getValue(I->first.first, I->first.second,
1708 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1709 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1710 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1711 if (Value* V = I->getOperand(i))
1712 if (Argument *A = dyn_cast<Argument>(V)) {
1713 std::map<Value*, Value*>::iterator It = ForwardRefMapping.find(A);
1714 if (It != ForwardRefMapping.end()) I->setOperand(i, It->second);
1717 while (!ForwardReferences.empty()) {
1718 std::map<std::pair<unsigned,unsigned>, Value*>::iterator I =
1719 ForwardReferences.begin();
1720 Value *PlaceHolder = I->second;
1721 ForwardReferences.erase(I);
1723 // Now that all the uses are gone, delete the placeholder...
1724 // If we couldn't find a def (error case), then leak a little
1725 // memory, because otherwise we can't remove all uses!
1729 // Clear out function-level types...
1730 FunctionTypes.clear();
1731 CompactionTypes.clear();
1732 CompactionValues.clear();
1733 freeTable(FunctionValues);
1735 if (Handler) Handler->handleFunctionEnd(F);
1738 /// This function parses LLVM functions lazily. It obtains the type of the
1739 /// function and records where the body of the function is in the bytecode
1740 /// buffer. The caller can then use the ParseNextFunction and
1741 /// ParseAllFunctionBodies to get handler events for the functions.
1742 void BytecodeReader::ParseFunctionLazily() {
1743 if (FunctionSignatureList.empty())
1744 error("FunctionSignatureList empty!");
1746 Function *Func = FunctionSignatureList.back();
1747 FunctionSignatureList.pop_back();
1749 // Save the information for future reading of the function
1750 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd);
1752 // This function has a body but it's not loaded so it appears `External'.
1753 // Mark it as a `Ghost' instead to notify the users that it has a body.
1754 Func->setLinkage(GlobalValue::GhostLinkage);
1756 // Pretend we've `parsed' this function
1760 /// The ParserFunction method lazily parses one function. Use this method to
1761 /// casue the parser to parse a specific function in the module. Note that
1762 /// this will remove the function from what is to be included by
1763 /// ParseAllFunctionBodies.
1764 /// @see ParseAllFunctionBodies
1765 /// @see ParseBytecode
1766 void BytecodeReader::ParseFunction(Function* Func) {
1767 // Find {start, end} pointers and slot in the map. If not there, we're done.
1768 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func);
1770 // Make sure we found it
1771 if (Fi == LazyFunctionLoadMap.end()) {
1772 error("Unrecognized function of type " + Func->getType()->getDescription());
1776 BlockStart = At = Fi->second.Buf;
1777 BlockEnd = Fi->second.EndBuf;
1778 assert(Fi->first == Func && "Found wrong function?");
1780 LazyFunctionLoadMap.erase(Fi);
1782 this->ParseFunctionBody(Func);
1785 /// The ParseAllFunctionBodies method parses through all the previously
1786 /// unparsed functions in the bytecode file. If you want to completely parse
1787 /// a bytecode file, this method should be called after Parsebytecode because
1788 /// Parsebytecode only records the locations in the bytecode file of where
1789 /// the function definitions are located. This function uses that information
1790 /// to materialize the functions.
1791 /// @see ParseBytecode
1792 void BytecodeReader::ParseAllFunctionBodies() {
1793 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin();
1794 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end();
1797 Function* Func = Fi->first;
1798 BlockStart = At = Fi->second.Buf;
1799 BlockEnd = Fi->second.EndBuf;
1800 this->ParseFunctionBody(Func);
1805 /// Parse the global type list
1806 void BytecodeReader::ParseGlobalTypes() {
1807 // Read the number of types
1808 unsigned NumEntries = read_vbr_uint();
1810 // Ignore the type plane identifier for types if the bc file is pre 1.3
1811 if (hasTypeDerivedFromValue)
1814 ParseTypes(ModuleTypes, NumEntries);
1817 /// Parse the Global info (types, global vars, constants)
1818 void BytecodeReader::ParseModuleGlobalInfo() {
1820 if (Handler) Handler->handleModuleGlobalsBegin();
1822 // Read global variables...
1823 unsigned VarType = read_vbr_uint();
1824 while (VarType != Type::VoidTyID) { // List is terminated by Void
1825 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 =
1826 // Linkage, bit4+ = slot#
1827 unsigned SlotNo = VarType >> 5;
1828 if (sanitizeTypeId(SlotNo))
1829 error("Invalid type (type type) for global var!");
1830 unsigned LinkageID = (VarType >> 2) & 7;
1831 bool isConstant = VarType & 1;
1832 bool hasInitializer = VarType & 2;
1833 GlobalValue::LinkageTypes Linkage;
1835 switch (LinkageID) {
1836 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1837 case 1: Linkage = GlobalValue::WeakLinkage; break;
1838 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1839 case 3: Linkage = GlobalValue::InternalLinkage; break;
1840 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1842 error("Unknown linkage type: " + utostr(LinkageID));
1843 Linkage = GlobalValue::InternalLinkage;
1847 const Type *Ty = getType(SlotNo);
1849 error("Global has no type! SlotNo=" + utostr(SlotNo));
1852 if (!isa<PointerType>(Ty)) {
1853 error("Global not a pointer type! Ty= " + Ty->getDescription());
1856 const Type *ElTy = cast<PointerType>(Ty)->getElementType();
1858 // Create the global variable...
1859 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage,
1861 insertValue(GV, SlotNo, ModuleValues);
1863 unsigned initSlot = 0;
1864 if (hasInitializer) {
1865 initSlot = read_vbr_uint();
1866 GlobalInits.push_back(std::make_pair(GV, initSlot));
1869 // Notify handler about the global value.
1871 Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo,initSlot);
1874 VarType = read_vbr_uint();
1877 // Read the function objects for all of the functions that are coming
1878 unsigned FnSignature = read_vbr_uint();
1880 if (hasNoFlagsForFunctions)
1881 FnSignature = (FnSignature << 5) + 1;
1883 // List is terminated by VoidTy.
1884 while ((FnSignature >> 5) != Type::VoidTyID) {
1885 const Type *Ty = getType(FnSignature >> 5);
1886 if (!isa<PointerType>(Ty) ||
1887 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) {
1888 error("Function not a pointer to function type! Ty = " +
1889 Ty->getDescription());
1892 // We create functions by passing the underlying FunctionType to create...
1893 const FunctionType* FTy =
1894 cast<FunctionType>(cast<PointerType>(Ty)->getElementType());
1897 // Insert the place holder.
1898 Function* Func = new Function(FTy, GlobalValue::ExternalLinkage,
1900 insertValue(Func, FnSignature >> 5, ModuleValues);
1902 // Flags are not used yet.
1903 unsigned Flags = FnSignature & 31;
1905 // Save this for later so we know type of lazily instantiated functions.
1906 // Note that known-external functions do not have FunctionInfo blocks, so we
1907 // do not add them to the FunctionSignatureList.
1908 if ((Flags & (1 << 4)) == 0)
1909 FunctionSignatureList.push_back(Func);
1911 if (Handler) Handler->handleFunctionDeclaration(Func);
1913 // Get the next function signature.
1914 FnSignature = read_vbr_uint();
1915 if (hasNoFlagsForFunctions)
1916 FnSignature = (FnSignature << 5) + 1;
1919 // Now that the function signature list is set up, reverse it so that we can
1920 // remove elements efficiently from the back of the vector.
1921 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end());
1923 // If this bytecode format has dependent library information in it ..
1924 if (!hasNoDependentLibraries) {
1925 // Read in the number of dependent library items that follow
1926 unsigned num_dep_libs = read_vbr_uint();
1927 std::string dep_lib;
1928 while( num_dep_libs-- ) {
1929 dep_lib = read_str();
1930 TheModule->addLibrary(dep_lib);
1932 Handler->handleDependentLibrary(dep_lib);
1936 // Read target triple and place into the module
1937 std::string triple = read_str();
1938 TheModule->setTargetTriple(triple);
1940 Handler->handleTargetTriple(triple);
1943 if (hasInconsistentModuleGlobalInfo)
1946 // This is for future proofing... in the future extra fields may be added that
1947 // we don't understand, so we transparently ignore them.
1951 if (Handler) Handler->handleModuleGlobalsEnd();
1954 /// Parse the version information and decode it by setting flags on the
1955 /// Reader that enable backward compatibility of the reader.
1956 void BytecodeReader::ParseVersionInfo() {
1957 unsigned Version = read_vbr_uint();
1959 // Unpack version number: low four bits are for flags, top bits = version
1960 Module::Endianness Endianness;
1961 Module::PointerSize PointerSize;
1962 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian;
1963 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32;
1965 bool hasNoEndianness = Version & 4;
1966 bool hasNoPointerSize = Version & 8;
1968 RevisionNum = Version >> 4;
1970 // Default values for the current bytecode version
1971 hasInconsistentModuleGlobalInfo = false;
1972 hasExplicitPrimitiveZeros = false;
1973 hasRestrictedGEPTypes = false;
1974 hasTypeDerivedFromValue = false;
1975 hasLongBlockHeaders = false;
1976 has32BitTypes = false;
1977 hasNoDependentLibraries = false;
1978 hasAlignment = false;
1979 hasInconsistentBBSlotNums = false;
1980 hasVBRByteTypes = false;
1981 hasUnnecessaryModuleBlockId = false;
1982 hasNoUndefValue = false;
1983 hasNoFlagsForFunctions = false;
1984 hasNoUnreachableInst = false;
1986 switch (RevisionNum) {
1987 case 0: // LLVM 1.0, 1.1 (Released)
1988 // Base LLVM 1.0 bytecode format.
1989 hasInconsistentModuleGlobalInfo = true;
1990 hasExplicitPrimitiveZeros = true;
1994 case 1: // LLVM 1.2 (Released)
1995 // LLVM 1.2 added explicit support for emitting strings efficiently.
1997 // Also, it fixed the problem where the size of the ModuleGlobalInfo block
1998 // included the size for the alignment at the end, where the rest of the
2001 // LLVM 1.2 and before required that GEP indices be ubyte constants for
2002 // structures and longs for sequential types.
2003 hasRestrictedGEPTypes = true;
2005 // LLVM 1.2 and before had the Type class derive from Value class. This
2006 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are
2007 // written differently because Types can no longer be part of the
2008 // type planes for Values.
2009 hasTypeDerivedFromValue = true;
2013 case 2: // 1.2.5 (Not Released)
2015 // LLVM 1.2 and earlier had two-word block headers. This is a bit wasteful,
2016 // especially for small files where the 8 bytes per block is a large
2017 // fraction of the total block size. In LLVM 1.3, the block type and length
2018 // are compressed into a single 32-bit unsigned integer. 27 bits for length,
2019 // 5 bits for block type.
2020 hasLongBlockHeaders = true;
2022 // LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
2023 // this has been reduced to vbr_uint24. It shouldn't make much difference
2024 // since we haven't run into a module with > 24 million types, but for
2025 // safety the 24-bit restriction has been enforced in 1.3 to free some bits
2026 // in various places and to ensure consistency.
2027 has32BitTypes = true;
2029 // LLVM 1.2 and earlier did not provide a target triple nor a list of
2030 // libraries on which the bytecode is dependent. LLVM 1.3 provides these
2031 // features, for use in future versions of LLVM.
2032 hasNoDependentLibraries = true;
2036 case 3: // LLVM 1.3 (Released)
2037 // LLVM 1.3 and earlier caused alignment bytes to be written on some block
2038 // boundaries and at the end of some strings. In extreme cases (e.g. lots
2039 // of GEP references to a constant array), this can increase the file size
2040 // by 30% or more. In version 1.4 alignment is done away with completely.
2041 hasAlignment = true;
2045 case 4: // 1.3.1 (Not Released)
2046 // In version 4, we did not support the 'undef' constant.
2047 hasNoUndefValue = true;
2049 // In version 4 and above, we did not include space for flags for functions
2050 // in the module info block.
2051 hasNoFlagsForFunctions = true;
2053 // In version 4 and above, we did not include the 'unreachable' instruction
2054 // in the opcode numbering in the bytecode file.
2055 hasNoUnreachableInst = true;
2060 case 5: // 1.x.x (Not Released)
2062 // FIXME: NONE of this is implemented yet!
2064 // In version 5, basic blocks have a minimum index of 0 whereas all the
2065 // other primitives have a minimum index of 1 (because 0 is the "null"
2066 // value. In version 5, we made this consistent.
2067 hasInconsistentBBSlotNums = true;
2069 // In version 5, the types SByte and UByte were encoded as vbr_uint so that
2070 // signed values > 63 and unsigned values >127 would be encoded as two
2071 // bytes. In version 5, they are encoded directly in a single byte.
2072 hasVBRByteTypes = true;
2074 // In version 5, modules begin with a "Module Block" which encodes a 4-byte
2075 // integer value 0x01 to identify the module block. This is unnecessary and
2076 // removed in version 5.
2077 hasUnnecessaryModuleBlockId = true;
2080 error("Unknown bytecode version number: " + itostr(RevisionNum));
2083 if (hasNoEndianness) Endianness = Module::AnyEndianness;
2084 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize;
2086 TheModule->setEndianness(Endianness);
2087 TheModule->setPointerSize(PointerSize);
2089 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize);
2092 /// Parse a whole module.
2093 void BytecodeReader::ParseModule() {
2094 unsigned Type, Size;
2096 FunctionSignatureList.clear(); // Just in case...
2098 // Read into instance variables...
2102 bool SeenModuleGlobalInfo = false;
2103 bool SeenGlobalTypePlane = false;
2104 BufPtr MyEnd = BlockEnd;
2105 while (At < MyEnd) {
2107 read_block(Type, Size);
2111 case BytecodeFormat::GlobalTypePlaneBlockID:
2112 if (SeenGlobalTypePlane)
2113 error("Two GlobalTypePlane Blocks Encountered!");
2117 SeenGlobalTypePlane = true;
2120 case BytecodeFormat::ModuleGlobalInfoBlockID:
2121 if (SeenModuleGlobalInfo)
2122 error("Two ModuleGlobalInfo Blocks Encountered!");
2123 ParseModuleGlobalInfo();
2124 SeenModuleGlobalInfo = true;
2127 case BytecodeFormat::ConstantPoolBlockID:
2128 ParseConstantPool(ModuleValues, ModuleTypes,false);
2131 case BytecodeFormat::FunctionBlockID:
2132 ParseFunctionLazily();
2135 case BytecodeFormat::SymbolTableBlockID:
2136 ParseSymbolTable(0, &TheModule->getSymbolTable());
2142 error("Unexpected Block of Type #" + utostr(Type) + " encountered!");
2150 // After the module constant pool has been read, we can safely initialize
2151 // global variables...
2152 while (!GlobalInits.empty()) {
2153 GlobalVariable *GV = GlobalInits.back().first;
2154 unsigned Slot = GlobalInits.back().second;
2155 GlobalInits.pop_back();
2157 // Look up the initializer value...
2158 // FIXME: Preserve this type ID!
2160 const llvm::PointerType* GVType = GV->getType();
2161 unsigned TypeSlot = getTypeSlot(GVType->getElementType());
2162 if (Constant *CV = getConstantValue(TypeSlot, Slot)) {
2163 if (GV->hasInitializer())
2164 error("Global *already* has an initializer?!");
2165 if (Handler) Handler->handleGlobalInitializer(GV,CV);
2166 GV->setInitializer(CV);
2168 error("Cannot find initializer value.");
2171 /// Make sure we pulled them all out. If we didn't then there's a declaration
2172 /// but a missing body. That's not allowed.
2173 if (!FunctionSignatureList.empty())
2174 error("Function declared, but bytecode stream ended before definition");
2177 /// This function completely parses a bytecode buffer given by the \p Buf
2178 /// and \p Length parameters.
2179 void BytecodeReader::ParseBytecode(BufPtr Buf, unsigned Length,
2180 const std::string &ModuleID) {
2184 At = MemStart = BlockStart = Buf;
2185 MemEnd = BlockEnd = Buf + Length;
2187 // Create the module
2188 TheModule = new Module(ModuleID);
2190 if (Handler) Handler->handleStart(TheModule, Length);
2192 // Read the four bytes of the signature.
2193 unsigned Sig = read_uint();
2195 // If this is a compressed file
2196 if (Sig == ('l' | ('l' << 8) | ('v' << 16) | ('c' << 24))) {
2198 // Invoke the decompression of the bytecode. Note that we have to skip the
2199 // file's magic number which is not part of the compressed block. Hence,
2200 // the Buf+4 and Length-4. The result goes into decompressedBlock, a data
2201 // member for retention until BytecodeReader is destructed.
2202 unsigned decompressedLength = Compressor::decompressToNewBuffer(
2203 (char*)Buf+4,Length-4,decompressedBlock);
2205 // We must adjust the buffer pointers used by the bytecode reader to point
2206 // into the new decompressed block. After decompression, the
2207 // decompressedBlock will point to a contiguous memory area that has
2208 // the decompressed data.
2209 At = MemStart = BlockStart = Buf = (BufPtr) decompressedBlock;
2210 MemEnd = BlockEnd = Buf + decompressedLength;
2212 // else if this isn't a regular (uncompressed) bytecode file, then its
2213 // and error, generate that now.
2214 } else if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) {
2215 error("Invalid bytecode signature: " + utohexstr(Sig));
2218 // Tell the handler we're starting a module
2219 if (Handler) Handler->handleModuleBegin(ModuleID);
2221 // Get the module block and size and verify. This is handled specially
2222 // because the module block/size is always written in long format. Other
2223 // blocks are written in short format so the read_block method is used.
2224 unsigned Type, Size;
2227 if (Type != BytecodeFormat::ModuleBlockID) {
2228 error("Expected Module Block! Type:" + utostr(Type) + ", Size:"
2232 // It looks like the darwin ranlib program is broken, and adds trailing
2233 // garbage to the end of some bytecode files. This hack allows the bc
2234 // reader to ignore trailing garbage on bytecode files.
2235 if (At + Size < MemEnd)
2236 MemEnd = BlockEnd = At+Size;
2238 if (At + Size != MemEnd)
2239 error("Invalid Top Level Block Length! Type:" + utostr(Type)
2240 + ", Size:" + utostr(Size));
2242 // Parse the module contents
2243 this->ParseModule();
2245 // Check for missing functions
2247 error("Function expected, but bytecode stream ended!");
2249 // Tell the handler we're done with the module
2251 Handler->handleModuleEnd(ModuleID);
2253 // Tell the handler we're finished the parse
2254 if (Handler) Handler->handleFinish();
2256 } catch (std::string& errstr) {
2257 if (Handler) Handler->handleError(errstr);
2261 if (decompressedBlock != 0 ) {
2262 ::free(decompressedBlock);
2263 decompressedBlock = 0;
2267 std::string msg("Unknown Exception Occurred");
2268 if (Handler) Handler->handleError(msg);
2272 if (decompressedBlock != 0) {
2273 ::free(decompressedBlock);
2274 decompressedBlock = 0;
2280 //===----------------------------------------------------------------------===//
2281 //=== Default Implementations of Handler Methods
2282 //===----------------------------------------------------------------------===//
2284 BytecodeHandler::~BytecodeHandler() {}