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 "Support/StringExtras.h"
34 /// @brief A class for maintaining the slot number definition
35 /// as a placeholder for the actual definition for forward constants defs.
36 class ConstantPlaceHolder : public ConstantExpr {
38 ConstantPlaceHolder(); // DO NOT IMPLEMENT
39 void operator=(const ConstantPlaceHolder &); // DO NOT IMPLEMENT
41 ConstantPlaceHolder(const Type *Ty, unsigned id)
42 : ConstantExpr(Instruction::UserOp1, Constant::getNullValue(Ty), Ty),
44 unsigned getID() { return ID; }
49 // Provide some details on error
50 inline void BytecodeReader::error(std::string err) {
52 err += itostr(RevisionNum) ;
54 err += itostr(At-MemStart);
59 //===----------------------------------------------------------------------===//
60 // Bytecode Reading Methods
61 //===----------------------------------------------------------------------===//
63 /// Determine if the current block being read contains any more data.
64 inline bool BytecodeReader::moreInBlock() {
68 /// Throw an error if we've read past the end of the current block
69 inline void BytecodeReader::checkPastBlockEnd(const char * block_name) {
71 error(std::string("Attempt to read past the end of ") + block_name + " block.");
74 /// Align the buffer position to a 32 bit boundary
75 inline void BytecodeReader::align32() {
77 At = (const unsigned char *)((unsigned long)(At+3) & (~3UL));
79 if (Handler) Handler->handleAlignment(At - Save);
81 error("Ran out of data while aligning!");
84 /// Read a whole unsigned integer
85 inline unsigned BytecodeReader::read_uint() {
87 error("Ran out of data reading uint!");
89 return At[-4] | (At[-3] << 8) | (At[-2] << 16) | (At[-1] << 24);
92 /// Read a variable-bit-rate encoded unsigned integer
93 inline unsigned BytecodeReader::read_vbr_uint() {
100 error("Ran out of data reading vbr_uint!");
101 Result |= (unsigned)((*At++) & 0x7F) << Shift;
103 } while (At[-1] & 0x80);
104 if (Handler) Handler->handleVBR32(At-Save);
108 /// Read a variable-bit-rate encoded unsigned 64-bit integer.
109 inline uint64_t BytecodeReader::read_vbr_uint64() {
116 error("Ran out of data reading vbr_uint64!");
117 Result |= (uint64_t)((*At++) & 0x7F) << Shift;
119 } while (At[-1] & 0x80);
120 if (Handler) Handler->handleVBR64(At-Save);
124 /// Read a variable-bit-rate encoded signed 64-bit integer.
125 inline int64_t BytecodeReader::read_vbr_int64() {
126 uint64_t R = read_vbr_uint64();
129 return -(int64_t)(R >> 1);
130 else // There is no such thing as -0 with integers. "-0" really means
131 // 0x8000000000000000.
134 return (int64_t)(R >> 1);
137 /// Read a pascal-style string (length followed by text)
138 inline std::string BytecodeReader::read_str() {
139 unsigned Size = read_vbr_uint();
140 const unsigned char *OldAt = At;
142 if (At > BlockEnd) // Size invalid?
143 error("Ran out of data reading a string!");
144 return std::string((char*)OldAt, Size);
147 /// Read an arbitrary block of data
148 inline void BytecodeReader::read_data(void *Ptr, void *End) {
149 unsigned char *Start = (unsigned char *)Ptr;
150 unsigned Amount = (unsigned char *)End - Start;
151 if (At+Amount > BlockEnd)
152 error("Ran out of data!");
153 std::copy(At, At+Amount, Start);
157 /// Read a float value in little-endian order
158 inline void BytecodeReader::read_float(float& FloatVal) {
159 /// FIXME: This is a broken implementation! It reads
160 /// it in a platform-specific endianess. Need to make
161 /// it little endian always.
162 read_data(&FloatVal, &FloatVal+1);
165 /// Read a double value in little-endian order
166 inline void BytecodeReader::read_double(double& DoubleVal) {
167 /// FIXME: This is a broken implementation! It reads
168 /// it in a platform-specific endianess. Need to make
169 /// it little endian always.
170 read_data(&DoubleVal, &DoubleVal+1);
173 /// Read a block header and obtain its type and size
174 inline void BytecodeReader::read_block(unsigned &Type, unsigned &Size) {
178 if (At + Size > BlockEnd)
179 error("Attempt to size a block past end of memory");
180 BlockEnd = At + Size;
181 if (Handler) Handler->handleBlock(Type, BlockStart, Size);
185 /// In LLVM 1.2 and before, Types were derived from Value and so they were
186 /// written as part of the type planes along with any other Value. In LLVM
187 /// 1.3 this changed so that Type does not derive from Value. Consequently,
188 /// the BytecodeReader's containers for Values can't contain Types because
189 /// there's no inheritance relationship. This means that the "Type Type"
190 /// plane is defunct along with the Type::TypeTyID TypeID. In LLVM 1.3
191 /// whenever a bytecode construct must have both types and values together,
192 /// the types are always read/written first and then the Values. Furthermore
193 /// since Type::TypeTyID no longer exists, its value (12) now corresponds to
194 /// Type::LabelTyID. In order to overcome this we must "sanitize" all the
195 /// type TypeIDs we encounter. For LLVM 1.3 bytecode files, there's no change.
196 /// For LLVM 1.2 and before, this function will decrement the type id by
197 /// one to account for the missing Type::TypeTyID enumerator if the value is
198 /// larger than 12 (Type::LabelTyID). If the value is exactly 12, then this
199 /// function returns true, otherwise false. This helps detect situations
200 /// where the pre 1.3 bytecode is indicating that what follows is a type.
201 /// @returns true iff type id corresponds to pre 1.3 "type type"
202 inline bool BytecodeReader::sanitizeTypeId(unsigned &TypeId) {
203 if (hasTypeDerivedFromValue) { /// do nothing if 1.3 or later
204 if (TypeId == Type::LabelTyID) {
205 TypeId = Type::VoidTyID; // sanitize it
206 return true; // indicate we got TypeTyID in pre 1.3 bytecode
207 } else if (TypeId > Type::LabelTyID)
208 --TypeId; // shift all planes down because type type plane is missing
213 /// Reads a vbr uint to read in a type id and does the necessary
214 /// conversion on it by calling sanitizeTypeId.
215 /// @returns true iff \p TypeId read corresponds to a pre 1.3 "type type"
216 /// @see sanitizeTypeId
217 inline bool BytecodeReader::read_typeid(unsigned &TypeId) {
218 TypeId = read_vbr_uint();
219 return sanitizeTypeId(TypeId);
222 //===----------------------------------------------------------------------===//
224 //===----------------------------------------------------------------------===//
226 /// Determine if a type id has an implicit null value
227 inline bool BytecodeReader::hasImplicitNull(unsigned TyID) {
228 if (!hasExplicitPrimitiveZeros)
229 return TyID != Type::LabelTyID && TyID != Type::VoidTyID;
230 return TyID >= Type::FirstDerivedTyID;
233 /// Obtain a type given a typeid and account for things like compaction tables,
234 /// function level vs module level, and the offsetting for the primitive types.
235 const Type *BytecodeReader::getType(unsigned ID) {
236 if (ID < Type::FirstDerivedTyID)
237 if (const Type *T = Type::getPrimitiveType((Type::TypeID)ID))
238 return T; // Asked for a primitive type...
240 // Otherwise, derived types need offset...
241 ID -= Type::FirstDerivedTyID;
243 if (!CompactionTypes.empty()) {
244 if (ID >= CompactionTypes.size())
245 error("Type ID out of range for compaction table!");
246 return CompactionTypes[ID];
249 // Is it a module-level type?
250 if (ID < ModuleTypes.size())
251 return ModuleTypes[ID].get();
253 // Nope, is it a function-level type?
254 ID -= ModuleTypes.size();
255 if (ID < FunctionTypes.size())
256 return FunctionTypes[ID].get();
258 error("Illegal type reference!");
262 /// Get a sanitized type id. This just makes sure that the \p ID
263 /// is both sanitized and not the "type type" of pre-1.3 bytecode.
264 /// @see sanitizeTypeId
265 inline const Type* BytecodeReader::getSanitizedType(unsigned& ID) {
266 if (sanitizeTypeId(ID))
267 error("Invalid type id encountered");
271 /// This method just saves some coding. It uses read_typeid to read
272 /// in a sanitized type id, errors that its not the type type, and
273 /// then calls getType to return the type value.
274 inline const Type* BytecodeReader::readSanitizedType() {
277 error("Invalid type id encountered");
281 /// Get the slot number associated with a type accounting for primitive
282 /// types, compaction tables, and function level vs module level.
283 unsigned BytecodeReader::getTypeSlot(const Type *Ty) {
284 if (Ty->isPrimitiveType())
285 return Ty->getTypeID();
287 // Scan the compaction table for the type if needed.
288 if (!CompactionTypes.empty()) {
289 std::vector<const Type*>::const_iterator I =
290 find(CompactionTypes.begin(), CompactionTypes.end(), Ty);
292 if (I == CompactionTypes.end())
293 error("Couldn't find type specified in compaction table!");
294 return Type::FirstDerivedTyID + (&*I - &CompactionTypes[0]);
297 // Check the function level types first...
298 TypeListTy::iterator I = find(FunctionTypes.begin(), FunctionTypes.end(), Ty);
300 if (I != FunctionTypes.end())
301 return Type::FirstDerivedTyID + ModuleTypes.size() +
302 (&*I - &FunctionTypes[0]);
304 // Check the module level types now...
305 I = find(ModuleTypes.begin(), ModuleTypes.end(), Ty);
306 if (I == ModuleTypes.end())
307 error("Didn't find type in ModuleTypes.");
308 return Type::FirstDerivedTyID + (&*I - &ModuleTypes[0]);
311 /// This is just like getType, but when a compaction table is in use, it is
312 /// ignored. It also ignores function level types.
314 const Type *BytecodeReader::getGlobalTableType(unsigned Slot) {
315 if (Slot < Type::FirstDerivedTyID) {
316 const Type *Ty = Type::getPrimitiveType((Type::TypeID)Slot);
318 error("Not a primitive type ID?");
321 Slot -= Type::FirstDerivedTyID;
322 if (Slot >= ModuleTypes.size())
323 error("Illegal compaction table type reference!");
324 return ModuleTypes[Slot];
327 /// This is just like getTypeSlot, but when a compaction table is in use, it
328 /// is ignored. It also ignores function level types.
329 unsigned BytecodeReader::getGlobalTableTypeSlot(const Type *Ty) {
330 if (Ty->isPrimitiveType())
331 return Ty->getTypeID();
332 TypeListTy::iterator I = find(ModuleTypes.begin(),
333 ModuleTypes.end(), Ty);
334 if (I == ModuleTypes.end())
335 error("Didn't find type in ModuleTypes.");
336 return Type::FirstDerivedTyID + (&*I - &ModuleTypes[0]);
339 /// Retrieve a value of a given type and slot number, possibly creating
340 /// it if it doesn't already exist.
341 Value * BytecodeReader::getValue(unsigned type, unsigned oNum, bool Create) {
342 assert(type != Type::LabelTyID && "getValue() cannot get blocks!");
345 // If there is a compaction table active, it defines the low-level numbers.
346 // If not, the module values define the low-level numbers.
347 if (CompactionValues.size() > type && !CompactionValues[type].empty()) {
348 if (Num < CompactionValues[type].size())
349 return CompactionValues[type][Num];
350 Num -= CompactionValues[type].size();
352 // By default, the global type id is the type id passed in
353 unsigned GlobalTyID = type;
355 // If the type plane was compactified, figure out the global type ID
356 // by adding the derived type ids and the distance.
357 if (!CompactionTypes.empty() && type >= Type::FirstDerivedTyID) {
358 const Type *Ty = CompactionTypes[type-Type::FirstDerivedTyID];
359 TypeListTy::iterator I =
360 find(ModuleTypes.begin(), ModuleTypes.end(), Ty);
361 assert(I != ModuleTypes.end());
362 GlobalTyID = Type::FirstDerivedTyID + (&*I - &ModuleTypes[0]);
365 if (hasImplicitNull(GlobalTyID)) {
367 return Constant::getNullValue(getType(type));
371 if (GlobalTyID < ModuleValues.size() && ModuleValues[GlobalTyID]) {
372 if (Num < ModuleValues[GlobalTyID]->size())
373 return ModuleValues[GlobalTyID]->getOperand(Num);
374 Num -= ModuleValues[GlobalTyID]->size();
378 if (FunctionValues.size() > type &&
379 FunctionValues[type] &&
380 Num < FunctionValues[type]->size())
381 return FunctionValues[type]->getOperand(Num);
383 if (!Create) return 0; // Do not create a placeholder?
385 std::pair<unsigned,unsigned> KeyValue(type, oNum);
386 ForwardReferenceMap::iterator I = ForwardReferences.lower_bound(KeyValue);
387 if (I != ForwardReferences.end() && I->first == KeyValue)
388 return I->second; // We have already created this placeholder
390 Value *Val = new Argument(getType(type));
391 ForwardReferences.insert(I, std::make_pair(KeyValue, Val));
395 /// This is just like getValue, but when a compaction table is in use, it
396 /// is ignored. Also, no forward references or other fancy features are
398 Value* BytecodeReader::getGlobalTableValue(const Type *Ty, unsigned SlotNo) {
399 // FIXME: getTypeSlot is inefficient!
400 unsigned TyID = getGlobalTableTypeSlot(Ty);
402 if (TyID != Type::LabelTyID) {
404 return Constant::getNullValue(Ty);
408 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0 ||
409 SlotNo >= ModuleValues[TyID]->size()) {
410 error("Corrupt compaction table entry!"
411 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
412 + utostr(ModuleValues.size()) + ", "
413 + utohexstr(intptr_t((void*)ModuleValues[TyID])) + ", "
414 + utostr(ModuleValues[TyID]->size()));
416 return ModuleValues[TyID]->getOperand(SlotNo);
419 /// Just like getValue, except that it returns a null pointer
420 /// only on error. It always returns a constant (meaning that if the value is
421 /// defined, but is not a constant, that is an error). If the specified
422 /// constant hasn't been parsed yet, a placeholder is defined and used.
423 /// Later, after the real value is parsed, the placeholder is eliminated.
424 Constant* BytecodeReader::getConstantValue(unsigned TypeSlot, unsigned Slot) {
425 if (Value *V = getValue(TypeSlot, Slot, false))
426 if (Constant *C = dyn_cast<Constant>(V))
427 return C; // If we already have the value parsed, just return it
429 error("Value for slot " + utostr(Slot) +
430 " is expected to be a constant!");
432 const Type *Ty = getType(TypeSlot);
433 std::pair<const Type*, unsigned> Key(Ty, Slot);
434 ConstantRefsType::iterator I = ConstantFwdRefs.lower_bound(Key);
436 if (I != ConstantFwdRefs.end() && I->first == Key) {
439 // Create a placeholder for the constant reference and
440 // keep track of the fact that we have a forward ref to recycle it
441 Constant *C = new ConstantPlaceHolder(Ty, Slot);
443 // Keep track of the fact that we have a forward ref to recycle it
444 ConstantFwdRefs.insert(I, std::make_pair(Key, C));
449 //===----------------------------------------------------------------------===//
450 // IR Construction Methods
451 //===----------------------------------------------------------------------===//
453 /// As values are created, they are inserted into the appropriate place
454 /// with this method. The ValueTable argument must be one of ModuleValues
455 /// or FunctionValues data members of this class.
456 unsigned BytecodeReader::insertValue(Value *Val, unsigned type,
457 ValueTable &ValueTab) {
458 assert((!isa<Constant>(Val) || !cast<Constant>(Val)->isNullValue()) ||
459 !hasImplicitNull(type) &&
460 "Cannot read null values from bytecode!");
462 if (ValueTab.size() <= type)
463 ValueTab.resize(type+1);
465 if (!ValueTab[type]) ValueTab[type] = new ValueList();
467 ValueTab[type]->push_back(Val);
469 bool HasOffset = hasImplicitNull(type);
470 return ValueTab[type]->size()-1 + HasOffset;
473 /// Insert the arguments of a function as new values in the reader.
474 void BytecodeReader::insertArguments(Function* F) {
475 const FunctionType *FT = F->getFunctionType();
476 Function::aiterator AI = F->abegin();
477 for (FunctionType::param_iterator It = FT->param_begin();
478 It != FT->param_end(); ++It, ++AI)
479 insertValue(AI, getTypeSlot(AI->getType()), FunctionValues);
482 //===----------------------------------------------------------------------===//
483 // Bytecode Parsing Methods
484 //===----------------------------------------------------------------------===//
486 /// This method parses a single instruction. The instruction is
487 /// inserted at the end of the \p BB provided. The arguments of
488 /// the instruction are provided in the \p Args vector.
489 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds,
493 // Clear instruction data
497 unsigned Op = read_uint();
499 // bits Instruction format: Common to all formats
500 // --------------------------
501 // 01-00: Opcode type, fixed to 1.
503 Opcode = (Op >> 2) & 63;
504 Oprnds.resize((Op >> 0) & 03);
506 // Extract the operands
507 switch (Oprnds.size()) {
509 // bits Instruction format:
510 // --------------------------
511 // 19-08: Resulting type plane
512 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
514 iType = (Op >> 8) & 4095;
515 Oprnds[0] = (Op >> 20) & 4095;
516 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands...
520 // bits Instruction format:
521 // --------------------------
522 // 15-08: Resulting type plane
526 iType = (Op >> 8) & 255;
527 Oprnds[0] = (Op >> 16) & 255;
528 Oprnds[1] = (Op >> 24) & 255;
531 // bits Instruction format:
532 // --------------------------
533 // 13-08: Resulting type plane
538 iType = (Op >> 8) & 63;
539 Oprnds[0] = (Op >> 14) & 63;
540 Oprnds[1] = (Op >> 20) & 63;
541 Oprnds[2] = (Op >> 26) & 63;
544 At -= 4; // Hrm, try this again...
545 Opcode = read_vbr_uint();
547 iType = read_vbr_uint();
549 unsigned NumOprnds = read_vbr_uint();
550 Oprnds.resize(NumOprnds);
553 error("Zero-argument instruction found; this is invalid.");
555 for (unsigned i = 0; i != NumOprnds; ++i)
556 Oprnds[i] = read_vbr_uint();
561 const Type *InstTy = getSanitizedType(iType);
563 // We have enough info to inform the handler now.
564 if (Handler) Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt);
566 // Declare the resulting instruction we'll build.
567 Instruction *Result = 0;
569 // Handle binary operators
570 if (Opcode >= Instruction::BinaryOpsBegin &&
571 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2)
572 Result = BinaryOperator::create((Instruction::BinaryOps)Opcode,
573 getValue(iType, Oprnds[0]),
574 getValue(iType, Oprnds[1]));
579 error("Illegal instruction read!");
581 case Instruction::VAArg:
582 Result = new VAArgInst(getValue(iType, Oprnds[0]),
583 getSanitizedType(Oprnds[1]));
585 case Instruction::VANext:
586 Result = new VANextInst(getValue(iType, Oprnds[0]),
587 getSanitizedType(Oprnds[1]));
589 case Instruction::Cast:
590 Result = new CastInst(getValue(iType, Oprnds[0]),
591 getSanitizedType(Oprnds[1]));
593 case Instruction::Select:
594 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]),
595 getValue(iType, Oprnds[1]),
596 getValue(iType, Oprnds[2]));
598 case Instruction::PHI: {
599 if (Oprnds.size() == 0 || (Oprnds.size() & 1))
600 error("Invalid phi node encountered!");
602 PHINode *PN = new PHINode(InstTy);
603 PN->op_reserve(Oprnds.size());
604 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2)
605 PN->addIncoming(getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1]));
610 case Instruction::Shl:
611 case Instruction::Shr:
612 Result = new ShiftInst((Instruction::OtherOps)Opcode,
613 getValue(iType, Oprnds[0]),
614 getValue(Type::UByteTyID, Oprnds[1]));
616 case Instruction::Ret:
617 if (Oprnds.size() == 0)
618 Result = new ReturnInst();
619 else if (Oprnds.size() == 1)
620 Result = new ReturnInst(getValue(iType, Oprnds[0]));
622 error("Unrecognized instruction!");
625 case Instruction::Br:
626 if (Oprnds.size() == 1)
627 Result = new BranchInst(getBasicBlock(Oprnds[0]));
628 else if (Oprnds.size() == 3)
629 Result = new BranchInst(getBasicBlock(Oprnds[0]),
630 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2]));
632 error("Invalid number of operands for a 'br' instruction!");
634 case Instruction::Switch: {
635 if (Oprnds.size() & 1)
636 error("Switch statement with odd number of arguments!");
638 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]),
639 getBasicBlock(Oprnds[1]));
640 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2)
641 I->addCase(cast<Constant>(getValue(iType, Oprnds[i])),
642 getBasicBlock(Oprnds[i+1]));
647 case Instruction::Call: {
648 if (Oprnds.size() == 0)
649 error("Invalid call instruction encountered!");
651 Value *F = getValue(iType, Oprnds[0]);
653 // Check to make sure we have a pointer to function type
654 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
655 if (PTy == 0) error("Call to non function pointer value!");
656 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
657 if (FTy == 0) error("Call to non function pointer value!");
659 std::vector<Value *> Params;
660 if (!FTy->isVarArg()) {
661 FunctionType::param_iterator It = FTy->param_begin();
663 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
664 if (It == FTy->param_end())
665 error("Invalid call instruction!");
666 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
668 if (It != FTy->param_end())
669 error("Invalid call instruction!");
671 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
673 unsigned FirstVariableOperand;
674 if (Oprnds.size() < FTy->getNumParams())
675 error("Call instruction missing operands!");
677 // Read all of the fixed arguments
678 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
679 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i]));
681 FirstVariableOperand = FTy->getNumParams();
683 if ((Oprnds.size()-FirstVariableOperand) & 1) // Must be pairs of type/value
684 error("Invalid call instruction!");
686 for (unsigned i = FirstVariableOperand, e = Oprnds.size();
688 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
691 Result = new CallInst(F, Params);
694 case Instruction::Invoke: {
695 if (Oprnds.size() < 3)
696 error("Invalid invoke instruction!");
697 Value *F = getValue(iType, Oprnds[0]);
699 // Check to make sure we have a pointer to function type
700 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
702 error("Invoke to non function pointer value!");
703 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
705 error("Invoke to non function pointer value!");
707 std::vector<Value *> Params;
708 BasicBlock *Normal, *Except;
710 if (!FTy->isVarArg()) {
711 Normal = getBasicBlock(Oprnds[1]);
712 Except = getBasicBlock(Oprnds[2]);
714 FunctionType::param_iterator It = FTy->param_begin();
715 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) {
716 if (It == FTy->param_end())
717 error("Invalid invoke instruction!");
718 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
720 if (It != FTy->param_end())
721 error("Invalid invoke instruction!");
723 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
725 Normal = getBasicBlock(Oprnds[0]);
726 Except = getBasicBlock(Oprnds[1]);
728 unsigned FirstVariableArgument = FTy->getNumParams()+2;
729 for (unsigned i = 2; i != FirstVariableArgument; ++i)
730 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)),
733 if (Oprnds.size()-FirstVariableArgument & 1) // Must be type/value pairs
734 error("Invalid invoke instruction!");
736 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2)
737 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
740 Result = new InvokeInst(F, Normal, Except, Params);
743 case Instruction::Malloc:
744 if (Oprnds.size() > 2)
745 error("Invalid malloc instruction!");
746 if (!isa<PointerType>(InstTy))
747 error("Invalid malloc instruction!");
749 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(),
750 Oprnds.size() ? getValue(Type::UIntTyID,
754 case Instruction::Alloca:
755 if (Oprnds.size() > 2)
756 error("Invalid alloca instruction!");
757 if (!isa<PointerType>(InstTy))
758 error("Invalid alloca instruction!");
760 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(),
761 Oprnds.size() ? getValue(Type::UIntTyID,
764 case Instruction::Free:
765 if (!isa<PointerType>(InstTy))
766 error("Invalid free instruction!");
767 Result = new FreeInst(getValue(iType, Oprnds[0]));
769 case Instruction::GetElementPtr: {
770 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy))
771 error("Invalid getelementptr instruction!");
773 std::vector<Value*> Idx;
775 const Type *NextTy = InstTy;
776 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
777 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy);
779 error("Invalid getelementptr instruction!");
781 unsigned ValIdx = Oprnds[i];
783 if (!hasRestrictedGEPTypes) {
784 // Struct indices are always uints, sequential type indices can be any
785 // of the 32 or 64-bit integer types. The actual choice of type is
786 // encoded in the low two bits of the slot number.
787 if (isa<StructType>(TopTy))
788 IdxTy = Type::UIntTyID;
790 switch (ValIdx & 3) {
792 case 0: IdxTy = Type::UIntTyID; break;
793 case 1: IdxTy = Type::IntTyID; break;
794 case 2: IdxTy = Type::ULongTyID; break;
795 case 3: IdxTy = Type::LongTyID; break;
800 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID;
803 Idx.push_back(getValue(IdxTy, ValIdx));
805 // Convert ubyte struct indices into uint struct indices.
806 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes)
807 if (ConstantUInt *C = dyn_cast<ConstantUInt>(Idx.back()))
808 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy);
810 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true);
813 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx);
817 case 62: // volatile load
818 case Instruction::Load:
819 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy))
820 error("Invalid load instruction!");
821 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62);
824 case 63: // volatile store
825 case Instruction::Store: {
826 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2)
827 error("Invalid store instruction!");
829 Value *Ptr = getValue(iType, Oprnds[1]);
830 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType();
831 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr,
835 case Instruction::Unwind:
836 if (Oprnds.size() != 0)
837 error("Invalid unwind instruction!");
838 Result = new UnwindInst();
840 } // end switch(Opcode)
843 if (Result->getType() == InstTy)
846 TypeSlot = getTypeSlot(Result->getType());
848 insertValue(Result, TypeSlot, FunctionValues);
849 BB->getInstList().push_back(Result);
852 /// Get a particular numbered basic block, which might be a forward reference.
853 /// This works together with ParseBasicBlock to handle these forward references
854 /// in a clean manner. This function is used when constructing phi, br, switch,
855 /// and other instructions that reference basic blocks. Blocks are numbered
856 /// sequentially as they appear in the function.
857 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) {
858 // Make sure there is room in the table...
859 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1);
861 // First check to see if this is a backwards reference, i.e., ParseBasicBlock
862 // has already created this block, or if the forward reference has already
864 if (ParsedBasicBlocks[ID])
865 return ParsedBasicBlocks[ID];
867 // Otherwise, the basic block has not yet been created. Do so and add it to
868 // the ParsedBasicBlocks list.
869 return ParsedBasicBlocks[ID] = new BasicBlock();
872 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time.
873 /// This method reads in one of the basicblock packets. This method is not used
874 /// for bytecode files after LLVM 1.0
875 /// @returns The basic block constructed.
876 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) {
877 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
881 if (ParsedBasicBlocks.size() == BlockNo)
882 ParsedBasicBlocks.push_back(BB = new BasicBlock());
883 else if (ParsedBasicBlocks[BlockNo] == 0)
884 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
886 BB = ParsedBasicBlocks[BlockNo];
888 std::vector<unsigned> Operands;
889 while (moreInBlock())
890 ParseInstruction(Operands, BB);
892 if (Handler) Handler->handleBasicBlockEnd(BlockNo);
896 /// Parse all of the BasicBlock's & Instruction's in the body of a function.
897 /// In post 1.0 bytecode files, we no longer emit basic block individually,
898 /// in order to avoid per-basic-block overhead.
899 /// @returns Rhe number of basic blocks encountered.
900 unsigned BytecodeReader::ParseInstructionList(Function* F) {
901 unsigned BlockNo = 0;
902 std::vector<unsigned> Args;
904 while (moreInBlock()) {
905 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
907 if (ParsedBasicBlocks.size() == BlockNo)
908 ParsedBasicBlocks.push_back(BB = new BasicBlock());
909 else if (ParsedBasicBlocks[BlockNo] == 0)
910 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
912 BB = ParsedBasicBlocks[BlockNo];
914 F->getBasicBlockList().push_back(BB);
916 // Read instructions into this basic block until we get to a terminator
917 while (moreInBlock() && !BB->getTerminator())
918 ParseInstruction(Args, BB);
920 if (!BB->getTerminator())
921 error("Non-terminated basic block found!");
923 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1);
929 /// Parse a symbol table. This works for both module level and function
930 /// level symbol tables. For function level symbol tables, the CurrentFunction
931 /// parameter must be non-zero and the ST parameter must correspond to
932 /// CurrentFunction's symbol table. For Module level symbol tables, the
933 /// CurrentFunction argument must be zero.
934 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction,
936 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST);
938 // Allow efficient basic block lookup by number.
939 std::vector<BasicBlock*> BBMap;
941 for (Function::iterator I = CurrentFunction->begin(),
942 E = CurrentFunction->end(); I != E; ++I)
945 /// In LLVM 1.3 we write types separately from values so
946 /// The types are always first in the symbol table. This is
947 /// because Type no longer derives from Value.
948 if (!hasTypeDerivedFromValue) {
949 // Symtab block header: [num entries]
950 unsigned NumEntries = read_vbr_uint();
951 for (unsigned i = 0; i < NumEntries; ++i) {
952 // Symtab entry: [def slot #][name]
953 unsigned slot = read_vbr_uint();
954 std::string Name = read_str();
955 const Type* T = getType(slot);
960 while (moreInBlock()) {
961 // Symtab block header: [num entries][type id number]
962 unsigned NumEntries = read_vbr_uint();
964 bool isTypeType = read_typeid(Typ);
965 const Type *Ty = getType(Typ);
967 for (unsigned i = 0; i != NumEntries; ++i) {
968 // Symtab entry: [def slot #][name]
969 unsigned slot = read_vbr_uint();
970 std::string Name = read_str();
972 // if we're reading a pre 1.3 bytecode file and the type plane
973 // is the "type type", handle it here
975 const Type* T = getType(slot);
977 error("Failed type look-up for name '" + Name + "'");
979 continue; // code below must be short circuited
982 if (Typ == Type::LabelTyID) {
983 if (slot < BBMap.size())
986 V = getValue(Typ, slot, false); // Find mapping...
989 error("Failed value look-up for name '" + Name + "'");
990 V->setName(Name, ST);
994 checkPastBlockEnd("Symbol Table");
995 if (Handler) Handler->handleSymbolTableEnd();
998 /// Read in the types portion of a compaction table.
999 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) {
1000 for (unsigned i = 0; i != NumEntries; ++i) {
1001 unsigned TypeSlot = 0;
1002 if (read_typeid(TypeSlot))
1003 error("Invalid type in compaction table: type type");
1004 const Type *Typ = getGlobalTableType(TypeSlot);
1005 CompactionTypes.push_back(Typ);
1006 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ);
1010 /// Parse a compaction table.
1011 void BytecodeReader::ParseCompactionTable() {
1013 // Notify handler that we're beginning a compaction table.
1014 if (Handler) Handler->handleCompactionTableBegin();
1016 // In LLVM 1.3 Type no longer derives from Value. So,
1017 // we always write them first in the compaction table
1018 // because they can't occupy a "type plane" where the
1020 if (! hasTypeDerivedFromValue) {
1021 unsigned NumEntries = read_vbr_uint();
1022 ParseCompactionTypes(NumEntries);
1025 // Compaction tables live in separate blocks so we have to loop
1026 // until we've read the whole thing.
1027 while (moreInBlock()) {
1028 // Read the number of Value* entries in the compaction table
1029 unsigned NumEntries = read_vbr_uint();
1031 unsigned isTypeType = false;
1033 // Decode the type from value read in. Most compaction table
1034 // planes will have one or two entries in them. If that's the
1035 // case then the length is encoded in the bottom two bits and
1036 // the higher bits encode the type. This saves another VBR value.
1037 if ((NumEntries & 3) == 3) {
1038 // In this case, both low-order bits are set (value 3). This
1039 // is a signal that the typeid follows.
1041 isTypeType = read_typeid(Ty);
1043 // In this case, the low-order bits specify the number of entries
1044 // and the high order bits specify the type.
1045 Ty = NumEntries >> 2;
1046 isTypeType = sanitizeTypeId(Ty);
1050 // if we're reading a pre 1.3 bytecode file and the type plane
1051 // is the "type type", handle it here
1053 ParseCompactionTypes(NumEntries);
1055 // Make sure we have enough room for the plane
1056 if (Ty >= CompactionValues.size())
1057 CompactionValues.resize(Ty+1);
1059 // Make sure the plane is empty or we have some kind of error
1060 if (!CompactionValues[Ty].empty())
1061 error("Compaction table plane contains multiple entries!");
1063 // Notify handler about the plane
1064 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries);
1066 // Convert the type slot to a type
1067 const Type *Typ = getType(Ty);
1069 // Push the implicit zero
1070 CompactionValues[Ty].push_back(Constant::getNullValue(Typ));
1072 // Read in each of the entries, put them in the compaction table
1073 // and notify the handler that we have a new compaction table value.
1074 for (unsigned i = 0; i != NumEntries; ++i) {
1075 unsigned ValSlot = read_vbr_uint();
1076 Value *V = getGlobalTableValue(Typ, ValSlot);
1077 CompactionValues[Ty].push_back(V);
1078 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot, Typ);
1082 // Notify handler that the compaction table is done.
1083 if (Handler) Handler->handleCompactionTableEnd();
1086 // Parse a single type. The typeid is read in first. If its a primitive type
1087 // then nothing else needs to be read, we know how to instantiate it. If its
1088 // a derived type, then additional data is read to fill out the type
1090 const Type *BytecodeReader::ParseType() {
1091 unsigned PrimType = 0;
1092 if (read_typeid(PrimType))
1093 error("Invalid type (type type) in type constants!");
1095 const Type *Result = 0;
1096 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType)))
1100 case Type::FunctionTyID: {
1101 const Type *RetType = readSanitizedType();
1103 unsigned NumParams = read_vbr_uint();
1105 std::vector<const Type*> Params;
1107 Params.push_back(readSanitizedType());
1109 bool isVarArg = Params.size() && Params.back() == Type::VoidTy;
1110 if (isVarArg) Params.pop_back();
1112 Result = FunctionType::get(RetType, Params, isVarArg);
1115 case Type::ArrayTyID: {
1116 const Type *ElementType = readSanitizedType();
1117 unsigned NumElements = read_vbr_uint();
1118 Result = ArrayType::get(ElementType, NumElements);
1121 case Type::StructTyID: {
1122 std::vector<const Type*> Elements;
1124 if (read_typeid(Typ))
1125 error("Invalid element type (type type) for structure!");
1127 while (Typ) { // List is terminated by void/0 typeid
1128 Elements.push_back(getType(Typ));
1129 if (read_typeid(Typ))
1130 error("Invalid element type (type type) for structure!");
1133 Result = StructType::get(Elements);
1136 case Type::PointerTyID: {
1137 Result = PointerType::get(readSanitizedType());
1141 case Type::OpaqueTyID: {
1142 Result = OpaqueType::get();
1147 error("Don't know how to deserialize primitive type " + utostr(PrimType));
1150 if (Handler) Handler->handleType(Result);
1154 // ParseType - We have to use this weird code to handle recursive
1155 // types. We know that recursive types will only reference the current slab of
1156 // values in the type plane, but they can forward reference types before they
1157 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might
1158 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix
1159 // this ugly problem, we pessimistically insert an opaque type for each type we
1160 // are about to read. This means that forward references will resolve to
1161 // something and when we reread the type later, we can replace the opaque type
1162 // with a new resolved concrete type.
1164 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){
1165 assert(Tab.size() == 0 && "should not have read type constants in before!");
1167 // Insert a bunch of opaque types to be resolved later...
1168 Tab.reserve(NumEntries);
1169 for (unsigned i = 0; i != NumEntries; ++i)
1170 Tab.push_back(OpaqueType::get());
1172 // Loop through reading all of the types. Forward types will make use of the
1173 // opaque types just inserted.
1175 for (unsigned i = 0; i != NumEntries; ++i) {
1176 const Type* NewTy = ParseType();
1177 const Type* OldTy = Tab[i].get();
1179 error("Couldn't parse type!");
1181 // Don't directly push the new type on the Tab. Instead we want to replace
1182 // the opaque type we previously inserted with the new concrete value. This
1183 // approach helps with forward references to types. The refinement from the
1184 // abstract (opaque) type to the new type causes all uses of the abstract
1185 // type to use the concrete type (NewTy). This will also cause the opaque
1186 // type to be deleted.
1187 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy);
1189 // This should have replaced the old opaque type with the new type in the
1190 // value table... or with a preexisting type that was already in the system.
1191 // Let's just make sure it did.
1192 assert(Tab[i] != OldTy && "refineAbstractType didn't work!");
1196 /// Parse a single constant value
1197 Constant *BytecodeReader::ParseConstantValue(unsigned TypeID) {
1198 // We must check for a ConstantExpr before switching by type because
1199 // a ConstantExpr can be of any type, and has no explicit value.
1201 // 0 if not expr; numArgs if is expr
1202 unsigned isExprNumArgs = read_vbr_uint();
1204 if (isExprNumArgs) {
1205 // FIXME: Encoding of constant exprs could be much more compact!
1206 std::vector<Constant*> ArgVec;
1207 ArgVec.reserve(isExprNumArgs);
1208 unsigned Opcode = read_vbr_uint();
1210 // Read the slot number and types of each of the arguments
1211 for (unsigned i = 0; i != isExprNumArgs; ++i) {
1212 unsigned ArgValSlot = read_vbr_uint();
1213 unsigned ArgTypeSlot = 0;
1214 if (read_typeid(ArgTypeSlot))
1215 error("Invalid argument type (type type) for constant value");
1217 // Get the arg value from its slot if it exists, otherwise a placeholder
1218 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot));
1221 // Construct a ConstantExpr of the appropriate kind
1222 if (isExprNumArgs == 1) { // All one-operand expressions
1223 if (Opcode != Instruction::Cast)
1224 error("Only Cast instruction has one argument for ConstantExpr");
1226 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID));
1227 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1229 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr
1230 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end());
1232 if (hasRestrictedGEPTypes) {
1233 const Type *BaseTy = ArgVec[0]->getType();
1234 generic_gep_type_iterator<std::vector<Constant*>::iterator>
1235 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()),
1236 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end());
1237 for (unsigned i = 0; GTI != E; ++GTI, ++i)
1238 if (isa<StructType>(*GTI)) {
1239 if (IdxList[i]->getType() != Type::UByteTy)
1240 error("Invalid index for getelementptr!");
1241 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy);
1245 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList);
1246 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1248 } else if (Opcode == Instruction::Select) {
1249 if (ArgVec.size() != 3)
1250 error("Select instruction must have three arguments.");
1251 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1],
1253 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1255 } else { // All other 2-operand expressions
1256 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]);
1257 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1262 // Ok, not an ConstantExpr. We now know how to read the given type...
1263 const Type *Ty = getType(TypeID);
1264 switch (Ty->getTypeID()) {
1265 case Type::BoolTyID: {
1266 unsigned Val = read_vbr_uint();
1267 if (Val != 0 && Val != 1)
1268 error("Invalid boolean value read.");
1269 Constant* Result = ConstantBool::get(Val == 1);
1270 if (Handler) Handler->handleConstantValue(Result);
1274 case Type::UByteTyID: // Unsigned integer types...
1275 case Type::UShortTyID:
1276 case Type::UIntTyID: {
1277 unsigned Val = read_vbr_uint();
1278 if (!ConstantUInt::isValueValidForType(Ty, Val))
1279 error("Invalid unsigned byte/short/int read.");
1280 Constant* Result = ConstantUInt::get(Ty, Val);
1281 if (Handler) Handler->handleConstantValue(Result);
1285 case Type::ULongTyID: {
1286 Constant* Result = ConstantUInt::get(Ty, read_vbr_uint64());
1287 if (Handler) Handler->handleConstantValue(Result);
1291 case Type::SByteTyID: // Signed integer types...
1292 case Type::ShortTyID:
1293 case Type::IntTyID: {
1294 case Type::LongTyID:
1295 int64_t Val = read_vbr_int64();
1296 if (!ConstantSInt::isValueValidForType(Ty, Val))
1297 error("Invalid signed byte/short/int/long read.");
1298 Constant* Result = ConstantSInt::get(Ty, Val);
1299 if (Handler) Handler->handleConstantValue(Result);
1303 case Type::FloatTyID: {
1306 Constant* Result = ConstantFP::get(Ty, Val);
1307 if (Handler) Handler->handleConstantValue(Result);
1311 case Type::DoubleTyID: {
1314 Constant* Result = ConstantFP::get(Ty, Val);
1315 if (Handler) Handler->handleConstantValue(Result);
1319 case Type::ArrayTyID: {
1320 const ArrayType *AT = cast<ArrayType>(Ty);
1321 unsigned NumElements = AT->getNumElements();
1322 unsigned TypeSlot = getTypeSlot(AT->getElementType());
1323 std::vector<Constant*> Elements;
1324 Elements.reserve(NumElements);
1325 while (NumElements--) // Read all of the elements of the constant.
1326 Elements.push_back(getConstantValue(TypeSlot,
1328 Constant* Result = ConstantArray::get(AT, Elements);
1329 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result);
1333 case Type::StructTyID: {
1334 const StructType *ST = cast<StructType>(Ty);
1336 std::vector<Constant *> Elements;
1337 Elements.reserve(ST->getNumElements());
1338 for (unsigned i = 0; i != ST->getNumElements(); ++i)
1339 Elements.push_back(getConstantValue(ST->getElementType(i),
1342 Constant* Result = ConstantStruct::get(ST, Elements);
1343 if (Handler) Handler->handleConstantStruct(ST, Elements, Result);
1347 case Type::PointerTyID: { // ConstantPointerRef value...
1348 const PointerType *PT = cast<PointerType>(Ty);
1349 unsigned Slot = read_vbr_uint();
1351 // Check to see if we have already read this global variable...
1352 Value *Val = getValue(TypeID, Slot, false);
1355 if (!(GV = dyn_cast<GlobalValue>(Val)))
1356 error("GlobalValue not in ValueTable!");
1358 error("Forward references are not allowed here.");
1361 if (Handler) Handler->handleConstantPointer(PT, Slot, GV );
1366 error("Don't know how to deserialize constant value of type '" +
1367 Ty->getDescription());
1373 /// Resolve references for constants. This function resolves the forward
1374 /// referenced constants in the ConstantFwdRefs map. It uses the
1375 /// replaceAllUsesWith method of Value class to substitute the placeholder
1376 /// instance with the actual instance.
1377 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Slot){
1378 ConstantRefsType::iterator I =
1379 ConstantFwdRefs.find(std::make_pair(NewV->getType(), Slot));
1380 if (I == ConstantFwdRefs.end()) return; // Never forward referenced?
1382 Value *PH = I->second; // Get the placeholder...
1383 PH->replaceAllUsesWith(NewV);
1384 delete PH; // Delete the old placeholder
1385 ConstantFwdRefs.erase(I); // Remove the map entry for it
1388 /// Parse the constant strings section.
1389 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){
1390 for (; NumEntries; --NumEntries) {
1392 if (read_typeid(Typ))
1393 error("Invalid type (type type) for string constant");
1394 const Type *Ty = getType(Typ);
1395 if (!isa<ArrayType>(Ty))
1396 error("String constant data invalid!");
1398 const ArrayType *ATy = cast<ArrayType>(Ty);
1399 if (ATy->getElementType() != Type::SByteTy &&
1400 ATy->getElementType() != Type::UByteTy)
1401 error("String constant data invalid!");
1403 // Read character data. The type tells us how long the string is.
1404 char Data[ATy->getNumElements()];
1405 read_data(Data, Data+ATy->getNumElements());
1407 std::vector<Constant*> Elements(ATy->getNumElements());
1408 if (ATy->getElementType() == Type::SByteTy)
1409 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1410 Elements[i] = ConstantSInt::get(Type::SByteTy, (signed char)Data[i]);
1412 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1413 Elements[i] = ConstantUInt::get(Type::UByteTy, (unsigned char)Data[i]);
1415 // Create the constant, inserting it as needed.
1416 Constant *C = ConstantArray::get(ATy, Elements);
1417 unsigned Slot = insertValue(C, Typ, Tab);
1418 ResolveReferencesToConstant(C, Slot);
1419 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C));
1423 /// Parse the constant pool.
1424 void BytecodeReader::ParseConstantPool(ValueTable &Tab,
1425 TypeListTy &TypeTab,
1427 if (Handler) Handler->handleGlobalConstantsBegin();
1429 /// In LLVM 1.3 Type does not derive from Value so the types
1430 /// do not occupy a plane. Consequently, we read the types
1431 /// first in the constant pool.
1432 if (isFunction && !hasTypeDerivedFromValue) {
1433 unsigned NumEntries = read_vbr_uint();
1434 ParseTypes(TypeTab, NumEntries);
1437 while (moreInBlock()) {
1438 unsigned NumEntries = read_vbr_uint();
1440 bool isTypeType = read_typeid(Typ);
1442 /// In LLVM 1.2 and before, Types were written to the
1443 /// bytecode file in the "Type Type" plane (#12).
1444 /// In 1.3 plane 12 is now the label plane. Handle this here.
1446 ParseTypes(TypeTab, NumEntries);
1447 } else if (Typ == Type::VoidTyID) {
1448 /// Use of Type::VoidTyID is a misnomer. It actually means
1449 /// that the following plane is constant strings
1450 assert(&Tab == &ModuleValues && "Cannot read strings in functions!");
1451 ParseStringConstants(NumEntries, Tab);
1453 for (unsigned i = 0; i < NumEntries; ++i) {
1454 Constant *C = ParseConstantValue(Typ);
1455 assert(C && "ParseConstantValue returned NULL!");
1456 unsigned Slot = insertValue(C, Typ, Tab);
1458 // If we are reading a function constant table, make sure that we adjust
1459 // the slot number to be the real global constant number.
1461 if (&Tab != &ModuleValues && Typ < ModuleValues.size() &&
1463 Slot += ModuleValues[Typ]->size();
1464 ResolveReferencesToConstant(C, Slot);
1468 checkPastBlockEnd("Constant Pool");
1469 if (Handler) Handler->handleGlobalConstantsEnd();
1472 /// Parse the contents of a function. Note that this function can be
1473 /// called lazily by materializeFunction
1474 /// @see materializeFunction
1475 void BytecodeReader::ParseFunctionBody(Function* F) {
1477 unsigned FuncSize = BlockEnd - At;
1478 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage;
1480 unsigned LinkageType = read_vbr_uint();
1481 switch (LinkageType) {
1482 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1483 case 1: Linkage = GlobalValue::WeakLinkage; break;
1484 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1485 case 3: Linkage = GlobalValue::InternalLinkage; break;
1486 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1488 error("Invalid linkage type for Function.");
1489 Linkage = GlobalValue::InternalLinkage;
1493 F->setLinkage(Linkage);
1494 if (Handler) Handler->handleFunctionBegin(F,FuncSize);
1496 // Keep track of how many basic blocks we have read in...
1497 unsigned BlockNum = 0;
1498 bool InsertedArguments = false;
1500 BufPtr MyEnd = BlockEnd;
1501 while (At < MyEnd) {
1502 unsigned Type, Size;
1504 read_block(Type, Size);
1507 case BytecodeFormat::ConstantPool:
1508 if (!InsertedArguments) {
1509 // Insert arguments into the value table before we parse the first basic
1510 // block in the function, but after we potentially read in the
1511 // compaction table.
1513 InsertedArguments = true;
1516 ParseConstantPool(FunctionValues, FunctionTypes, true);
1519 case BytecodeFormat::CompactionTable:
1520 ParseCompactionTable();
1523 case BytecodeFormat::BasicBlock: {
1524 if (!InsertedArguments) {
1525 // Insert arguments into the value table before we parse the first basic
1526 // block in the function, but after we potentially read in the
1527 // compaction table.
1529 InsertedArguments = true;
1532 BasicBlock *BB = ParseBasicBlock(BlockNum++);
1533 F->getBasicBlockList().push_back(BB);
1537 case BytecodeFormat::InstructionList: {
1538 // Insert arguments into the value table before we parse the instruction
1539 // list for the function, but after we potentially read in the compaction
1541 if (!InsertedArguments) {
1543 InsertedArguments = true;
1547 error("Already parsed basic blocks!");
1548 BlockNum = ParseInstructionList(F);
1552 case BytecodeFormat::SymbolTable:
1553 ParseSymbolTable(F, &F->getSymbolTable());
1559 error("Wrapped around reading bytecode.");
1564 // Malformed bc file if read past end of block.
1568 // Make sure there were no references to non-existant basic blocks.
1569 if (BlockNum != ParsedBasicBlocks.size())
1570 error("Illegal basic block operand reference");
1572 ParsedBasicBlocks.clear();
1574 // Resolve forward references. Replace any uses of a forward reference value
1575 // with the real value.
1577 // replaceAllUsesWith is very inefficient for instructions which have a LARGE
1578 // number of operands. PHI nodes often have forward references, and can also
1579 // often have a very large number of operands.
1581 // FIXME: REEVALUATE. replaceAllUsesWith is _much_ faster now, and this code
1582 // should be simplified back to using it!
1584 std::map<Value*, Value*> ForwardRefMapping;
1585 for (std::map<std::pair<unsigned,unsigned>, Value*>::iterator
1586 I = ForwardReferences.begin(), E = ForwardReferences.end();
1588 ForwardRefMapping[I->second] = getValue(I->first.first, I->first.second,
1591 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
1592 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
1593 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1594 if (Argument *A = dyn_cast<Argument>(I->getOperand(i))) {
1595 std::map<Value*, Value*>::iterator It = ForwardRefMapping.find(A);
1596 if (It != ForwardRefMapping.end()) I->setOperand(i, It->second);
1599 while (!ForwardReferences.empty()) {
1600 std::map<std::pair<unsigned,unsigned>, Value*>::iterator I =
1601 ForwardReferences.begin();
1602 Value *PlaceHolder = I->second;
1603 ForwardReferences.erase(I);
1605 // Now that all the uses are gone, delete the placeholder...
1606 // If we couldn't find a def (error case), then leak a little
1607 // memory, because otherwise we can't remove all uses!
1611 // Clear out function-level types...
1612 FunctionTypes.clear();
1613 CompactionTypes.clear();
1614 CompactionValues.clear();
1615 freeTable(FunctionValues);
1617 if (Handler) Handler->handleFunctionEnd(F);
1620 /// This function parses LLVM functions lazily. It obtains the type of the
1621 /// function and records where the body of the function is in the bytecode
1622 /// buffer. The caller can then use the ParseNextFunction and
1623 /// ParseAllFunctionBodies to get handler events for the functions.
1624 void BytecodeReader::ParseFunctionLazily() {
1625 if (FunctionSignatureList.empty())
1626 error("FunctionSignatureList empty!");
1628 Function *Func = FunctionSignatureList.back();
1629 FunctionSignatureList.pop_back();
1631 // Save the information for future reading of the function
1632 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd);
1634 // Pretend we've `parsed' this function
1638 /// The ParserFunction method lazily parses one function. Use this method to
1639 /// casue the parser to parse a specific function in the module. Note that
1640 /// this will remove the function from what is to be included by
1641 /// ParseAllFunctionBodies.
1642 /// @see ParseAllFunctionBodies
1643 /// @see ParseBytecode
1644 void BytecodeReader::ParseFunction(Function* Func) {
1645 // Find {start, end} pointers and slot in the map. If not there, we're done.
1646 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func);
1648 // Make sure we found it
1649 if (Fi == LazyFunctionLoadMap.end()) {
1650 error("Unrecognized function of type " + Func->getType()->getDescription());
1654 BlockStart = At = Fi->second.Buf;
1655 BlockEnd = Fi->second.EndBuf;
1656 assert(Fi->first == Func && "Found wrong function?");
1658 LazyFunctionLoadMap.erase(Fi);
1660 this->ParseFunctionBody(Func);
1663 /// The ParseAllFunctionBodies method parses through all the previously
1664 /// unparsed functions in the bytecode file. If you want to completely parse
1665 /// a bytecode file, this method should be called after Parsebytecode because
1666 /// Parsebytecode only records the locations in the bytecode file of where
1667 /// the function definitions are located. This function uses that information
1668 /// to materialize the functions.
1669 /// @see ParseBytecode
1670 void BytecodeReader::ParseAllFunctionBodies() {
1671 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin();
1672 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end();
1675 Function* Func = Fi->first;
1676 BlockStart = At = Fi->second.Buf;
1677 BlockEnd = Fi->second.EndBuf;
1678 this->ParseFunctionBody(Func);
1683 /// Parse the global type list
1684 void BytecodeReader::ParseGlobalTypes() {
1685 // Read the number of types
1686 unsigned NumEntries = read_vbr_uint();
1688 // Ignore the type plane identifier for types if the bc file is pre 1.3
1689 if (hasTypeDerivedFromValue)
1692 ParseTypes(ModuleTypes, NumEntries);
1695 /// Parse the Global info (types, global vars, constants)
1696 void BytecodeReader::ParseModuleGlobalInfo() {
1698 if (Handler) Handler->handleModuleGlobalsBegin();
1700 // Read global variables...
1701 unsigned VarType = read_vbr_uint();
1702 while (VarType != Type::VoidTyID) { // List is terminated by Void
1703 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 =
1704 // Linkage, bit4+ = slot#
1705 unsigned SlotNo = VarType >> 5;
1706 if (sanitizeTypeId(SlotNo))
1707 error("Invalid type (type type) for global var!");
1708 unsigned LinkageID = (VarType >> 2) & 7;
1709 bool isConstant = VarType & 1;
1710 bool hasInitializer = VarType & 2;
1711 GlobalValue::LinkageTypes Linkage;
1713 switch (LinkageID) {
1714 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1715 case 1: Linkage = GlobalValue::WeakLinkage; break;
1716 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1717 case 3: Linkage = GlobalValue::InternalLinkage; break;
1718 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1720 error("Unknown linkage type: " + utostr(LinkageID));
1721 Linkage = GlobalValue::InternalLinkage;
1725 const Type *Ty = getType(SlotNo);
1727 error("Global has no type! SlotNo=" + utostr(SlotNo));
1730 if (!isa<PointerType>(Ty)) {
1731 error("Global not a pointer type! Ty= " + Ty->getDescription());
1734 const Type *ElTy = cast<PointerType>(Ty)->getElementType();
1736 // Create the global variable...
1737 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage,
1739 insertValue(GV, SlotNo, ModuleValues);
1741 unsigned initSlot = 0;
1742 if (hasInitializer) {
1743 initSlot = read_vbr_uint();
1744 GlobalInits.push_back(std::make_pair(GV, initSlot));
1747 // Notify handler about the global value.
1748 if (Handler) Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo, initSlot);
1751 VarType = read_vbr_uint();
1754 // Read the function objects for all of the functions that are coming
1755 unsigned FnSignature = 0;
1756 if (read_typeid(FnSignature))
1757 error("Invalid function type (type type) found");
1759 while (FnSignature != Type::VoidTyID) { // List is terminated by Void
1760 const Type *Ty = getType(FnSignature);
1761 if (!isa<PointerType>(Ty) ||
1762 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) {
1763 error("Function not a pointer to function type! Ty = " +
1764 Ty->getDescription());
1765 // FIXME: what should Ty be if handler continues?
1768 // We create functions by passing the underlying FunctionType to create...
1769 const FunctionType* FTy =
1770 cast<FunctionType>(cast<PointerType>(Ty)->getElementType());
1772 // Insert the place hodler
1773 Function* Func = new Function(FTy, GlobalValue::InternalLinkage,
1775 insertValue(Func, FnSignature, ModuleValues);
1777 // Save this for later so we know type of lazily instantiated functions
1778 FunctionSignatureList.push_back(Func);
1780 if (Handler) Handler->handleFunctionDeclaration(Func);
1782 // Get Next function signature
1783 if (read_typeid(FnSignature))
1784 error("Invalid function type (type type) found");
1787 if (hasInconsistentModuleGlobalInfo)
1790 // Now that the function signature list is set up, reverse it so that we can
1791 // remove elements efficiently from the back of the vector.
1792 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end());
1794 // This is for future proofing... in the future extra fields may be added that
1795 // we don't understand, so we transparently ignore them.
1799 if (Handler) Handler->handleModuleGlobalsEnd();
1802 /// Parse the version information and decode it by setting flags on the
1803 /// Reader that enable backward compatibility of the reader.
1804 void BytecodeReader::ParseVersionInfo() {
1805 unsigned Version = read_vbr_uint();
1807 // Unpack version number: low four bits are for flags, top bits = version
1808 Module::Endianness Endianness;
1809 Module::PointerSize PointerSize;
1810 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian;
1811 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32;
1813 bool hasNoEndianness = Version & 4;
1814 bool hasNoPointerSize = Version & 8;
1816 RevisionNum = Version >> 4;
1818 // Default values for the current bytecode version
1819 hasInconsistentModuleGlobalInfo = false;
1820 hasExplicitPrimitiveZeros = false;
1821 hasRestrictedGEPTypes = false;
1822 hasTypeDerivedFromValue = false;
1824 switch (RevisionNum) {
1825 case 0: // LLVM 1.0, 1.1 release version
1826 // Base LLVM 1.0 bytecode format.
1827 hasInconsistentModuleGlobalInfo = true;
1828 hasExplicitPrimitiveZeros = true;
1831 case 1: // LLVM 1.2 release version
1832 // LLVM 1.2 added explicit support for emitting strings efficiently.
1834 // Also, it fixed the problem where the size of the ModuleGlobalInfo block
1835 // included the size for the alignment at the end, where the rest of the
1838 // LLVM 1.2 and before required that GEP indices be ubyte constants for
1839 // structures and longs for sequential types.
1840 hasRestrictedGEPTypes = true;
1842 // LLVM 1.2 and before had the Type class derive from Value class. This
1843 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are
1844 // written differently because Types can no longer be part of the
1845 // type planes for Values.
1846 hasTypeDerivedFromValue = true;
1849 case 2: // LLVM 1.3 release version
1853 error("Unknown bytecode version number: " + itostr(RevisionNum));
1856 if (hasNoEndianness) Endianness = Module::AnyEndianness;
1857 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize;
1859 TheModule->setEndianness(Endianness);
1860 TheModule->setPointerSize(PointerSize);
1862 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize);
1865 /// Parse a whole module.
1866 void BytecodeReader::ParseModule() {
1867 unsigned Type, Size;
1869 FunctionSignatureList.clear(); // Just in case...
1871 // Read into instance variables...
1873 align32(); /// FIXME: Is this redundant? VI is first and 4 bytes!
1875 bool SeenModuleGlobalInfo = false;
1876 bool SeenGlobalTypePlane = false;
1877 BufPtr MyEnd = BlockEnd;
1878 while (At < MyEnd) {
1880 read_block(Type, Size);
1884 case BytecodeFormat::GlobalTypePlane:
1885 if (SeenGlobalTypePlane)
1886 error("Two GlobalTypePlane Blocks Encountered!");
1889 SeenGlobalTypePlane = true;
1892 case BytecodeFormat::ModuleGlobalInfo:
1893 if (SeenModuleGlobalInfo)
1894 error("Two ModuleGlobalInfo Blocks Encountered!");
1895 ParseModuleGlobalInfo();
1896 SeenModuleGlobalInfo = true;
1899 case BytecodeFormat::ConstantPool:
1900 ParseConstantPool(ModuleValues, ModuleTypes,false);
1903 case BytecodeFormat::Function:
1904 ParseFunctionLazily();
1907 case BytecodeFormat::SymbolTable:
1908 ParseSymbolTable(0, &TheModule->getSymbolTable());
1914 error("Unexpected Block of Type #" + utostr(Type) + " encountered!");
1922 // After the module constant pool has been read, we can safely initialize
1923 // global variables...
1924 while (!GlobalInits.empty()) {
1925 GlobalVariable *GV = GlobalInits.back().first;
1926 unsigned Slot = GlobalInits.back().second;
1927 GlobalInits.pop_back();
1929 // Look up the initializer value...
1930 // FIXME: Preserve this type ID!
1932 const llvm::PointerType* GVType = GV->getType();
1933 unsigned TypeSlot = getTypeSlot(GVType->getElementType());
1934 if (Constant *CV = getConstantValue(TypeSlot, Slot)) {
1935 if (GV->hasInitializer())
1936 error("Global *already* has an initializer?!");
1937 if (Handler) Handler->handleGlobalInitializer(GV,CV);
1938 GV->setInitializer(CV);
1940 error("Cannot find initializer value.");
1943 /// Make sure we pulled them all out. If we didn't then there's a declaration
1944 /// but a missing body. That's not allowed.
1945 if (!FunctionSignatureList.empty())
1946 error("Function declared, but bytecode stream ended before definition");
1949 /// This function completely parses a bytecode buffer given by the \p Buf
1950 /// and \p Length parameters.
1951 void BytecodeReader::ParseBytecode(BufPtr Buf, unsigned Length,
1952 const std::string &ModuleID,
1953 bool processFunctions) {
1956 At = MemStart = BlockStart = Buf;
1957 MemEnd = BlockEnd = Buf + Length;
1959 // Create the module
1960 TheModule = new Module(ModuleID);
1962 if (Handler) Handler->handleStart(TheModule, Length);
1964 // Read and check signature...
1965 unsigned Sig = read_uint();
1966 if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) {
1967 error("Invalid bytecode signature: " + utostr(Sig));
1971 // Tell the handler we're starting a module
1972 if (Handler) Handler->handleModuleBegin(ModuleID);
1974 // Get the module block and size and verify
1975 unsigned Type, Size;
1976 read_block(Type, Size);
1977 if (Type != BytecodeFormat::Module) {
1978 error("Expected Module Block! Type:" + utostr(Type) + ", Size:"
1981 if (At + Size != MemEnd) {
1982 error("Invalid Top Level Block Length! Type:" + utostr(Type)
1983 + ", Size:" + utostr(Size));
1986 // Parse the module contents
1987 this->ParseModule();
1989 // Check for missing functions
1991 error("Function expected, but bytecode stream ended!");
1993 // Process all the function bodies now, if requested
1994 if (processFunctions)
1995 ParseAllFunctionBodies();
1997 // Tell the handler we're done with the module
1999 Handler->handleModuleEnd(ModuleID);
2001 // Tell the handler we're finished the parse
2002 if (Handler) Handler->handleFinish();
2004 } catch (std::string& errstr) {
2005 if (Handler) Handler->handleError(errstr);
2011 std::string msg("Unknown Exception Occurred");
2012 if (Handler) Handler->handleError(msg);
2020 //===----------------------------------------------------------------------===//
2021 //=== Default Implementations of Handler Methods
2022 //===----------------------------------------------------------------------===//
2024 BytecodeHandler::~BytecodeHandler() {}