1 //===- Reader.cpp - Code to read bytecode files ---------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This library implements the functionality defined in llvm/Bytecode/Reader.h
12 // Note that this library should be as fast as possible, reentrant, and
15 // TODO: Allow passing in an option to ignore the symbol table
17 //===----------------------------------------------------------------------===//
20 #include "llvm/Assembly/AutoUpgrade.h"
21 #include "llvm/Bytecode/BytecodeHandler.h"
22 #include "llvm/BasicBlock.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/InlineAsm.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/SymbolTable.h"
28 #include "llvm/Bytecode/Format.h"
29 #include "llvm/Config/alloca.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Support/Compressor.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/ADT/StringExtras.h"
39 /// @brief A class for maintaining the slot number definition
40 /// as a placeholder for the actual definition for forward constants defs.
41 class ConstantPlaceHolder : public ConstantExpr {
42 ConstantPlaceHolder(); // DO NOT IMPLEMENT
43 void operator=(const ConstantPlaceHolder &); // DO NOT IMPLEMENT
46 ConstantPlaceHolder(const Type *Ty)
47 : ConstantExpr(Ty, Instruction::UserOp1, &Op, 1),
48 Op(UndefValue::get(Type::IntTy), this) {
53 // Provide some details on error
54 inline void BytecodeReader::error(const std::string& err) {
55 ErrorMsg = err + " (Vers=" + itostr(RevisionNum) + ", Pos="
56 + 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 *)((intptr_t)(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.
165 FloatVal = BitsToFloat(At[0] | (At[1] << 8) | (At[2] << 16) | (At[3] << 24));
166 At+=sizeof(uint32_t);
169 /// Read a double value in little-endian order
170 inline void BytecodeReader::read_double(double& DoubleVal) {
171 /// FIXME: This isn't optimal, it has size problems on some platforms
172 /// where FP is not IEEE.
173 DoubleVal = BitsToDouble((uint64_t(At[0]) << 0) | (uint64_t(At[1]) << 8) |
174 (uint64_t(At[2]) << 16) | (uint64_t(At[3]) << 24) |
175 (uint64_t(At[4]) << 32) | (uint64_t(At[5]) << 40) |
176 (uint64_t(At[6]) << 48) | (uint64_t(At[7]) << 56));
177 At+=sizeof(uint64_t);
180 /// Read a block header and obtain its type and size
181 inline void BytecodeReader::read_block(unsigned &Type, unsigned &Size) {
182 if ( hasLongBlockHeaders ) {
186 case BytecodeFormat::Reserved_DoNotUse :
187 error("Reserved_DoNotUse used as Module Type?");
188 Type = BytecodeFormat::ModuleBlockID; break;
189 case BytecodeFormat::Module:
190 Type = BytecodeFormat::ModuleBlockID; break;
191 case BytecodeFormat::Function:
192 Type = BytecodeFormat::FunctionBlockID; break;
193 case BytecodeFormat::ConstantPool:
194 Type = BytecodeFormat::ConstantPoolBlockID; break;
195 case BytecodeFormat::SymbolTable:
196 Type = BytecodeFormat::SymbolTableBlockID; break;
197 case BytecodeFormat::ModuleGlobalInfo:
198 Type = BytecodeFormat::ModuleGlobalInfoBlockID; break;
199 case BytecodeFormat::GlobalTypePlane:
200 Type = BytecodeFormat::GlobalTypePlaneBlockID; break;
201 case BytecodeFormat::InstructionList:
202 Type = BytecodeFormat::InstructionListBlockID; break;
203 case BytecodeFormat::CompactionTable:
204 Type = BytecodeFormat::CompactionTableBlockID; break;
205 case BytecodeFormat::BasicBlock:
206 /// This block type isn't used after version 1.1. However, we have to
207 /// still allow the value in case this is an old bc format file.
208 /// We just let its value creep thru.
211 error("Invalid block id found: " + utostr(Type));
216 Type = Size & 0x1F; // mask low order five bits
217 Size >>= 5; // get rid of five low order bits, leaving high 27
220 if (At + Size > BlockEnd)
221 error("Attempt to size a block past end of memory");
222 BlockEnd = At + Size;
223 if (Handler) Handler->handleBlock(Type, BlockStart, Size);
227 /// In LLVM 1.2 and before, Types were derived from Value and so they were
228 /// written as part of the type planes along with any other Value. In LLVM
229 /// 1.3 this changed so that Type does not derive from Value. Consequently,
230 /// the BytecodeReader's containers for Values can't contain Types because
231 /// there's no inheritance relationship. This means that the "Type Type"
232 /// plane is defunct along with the Type::TypeTyID TypeID. In LLVM 1.3
233 /// whenever a bytecode construct must have both types and values together,
234 /// the types are always read/written first and then the Values. Furthermore
235 /// since Type::TypeTyID no longer exists, its value (12) now corresponds to
236 /// Type::LabelTyID. In order to overcome this we must "sanitize" all the
237 /// type TypeIDs we encounter. For LLVM 1.3 bytecode files, there's no change.
238 /// For LLVM 1.2 and before, this function will decrement the type id by
239 /// one to account for the missing Type::TypeTyID enumerator if the value is
240 /// larger than 12 (Type::LabelTyID). If the value is exactly 12, then this
241 /// function returns true, otherwise false. This helps detect situations
242 /// where the pre 1.3 bytecode is indicating that what follows is a type.
243 /// @returns true iff type id corresponds to pre 1.3 "type type"
244 inline bool BytecodeReader::sanitizeTypeId(unsigned &TypeId) {
245 if (hasTypeDerivedFromValue) { /// do nothing if 1.3 or later
246 if (TypeId == Type::LabelTyID) {
247 TypeId = Type::VoidTyID; // sanitize it
248 return true; // indicate we got TypeTyID in pre 1.3 bytecode
249 } else if (TypeId > Type::LabelTyID)
250 --TypeId; // shift all planes down because type type plane is missing
255 /// Reads a vbr uint to read in a type id and does the necessary
256 /// conversion on it by calling sanitizeTypeId.
257 /// @returns true iff \p TypeId read corresponds to a pre 1.3 "type type"
258 /// @see sanitizeTypeId
259 inline bool BytecodeReader::read_typeid(unsigned &TypeId) {
260 TypeId = read_vbr_uint();
261 if ( !has32BitTypes )
262 if ( TypeId == 0x00FFFFFF )
263 TypeId = read_vbr_uint();
264 return sanitizeTypeId(TypeId);
267 //===----------------------------------------------------------------------===//
269 //===----------------------------------------------------------------------===//
271 /// Determine if a type id has an implicit null value
272 inline bool BytecodeReader::hasImplicitNull(unsigned TyID) {
273 if (!hasExplicitPrimitiveZeros)
274 return TyID != Type::LabelTyID && TyID != Type::VoidTyID;
275 return TyID >= Type::FirstDerivedTyID;
278 /// Obtain a type given a typeid and account for things like compaction tables,
279 /// function level vs module level, and the offsetting for the primitive types.
280 const Type *BytecodeReader::getType(unsigned ID) {
281 if (ID < Type::FirstDerivedTyID)
282 if (const Type *T = Type::getPrimitiveType((Type::TypeID)ID))
283 return T; // Asked for a primitive type...
285 // Otherwise, derived types need offset...
286 ID -= Type::FirstDerivedTyID;
288 if (!CompactionTypes.empty()) {
289 if (ID >= CompactionTypes.size())
290 error("Type ID out of range for compaction table!");
291 return CompactionTypes[ID].first;
294 // Is it a module-level type?
295 if (ID < ModuleTypes.size())
296 return ModuleTypes[ID].get();
298 // Nope, is it a function-level type?
299 ID -= ModuleTypes.size();
300 if (ID < FunctionTypes.size())
301 return FunctionTypes[ID].get();
303 error("Illegal type reference!");
307 /// Get a sanitized type id. This just makes sure that the \p ID
308 /// is both sanitized and not the "type type" of pre-1.3 bytecode.
309 /// @see sanitizeTypeId
310 inline const Type* BytecodeReader::getSanitizedType(unsigned& ID) {
311 if (sanitizeTypeId(ID))
312 error("Invalid type id encountered");
316 /// This method just saves some coding. It uses read_typeid to read
317 /// in a sanitized type id, errors that its not the type type, and
318 /// then calls getType to return the type value.
319 inline const Type* BytecodeReader::readSanitizedType() {
322 error("Invalid type id encountered");
326 /// Get the slot number associated with a type accounting for primitive
327 /// types, compaction tables, and function level vs module level.
328 unsigned BytecodeReader::getTypeSlot(const Type *Ty) {
329 if (Ty->isPrimitiveType())
330 return Ty->getTypeID();
332 // Scan the compaction table for the type if needed.
333 if (!CompactionTypes.empty()) {
334 for (unsigned i = 0, e = CompactionTypes.size(); i != e; ++i)
335 if (CompactionTypes[i].first == Ty)
336 return Type::FirstDerivedTyID + i;
338 error("Couldn't find type specified in compaction table!");
341 // Check the function level types first...
342 TypeListTy::iterator I = std::find(FunctionTypes.begin(),
343 FunctionTypes.end(), Ty);
345 if (I != FunctionTypes.end())
346 return Type::FirstDerivedTyID + ModuleTypes.size() +
347 (&*I - &FunctionTypes[0]);
349 // If we don't have our cache yet, build it now.
350 if (ModuleTypeIDCache.empty()) {
352 ModuleTypeIDCache.reserve(ModuleTypes.size());
353 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end();
355 ModuleTypeIDCache.push_back(std::make_pair(*I, N));
357 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end());
360 // Binary search the cache for the entry.
361 std::vector<std::pair<const Type*, unsigned> >::iterator IT =
362 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(),
363 std::make_pair(Ty, 0U));
364 if (IT == ModuleTypeIDCache.end() || IT->first != Ty)
365 error("Didn't find type in ModuleTypes.");
367 return Type::FirstDerivedTyID + IT->second;
370 /// This is just like getType, but when a compaction table is in use, it is
371 /// ignored. It also ignores function level types.
373 const Type *BytecodeReader::getGlobalTableType(unsigned Slot) {
374 if (Slot < Type::FirstDerivedTyID) {
375 const Type *Ty = Type::getPrimitiveType((Type::TypeID)Slot);
377 error("Not a primitive type ID?");
380 Slot -= Type::FirstDerivedTyID;
381 if (Slot >= ModuleTypes.size())
382 error("Illegal compaction table type reference!");
383 return ModuleTypes[Slot];
386 /// This is just like getTypeSlot, but when a compaction table is in use, it
387 /// is ignored. It also ignores function level types.
388 unsigned BytecodeReader::getGlobalTableTypeSlot(const Type *Ty) {
389 if (Ty->isPrimitiveType())
390 return Ty->getTypeID();
392 // If we don't have our cache yet, build it now.
393 if (ModuleTypeIDCache.empty()) {
395 ModuleTypeIDCache.reserve(ModuleTypes.size());
396 for (TypeListTy::iterator I = ModuleTypes.begin(), E = ModuleTypes.end();
398 ModuleTypeIDCache.push_back(std::make_pair(*I, N));
400 std::sort(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end());
403 // Binary search the cache for the entry.
404 std::vector<std::pair<const Type*, unsigned> >::iterator IT =
405 std::lower_bound(ModuleTypeIDCache.begin(), ModuleTypeIDCache.end(),
406 std::make_pair(Ty, 0U));
407 if (IT == ModuleTypeIDCache.end() || IT->first != Ty)
408 error("Didn't find type in ModuleTypes.");
410 return Type::FirstDerivedTyID + IT->second;
413 /// Retrieve a value of a given type and slot number, possibly creating
414 /// it if it doesn't already exist.
415 Value * BytecodeReader::getValue(unsigned type, unsigned oNum, bool Create) {
416 assert(type != Type::LabelTyID && "getValue() cannot get blocks!");
419 // If there is a compaction table active, it defines the low-level numbers.
420 // If not, the module values define the low-level numbers.
421 if (CompactionValues.size() > type && !CompactionValues[type].empty()) {
422 if (Num < CompactionValues[type].size())
423 return CompactionValues[type][Num];
424 Num -= CompactionValues[type].size();
426 // By default, the global type id is the type id passed in
427 unsigned GlobalTyID = type;
429 // If the type plane was compactified, figure out the global type ID by
430 // adding the derived type ids and the distance.
431 if (!CompactionTypes.empty() && type >= Type::FirstDerivedTyID)
432 GlobalTyID = CompactionTypes[type-Type::FirstDerivedTyID].second;
434 if (hasImplicitNull(GlobalTyID)) {
435 const Type *Ty = getType(type);
436 if (!isa<OpaqueType>(Ty)) {
438 return Constant::getNullValue(Ty);
443 if (GlobalTyID < ModuleValues.size() && ModuleValues[GlobalTyID]) {
444 if (Num < ModuleValues[GlobalTyID]->size())
445 return ModuleValues[GlobalTyID]->getOperand(Num);
446 Num -= ModuleValues[GlobalTyID]->size();
450 if (FunctionValues.size() > type &&
451 FunctionValues[type] &&
452 Num < FunctionValues[type]->size())
453 return FunctionValues[type]->getOperand(Num);
455 if (!Create) return 0; // Do not create a placeholder?
457 // Did we already create a place holder?
458 std::pair<unsigned,unsigned> KeyValue(type, oNum);
459 ForwardReferenceMap::iterator I = ForwardReferences.lower_bound(KeyValue);
460 if (I != ForwardReferences.end() && I->first == KeyValue)
461 return I->second; // We have already created this placeholder
463 // If the type exists (it should)
464 if (const Type* Ty = getType(type)) {
465 // Create the place holder
466 Value *Val = new Argument(Ty);
467 ForwardReferences.insert(I, std::make_pair(KeyValue, Val));
470 error("Can't create placeholder for value of type slot #" + utostr(type));
471 return 0; // just silence warning, error calls longjmp
474 /// This is just like getValue, but when a compaction table is in use, it
475 /// is ignored. Also, no forward references or other fancy features are
477 Value* BytecodeReader::getGlobalTableValue(unsigned TyID, unsigned SlotNo) {
479 return Constant::getNullValue(getType(TyID));
481 if (!CompactionTypes.empty() && TyID >= Type::FirstDerivedTyID) {
482 TyID -= Type::FirstDerivedTyID;
483 if (TyID >= CompactionTypes.size())
484 error("Type ID out of range for compaction table!");
485 TyID = CompactionTypes[TyID].second;
490 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0 ||
491 SlotNo >= ModuleValues[TyID]->size()) {
492 if (TyID >= ModuleValues.size() || ModuleValues[TyID] == 0)
493 error("Corrupt compaction table entry!"
494 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
495 + utostr(ModuleValues.size()));
497 error("Corrupt compaction table entry!"
498 + utostr(TyID) + ", " + utostr(SlotNo) + ": "
499 + utostr(ModuleValues.size()) + ", "
500 + utohexstr(reinterpret_cast<uint64_t>(((void*)ModuleValues[TyID])))
502 + utostr(ModuleValues[TyID]->size()));
504 return ModuleValues[TyID]->getOperand(SlotNo);
507 /// Just like getValue, except that it returns a null pointer
508 /// only on error. It always returns a constant (meaning that if the value is
509 /// defined, but is not a constant, that is an error). If the specified
510 /// constant hasn't been parsed yet, a placeholder is defined and used.
511 /// Later, after the real value is parsed, the placeholder is eliminated.
512 Constant* BytecodeReader::getConstantValue(unsigned TypeSlot, unsigned Slot) {
513 if (Value *V = getValue(TypeSlot, Slot, false))
514 if (Constant *C = dyn_cast<Constant>(V))
515 return C; // If we already have the value parsed, just return it
517 error("Value for slot " + utostr(Slot) +
518 " is expected to be a constant!");
520 std::pair<unsigned, unsigned> Key(TypeSlot, Slot);
521 ConstantRefsType::iterator I = ConstantFwdRefs.lower_bound(Key);
523 if (I != ConstantFwdRefs.end() && I->first == Key) {
526 // Create a placeholder for the constant reference and
527 // keep track of the fact that we have a forward ref to recycle it
528 Constant *C = new ConstantPlaceHolder(getType(TypeSlot));
530 // Keep track of the fact that we have a forward ref to recycle it
531 ConstantFwdRefs.insert(I, std::make_pair(Key, C));
536 //===----------------------------------------------------------------------===//
537 // IR Construction Methods
538 //===----------------------------------------------------------------------===//
540 /// As values are created, they are inserted into the appropriate place
541 /// with this method. The ValueTable argument must be one of ModuleValues
542 /// or FunctionValues data members of this class.
543 unsigned BytecodeReader::insertValue(Value *Val, unsigned type,
544 ValueTable &ValueTab) {
545 if (ValueTab.size() <= type)
546 ValueTab.resize(type+1);
548 if (!ValueTab[type]) ValueTab[type] = new ValueList();
550 ValueTab[type]->push_back(Val);
552 bool HasOffset = hasImplicitNull(type) && !isa<OpaqueType>(Val->getType());
553 return ValueTab[type]->size()-1 + HasOffset;
556 /// Insert the arguments of a function as new values in the reader.
557 void BytecodeReader::insertArguments(Function* F) {
558 const FunctionType *FT = F->getFunctionType();
559 Function::arg_iterator AI = F->arg_begin();
560 for (FunctionType::param_iterator It = FT->param_begin();
561 It != FT->param_end(); ++It, ++AI)
562 insertValue(AI, getTypeSlot(AI->getType()), FunctionValues);
565 //===----------------------------------------------------------------------===//
566 // Bytecode Parsing Methods
567 //===----------------------------------------------------------------------===//
569 /// This method parses a single instruction. The instruction is
570 /// inserted at the end of the \p BB provided. The arguments of
571 /// the instruction are provided in the \p Oprnds vector.
572 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds,
576 // Clear instruction data
580 unsigned Op = read_uint();
582 // bits Instruction format: Common to all formats
583 // --------------------------
584 // 01-00: Opcode type, fixed to 1.
586 Opcode = (Op >> 2) & 63;
587 Oprnds.resize((Op >> 0) & 03);
589 // Extract the operands
590 switch (Oprnds.size()) {
592 // bits Instruction format:
593 // --------------------------
594 // 19-08: Resulting type plane
595 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
597 iType = (Op >> 8) & 4095;
598 Oprnds[0] = (Op >> 20) & 4095;
599 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands...
603 // bits Instruction format:
604 // --------------------------
605 // 15-08: Resulting type plane
609 iType = (Op >> 8) & 255;
610 Oprnds[0] = (Op >> 16) & 255;
611 Oprnds[1] = (Op >> 24) & 255;
614 // bits Instruction format:
615 // --------------------------
616 // 13-08: Resulting type plane
621 iType = (Op >> 8) & 63;
622 Oprnds[0] = (Op >> 14) & 63;
623 Oprnds[1] = (Op >> 20) & 63;
624 Oprnds[2] = (Op >> 26) & 63;
627 At -= 4; // Hrm, try this again...
628 Opcode = read_vbr_uint();
630 iType = read_vbr_uint();
632 unsigned NumOprnds = read_vbr_uint();
633 Oprnds.resize(NumOprnds);
636 error("Zero-argument instruction found; this is invalid.");
638 for (unsigned i = 0; i != NumOprnds; ++i)
639 Oprnds[i] = read_vbr_uint();
644 const Type *InstTy = getSanitizedType(iType);
646 // We have enough info to inform the handler now.
647 if (Handler) Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt);
649 // Declare the resulting instruction we'll build.
650 Instruction *Result = 0;
652 // If this is a bytecode format that did not include the unreachable
653 // instruction, bump up all opcodes numbers to make space.
654 if (hasNoUnreachableInst) {
655 if (Opcode >= Instruction::Unreachable &&
661 // Handle binary operators
662 if (Opcode >= Instruction::BinaryOpsBegin &&
663 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2)
664 Result = BinaryOperator::create((Instruction::BinaryOps)Opcode,
665 getValue(iType, Oprnds[0]),
666 getValue(iType, Oprnds[1]));
672 error("Illegal instruction read!");
674 case Instruction::VAArg:
675 Result = new VAArgInst(getValue(iType, Oprnds[0]),
676 getSanitizedType(Oprnds[1]));
678 case 32: { //VANext_old
679 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
680 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy,
684 //foo = alloca 1 of t
689 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vanext.fix");
690 BB->getInstList().push_back(foo);
691 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
692 BB->getInstList().push_back(bar);
693 BB->getInstList().push_back(new StoreInst(bar, foo));
694 Instruction* tmp = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
695 BB->getInstList().push_back(tmp);
696 Result = new LoadInst(foo);
699 case 33: { //VAArg_old
700 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
701 Function* NF = TheModule->getOrInsertFunction("llvm.va_copy", ArgTy, ArgTy,
705 //foo = alloca 1 of t
709 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vaarg.fix");
710 BB->getInstList().push_back(foo);
711 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
712 BB->getInstList().push_back(bar);
713 BB->getInstList().push_back(new StoreInst(bar, foo));
714 Result = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
717 case Instruction::ExtractElement: {
718 if (Oprnds.size() != 2)
719 error("Invalid extractelement instruction!");
720 Value *V1 = getValue(iType, Oprnds[0]);
721 Value *V2 = getValue(Type::UIntTyID, Oprnds[1]);
723 if (!ExtractElementInst::isValidOperands(V1, V2))
724 error("Invalid extractelement instruction!");
726 Result = new ExtractElementInst(V1, V2);
729 case Instruction::InsertElement: {
730 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy);
731 if (!PackedTy || Oprnds.size() != 3)
732 error("Invalid insertelement instruction!");
734 Value *V1 = getValue(iType, Oprnds[0]);
735 Value *V2 = getValue(getTypeSlot(PackedTy->getElementType()), Oprnds[1]);
736 Value *V3 = getValue(Type::UIntTyID, Oprnds[2]);
738 if (!InsertElementInst::isValidOperands(V1, V2, V3))
739 error("Invalid insertelement instruction!");
740 Result = new InsertElementInst(V1, V2, V3);
743 case Instruction::ShuffleVector: {
744 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy);
745 if (!PackedTy || Oprnds.size() != 3)
746 error("Invalid shufflevector instruction!");
747 Value *V1 = getValue(iType, Oprnds[0]);
748 Value *V2 = getValue(iType, Oprnds[1]);
749 const PackedType *EltTy =
750 PackedType::get(Type::UIntTy, PackedTy->getNumElements());
751 Value *V3 = getValue(getTypeSlot(EltTy), Oprnds[2]);
752 if (!ShuffleVectorInst::isValidOperands(V1, V2, V3))
753 error("Invalid shufflevector instruction!");
754 Result = new ShuffleVectorInst(V1, V2, V3);
757 case Instruction::Cast:
758 Result = new CastInst(getValue(iType, Oprnds[0]),
759 getSanitizedType(Oprnds[1]));
761 case Instruction::Select:
762 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]),
763 getValue(iType, Oprnds[1]),
764 getValue(iType, Oprnds[2]));
766 case Instruction::PHI: {
767 if (Oprnds.size() == 0 || (Oprnds.size() & 1))
768 error("Invalid phi node encountered!");
770 PHINode *PN = new PHINode(InstTy);
771 PN->reserveOperandSpace(Oprnds.size());
772 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2)
773 PN->addIncoming(getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1]));
778 case Instruction::Shl:
779 case Instruction::Shr:
780 Result = new ShiftInst((Instruction::OtherOps)Opcode,
781 getValue(iType, Oprnds[0]),
782 getValue(Type::UByteTyID, Oprnds[1]));
784 case Instruction::Ret:
785 if (Oprnds.size() == 0)
786 Result = new ReturnInst();
787 else if (Oprnds.size() == 1)
788 Result = new ReturnInst(getValue(iType, Oprnds[0]));
790 error("Unrecognized instruction!");
793 case Instruction::Br:
794 if (Oprnds.size() == 1)
795 Result = new BranchInst(getBasicBlock(Oprnds[0]));
796 else if (Oprnds.size() == 3)
797 Result = new BranchInst(getBasicBlock(Oprnds[0]),
798 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2]));
800 error("Invalid number of operands for a 'br' instruction!");
802 case Instruction::Switch: {
803 if (Oprnds.size() & 1)
804 error("Switch statement with odd number of arguments!");
806 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]),
807 getBasicBlock(Oprnds[1]),
809 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2)
810 I->addCase(cast<ConstantInt>(getValue(iType, Oprnds[i])),
811 getBasicBlock(Oprnds[i+1]));
816 case 58: // Call with extra operand for calling conv
817 case 59: // tail call, Fast CC
818 case 60: // normal call, Fast CC
819 case 61: // tail call, C Calling Conv
820 case Instruction::Call: { // Normal Call, C Calling Convention
821 if (Oprnds.size() == 0)
822 error("Invalid call instruction encountered!");
824 Value *F = getValue(iType, Oprnds[0]);
826 unsigned CallingConv = CallingConv::C;
827 bool isTailCall = false;
829 if (Opcode == 61 || Opcode == 59)
833 isTailCall = Oprnds.back() & 1;
834 CallingConv = Oprnds.back() >> 1;
836 } else if (Opcode == 59 || Opcode == 60) {
837 CallingConv = CallingConv::Fast;
840 // Check to make sure we have a pointer to function type
841 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
842 if (PTy == 0) error("Call to non function pointer value!");
843 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
844 if (FTy == 0) error("Call to non function pointer value!");
846 std::vector<Value *> Params;
847 if (!FTy->isVarArg()) {
848 FunctionType::param_iterator It = FTy->param_begin();
850 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
851 if (It == FTy->param_end())
852 error("Invalid call instruction!");
853 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
855 if (It != FTy->param_end())
856 error("Invalid call instruction!");
858 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
860 unsigned FirstVariableOperand;
861 if (Oprnds.size() < FTy->getNumParams())
862 error("Call instruction missing operands!");
864 // Read all of the fixed arguments
865 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
866 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i]));
868 FirstVariableOperand = FTy->getNumParams();
870 if ((Oprnds.size()-FirstVariableOperand) & 1)
871 error("Invalid call instruction!"); // Must be pairs of type/value
873 for (unsigned i = FirstVariableOperand, e = Oprnds.size();
875 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
878 Result = new CallInst(F, Params);
879 if (isTailCall) cast<CallInst>(Result)->setTailCall();
880 if (CallingConv) cast<CallInst>(Result)->setCallingConv(CallingConv);
883 case 56: // Invoke with encoded CC
884 case 57: // Invoke Fast CC
885 case Instruction::Invoke: { // Invoke C CC
886 if (Oprnds.size() < 3)
887 error("Invalid invoke instruction!");
888 Value *F = getValue(iType, Oprnds[0]);
890 // Check to make sure we have a pointer to function type
891 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
893 error("Invoke to non function pointer value!");
894 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
896 error("Invoke to non function pointer value!");
898 std::vector<Value *> Params;
899 BasicBlock *Normal, *Except;
900 unsigned CallingConv = CallingConv::C;
903 CallingConv = CallingConv::Fast;
904 else if (Opcode == 56) {
905 CallingConv = Oprnds.back();
909 if (!FTy->isVarArg()) {
910 Normal = getBasicBlock(Oprnds[1]);
911 Except = getBasicBlock(Oprnds[2]);
913 FunctionType::param_iterator It = FTy->param_begin();
914 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) {
915 if (It == FTy->param_end())
916 error("Invalid invoke instruction!");
917 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
919 if (It != FTy->param_end())
920 error("Invalid invoke instruction!");
922 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
924 Normal = getBasicBlock(Oprnds[0]);
925 Except = getBasicBlock(Oprnds[1]);
927 unsigned FirstVariableArgument = FTy->getNumParams()+2;
928 for (unsigned i = 2; i != FirstVariableArgument; ++i)
929 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)),
932 if (Oprnds.size()-FirstVariableArgument & 1) // Must be type/value pairs
933 error("Invalid invoke instruction!");
935 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2)
936 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
939 Result = new InvokeInst(F, Normal, Except, Params);
940 if (CallingConv) cast<InvokeInst>(Result)->setCallingConv(CallingConv);
943 case Instruction::Malloc: {
945 if (Oprnds.size() == 2)
946 Align = (1 << Oprnds[1]) >> 1;
947 else if (Oprnds.size() > 2)
948 error("Invalid malloc instruction!");
949 if (!isa<PointerType>(InstTy))
950 error("Invalid malloc instruction!");
952 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(),
953 getValue(Type::UIntTyID, Oprnds[0]), Align);
957 case Instruction::Alloca: {
959 if (Oprnds.size() == 2)
960 Align = (1 << Oprnds[1]) >> 1;
961 else if (Oprnds.size() > 2)
962 error("Invalid alloca instruction!");
963 if (!isa<PointerType>(InstTy))
964 error("Invalid alloca instruction!");
966 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(),
967 getValue(Type::UIntTyID, Oprnds[0]), Align);
970 case Instruction::Free:
971 if (!isa<PointerType>(InstTy))
972 error("Invalid free instruction!");
973 Result = new FreeInst(getValue(iType, Oprnds[0]));
975 case Instruction::GetElementPtr: {
976 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy))
977 error("Invalid getelementptr instruction!");
979 std::vector<Value*> Idx;
981 const Type *NextTy = InstTy;
982 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
983 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy);
985 error("Invalid getelementptr instruction!");
987 unsigned ValIdx = Oprnds[i];
989 if (!hasRestrictedGEPTypes) {
990 // Struct indices are always uints, sequential type indices can be any
991 // of the 32 or 64-bit integer types. The actual choice of type is
992 // encoded in the low two bits of the slot number.
993 if (isa<StructType>(TopTy))
994 IdxTy = Type::UIntTyID;
996 switch (ValIdx & 3) {
998 case 0: IdxTy = Type::UIntTyID; break;
999 case 1: IdxTy = Type::IntTyID; break;
1000 case 2: IdxTy = Type::ULongTyID; break;
1001 case 3: IdxTy = Type::LongTyID; break;
1006 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID;
1009 Idx.push_back(getValue(IdxTy, ValIdx));
1011 // Convert ubyte struct indices into uint struct indices.
1012 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes)
1013 if (ConstantInt *C = dyn_cast<ConstantInt>(Idx.back()))
1014 if (C->getType() == Type::UByteTy)
1015 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy);
1017 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true);
1020 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx);
1024 case 62: // volatile load
1025 case Instruction::Load:
1026 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy))
1027 error("Invalid load instruction!");
1028 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62);
1031 case 63: // volatile store
1032 case Instruction::Store: {
1033 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2)
1034 error("Invalid store instruction!");
1036 Value *Ptr = getValue(iType, Oprnds[1]);
1037 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType();
1038 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr,
1042 case Instruction::Unwind:
1043 if (Oprnds.size() != 0) error("Invalid unwind instruction!");
1044 Result = new UnwindInst();
1046 case Instruction::Unreachable:
1047 if (Oprnds.size() != 0) error("Invalid unreachable instruction!");
1048 Result = new UnreachableInst();
1050 } // end switch(Opcode)
1052 BB->getInstList().push_back(Result);
1055 if (Result->getType() == InstTy)
1058 TypeSlot = getTypeSlot(Result->getType());
1060 insertValue(Result, TypeSlot, FunctionValues);
1063 /// Get a particular numbered basic block, which might be a forward reference.
1064 /// This works together with ParseBasicBlock to handle these forward references
1065 /// in a clean manner. This function is used when constructing phi, br, switch,
1066 /// and other instructions that reference basic blocks. Blocks are numbered
1067 /// sequentially as they appear in the function.
1068 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) {
1069 // Make sure there is room in the table...
1070 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1);
1072 // First check to see if this is a backwards reference, i.e., ParseBasicBlock
1073 // has already created this block, or if the forward reference has already
1075 if (ParsedBasicBlocks[ID])
1076 return ParsedBasicBlocks[ID];
1078 // Otherwise, the basic block has not yet been created. Do so and add it to
1079 // the ParsedBasicBlocks list.
1080 return ParsedBasicBlocks[ID] = new BasicBlock();
1083 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time.
1084 /// This method reads in one of the basicblock packets. This method is not used
1085 /// for bytecode files after LLVM 1.0
1086 /// @returns The basic block constructed.
1087 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) {
1088 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1092 if (ParsedBasicBlocks.size() == BlockNo)
1093 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1094 else if (ParsedBasicBlocks[BlockNo] == 0)
1095 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1097 BB = ParsedBasicBlocks[BlockNo];
1099 std::vector<unsigned> Operands;
1100 while (moreInBlock())
1101 ParseInstruction(Operands, BB);
1103 if (Handler) Handler->handleBasicBlockEnd(BlockNo);
1107 /// Parse all of the BasicBlock's & Instruction's in the body of a function.
1108 /// In post 1.0 bytecode files, we no longer emit basic block individually,
1109 /// in order to avoid per-basic-block overhead.
1110 /// @returns Rhe number of basic blocks encountered.
1111 unsigned BytecodeReader::ParseInstructionList(Function* F) {
1112 unsigned BlockNo = 0;
1113 std::vector<unsigned> Args;
1115 while (moreInBlock()) {
1116 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1118 if (ParsedBasicBlocks.size() == BlockNo)
1119 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1120 else if (ParsedBasicBlocks[BlockNo] == 0)
1121 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1123 BB = ParsedBasicBlocks[BlockNo];
1125 F->getBasicBlockList().push_back(BB);
1127 // Read instructions into this basic block until we get to a terminator
1128 while (moreInBlock() && !BB->getTerminator())
1129 ParseInstruction(Args, BB);
1131 if (!BB->getTerminator())
1132 error("Non-terminated basic block found!");
1134 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1);
1140 /// Parse a symbol table. This works for both module level and function
1141 /// level symbol tables. For function level symbol tables, the CurrentFunction
1142 /// parameter must be non-zero and the ST parameter must correspond to
1143 /// CurrentFunction's symbol table. For Module level symbol tables, the
1144 /// CurrentFunction argument must be zero.
1145 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction,
1147 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST);
1149 // Allow efficient basic block lookup by number.
1150 std::vector<BasicBlock*> BBMap;
1151 if (CurrentFunction)
1152 for (Function::iterator I = CurrentFunction->begin(),
1153 E = CurrentFunction->end(); I != E; ++I)
1156 /// In LLVM 1.3 we write types separately from values so
1157 /// The types are always first in the symbol table. This is
1158 /// because Type no longer derives from Value.
1159 if (!hasTypeDerivedFromValue) {
1160 // Symtab block header: [num entries]
1161 unsigned NumEntries = read_vbr_uint();
1162 for (unsigned i = 0; i < NumEntries; ++i) {
1163 // Symtab entry: [def slot #][name]
1164 unsigned slot = read_vbr_uint();
1165 std::string Name = read_str();
1166 const Type* T = getType(slot);
1167 ST->insert(Name, T);
1171 while (moreInBlock()) {
1172 // Symtab block header: [num entries][type id number]
1173 unsigned NumEntries = read_vbr_uint();
1175 bool isTypeType = read_typeid(Typ);
1176 const Type *Ty = getType(Typ);
1178 for (unsigned i = 0; i != NumEntries; ++i) {
1179 // Symtab entry: [def slot #][name]
1180 unsigned slot = read_vbr_uint();
1181 std::string Name = read_str();
1183 // if we're reading a pre 1.3 bytecode file and the type plane
1184 // is the "type type", handle it here
1186 const Type* T = getType(slot);
1188 error("Failed type look-up for name '" + Name + "'");
1189 ST->insert(Name, T);
1190 continue; // code below must be short circuited
1193 if (Typ == Type::LabelTyID) {
1194 if (slot < BBMap.size())
1197 V = getValue(Typ, slot, false); // Find mapping...
1200 error("Failed value look-up for name '" + Name + "'");
1205 checkPastBlockEnd("Symbol Table");
1206 if (Handler) Handler->handleSymbolTableEnd();
1209 /// Read in the types portion of a compaction table.
1210 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) {
1211 for (unsigned i = 0; i != NumEntries; ++i) {
1212 unsigned TypeSlot = 0;
1213 if (read_typeid(TypeSlot))
1214 error("Invalid type in compaction table: type type");
1215 const Type *Typ = getGlobalTableType(TypeSlot);
1216 CompactionTypes.push_back(std::make_pair(Typ, TypeSlot));
1217 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ);
1221 /// Parse a compaction table.
1222 void BytecodeReader::ParseCompactionTable() {
1224 // Notify handler that we're beginning a compaction table.
1225 if (Handler) Handler->handleCompactionTableBegin();
1227 // In LLVM 1.3 Type no longer derives from Value. So,
1228 // we always write them first in the compaction table
1229 // because they can't occupy a "type plane" where the
1231 if (! hasTypeDerivedFromValue) {
1232 unsigned NumEntries = read_vbr_uint();
1233 ParseCompactionTypes(NumEntries);
1236 // Compaction tables live in separate blocks so we have to loop
1237 // until we've read the whole thing.
1238 while (moreInBlock()) {
1239 // Read the number of Value* entries in the compaction table
1240 unsigned NumEntries = read_vbr_uint();
1242 unsigned isTypeType = false;
1244 // Decode the type from value read in. Most compaction table
1245 // planes will have one or two entries in them. If that's the
1246 // case then the length is encoded in the bottom two bits and
1247 // the higher bits encode the type. This saves another VBR value.
1248 if ((NumEntries & 3) == 3) {
1249 // In this case, both low-order bits are set (value 3). This
1250 // is a signal that the typeid follows.
1252 isTypeType = read_typeid(Ty);
1254 // In this case, the low-order bits specify the number of entries
1255 // and the high order bits specify the type.
1256 Ty = NumEntries >> 2;
1257 isTypeType = sanitizeTypeId(Ty);
1261 // if we're reading a pre 1.3 bytecode file and the type plane
1262 // is the "type type", handle it here
1264 ParseCompactionTypes(NumEntries);
1266 // Make sure we have enough room for the plane.
1267 if (Ty >= CompactionValues.size())
1268 CompactionValues.resize(Ty+1);
1270 // Make sure the plane is empty or we have some kind of error.
1271 if (!CompactionValues[Ty].empty())
1272 error("Compaction table plane contains multiple entries!");
1274 // Notify handler about the plane.
1275 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries);
1277 // Push the implicit zero.
1278 CompactionValues[Ty].push_back(Constant::getNullValue(getType(Ty)));
1280 // Read in each of the entries, put them in the compaction table
1281 // and notify the handler that we have a new compaction table value.
1282 for (unsigned i = 0; i != NumEntries; ++i) {
1283 unsigned ValSlot = read_vbr_uint();
1284 Value *V = getGlobalTableValue(Ty, ValSlot);
1285 CompactionValues[Ty].push_back(V);
1286 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot);
1290 // Notify handler that the compaction table is done.
1291 if (Handler) Handler->handleCompactionTableEnd();
1294 // Parse a single type. The typeid is read in first. If its a primitive type
1295 // then nothing else needs to be read, we know how to instantiate it. If its
1296 // a derived type, then additional data is read to fill out the type
1298 const Type *BytecodeReader::ParseType() {
1299 unsigned PrimType = 0;
1300 if (read_typeid(PrimType))
1301 error("Invalid type (type type) in type constants!");
1303 const Type *Result = 0;
1304 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType)))
1308 case Type::FunctionTyID: {
1309 const Type *RetType = readSanitizedType();
1311 unsigned NumParams = read_vbr_uint();
1313 std::vector<const Type*> Params;
1315 Params.push_back(readSanitizedType());
1317 bool isVarArg = Params.size() && Params.back() == Type::VoidTy;
1318 if (isVarArg) Params.pop_back();
1320 Result = FunctionType::get(RetType, Params, isVarArg);
1323 case Type::ArrayTyID: {
1324 const Type *ElementType = readSanitizedType();
1325 unsigned NumElements = read_vbr_uint();
1326 Result = ArrayType::get(ElementType, NumElements);
1329 case Type::PackedTyID: {
1330 const Type *ElementType = readSanitizedType();
1331 unsigned NumElements = read_vbr_uint();
1332 Result = PackedType::get(ElementType, NumElements);
1335 case Type::StructTyID: {
1336 std::vector<const Type*> Elements;
1338 if (read_typeid(Typ))
1339 error("Invalid element type (type type) for structure!");
1341 while (Typ) { // List is terminated by void/0 typeid
1342 Elements.push_back(getType(Typ));
1343 if (read_typeid(Typ))
1344 error("Invalid element type (type type) for structure!");
1347 Result = StructType::get(Elements);
1350 case Type::PointerTyID: {
1351 Result = PointerType::get(readSanitizedType());
1355 case Type::OpaqueTyID: {
1356 Result = OpaqueType::get();
1361 error("Don't know how to deserialize primitive type " + utostr(PrimType));
1364 if (Handler) Handler->handleType(Result);
1368 // ParseTypes - We have to use this weird code to handle recursive
1369 // types. We know that recursive types will only reference the current slab of
1370 // values in the type plane, but they can forward reference types before they
1371 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might
1372 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix
1373 // this ugly problem, we pessimistically insert an opaque type for each type we
1374 // are about to read. This means that forward references will resolve to
1375 // something and when we reread the type later, we can replace the opaque type
1376 // with a new resolved concrete type.
1378 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){
1379 assert(Tab.size() == 0 && "should not have read type constants in before!");
1381 // Insert a bunch of opaque types to be resolved later...
1382 Tab.reserve(NumEntries);
1383 for (unsigned i = 0; i != NumEntries; ++i)
1384 Tab.push_back(OpaqueType::get());
1387 Handler->handleTypeList(NumEntries);
1389 // If we are about to resolve types, make sure the type cache is clear.
1391 ModuleTypeIDCache.clear();
1393 // Loop through reading all of the types. Forward types will make use of the
1394 // opaque types just inserted.
1396 for (unsigned i = 0; i != NumEntries; ++i) {
1397 const Type* NewTy = ParseType();
1398 const Type* OldTy = Tab[i].get();
1400 error("Couldn't parse type!");
1402 // Don't directly push the new type on the Tab. Instead we want to replace
1403 // the opaque type we previously inserted with the new concrete value. This
1404 // approach helps with forward references to types. The refinement from the
1405 // abstract (opaque) type to the new type causes all uses of the abstract
1406 // type to use the concrete type (NewTy). This will also cause the opaque
1407 // type to be deleted.
1408 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy);
1410 // This should have replaced the old opaque type with the new type in the
1411 // value table... or with a preexisting type that was already in the system.
1412 // Let's just make sure it did.
1413 assert(Tab[i] != OldTy && "refineAbstractType didn't work!");
1417 /// Parse a single constant value
1418 Value *BytecodeReader::ParseConstantPoolValue(unsigned TypeID) {
1419 // We must check for a ConstantExpr before switching by type because
1420 // a ConstantExpr can be of any type, and has no explicit value.
1422 // 0 if not expr; numArgs if is expr
1423 unsigned isExprNumArgs = read_vbr_uint();
1425 if (isExprNumArgs) {
1426 if (!hasNoUndefValue) {
1427 // 'undef' is encoded with 'exprnumargs' == 1.
1428 if (isExprNumArgs == 1)
1429 return UndefValue::get(getType(TypeID));
1431 // Inline asm is encoded with exprnumargs == ~0U.
1432 if (isExprNumArgs == ~0U) {
1433 std::string AsmStr = read_str();
1434 std::string ConstraintStr = read_str();
1435 unsigned Flags = read_vbr_uint();
1437 const PointerType *PTy = dyn_cast<PointerType>(getType(TypeID));
1438 const FunctionType *FTy =
1439 PTy ? dyn_cast<FunctionType>(PTy->getElementType()) : 0;
1441 if (!FTy || !InlineAsm::Verify(FTy, ConstraintStr))
1442 error("Invalid constraints for inline asm");
1444 error("Invalid flags for inline asm");
1445 bool HasSideEffects = Flags & 1;
1446 return InlineAsm::get(FTy, AsmStr, ConstraintStr, HasSideEffects);
1452 // FIXME: Encoding of constant exprs could be much more compact!
1453 std::vector<Constant*> ArgVec;
1454 ArgVec.reserve(isExprNumArgs);
1455 unsigned Opcode = read_vbr_uint();
1457 // Bytecode files before LLVM 1.4 need have a missing terminator inst.
1458 if (hasNoUnreachableInst) Opcode++;
1460 // Read the slot number and types of each of the arguments
1461 for (unsigned i = 0; i != isExprNumArgs; ++i) {
1462 unsigned ArgValSlot = read_vbr_uint();
1463 unsigned ArgTypeSlot = 0;
1464 if (read_typeid(ArgTypeSlot))
1465 error("Invalid argument type (type type) for constant value");
1467 // Get the arg value from its slot if it exists, otherwise a placeholder
1468 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot));
1471 // Construct a ConstantExpr of the appropriate kind
1472 if (isExprNumArgs == 1) { // All one-operand expressions
1473 if (Opcode != Instruction::Cast)
1474 error("Only cast instruction has one argument for ConstantExpr");
1476 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID));
1477 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1479 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr
1480 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end());
1482 if (hasRestrictedGEPTypes) {
1483 const Type *BaseTy = ArgVec[0]->getType();
1484 generic_gep_type_iterator<std::vector<Constant*>::iterator>
1485 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()),
1486 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end());
1487 for (unsigned i = 0; GTI != E; ++GTI, ++i)
1488 if (isa<StructType>(*GTI)) {
1489 if (IdxList[i]->getType() != Type::UByteTy)
1490 error("Invalid index for getelementptr!");
1491 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy);
1495 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList);
1496 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1498 } else if (Opcode == Instruction::Select) {
1499 if (ArgVec.size() != 3)
1500 error("Select instruction must have three arguments.");
1501 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1],
1503 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1505 } else if (Opcode == Instruction::ExtractElement) {
1506 if (ArgVec.size() != 2 ||
1507 !ExtractElementInst::isValidOperands(ArgVec[0], ArgVec[1]))
1508 error("Invalid extractelement constand expr arguments");
1509 Constant* Result = ConstantExpr::getExtractElement(ArgVec[0], ArgVec[1]);
1510 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1512 } else if (Opcode == Instruction::InsertElement) {
1513 if (ArgVec.size() != 3 ||
1514 !InsertElementInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2]))
1515 error("Invalid insertelement constand expr arguments");
1518 ConstantExpr::getInsertElement(ArgVec[0], ArgVec[1], ArgVec[2]);
1519 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1521 } else if (Opcode == Instruction::ShuffleVector) {
1522 if (ArgVec.size() != 3 ||
1523 !ShuffleVectorInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2]))
1524 error("Invalid shufflevector constant expr arguments.");
1526 ConstantExpr::getShuffleVector(ArgVec[0], ArgVec[1], ArgVec[2]);
1527 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1529 } else { // All other 2-operand expressions
1530 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]);
1531 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1536 // Ok, not an ConstantExpr. We now know how to read the given type...
1537 const Type *Ty = getType(TypeID);
1538 Constant *Result = 0;
1539 switch (Ty->getTypeID()) {
1540 case Type::BoolTyID: {
1541 unsigned Val = read_vbr_uint();
1542 if (Val != 0 && Val != 1)
1543 error("Invalid boolean value read.");
1544 Result = ConstantBool::get(Val == 1);
1545 if (Handler) Handler->handleConstantValue(Result);
1549 case Type::UByteTyID: // Unsigned integer types...
1550 case Type::UShortTyID:
1551 case Type::UIntTyID: {
1552 unsigned Val = read_vbr_uint();
1553 if (!ConstantInt::isValueValidForType(Ty, uint64_t(Val)))
1554 error("Invalid unsigned byte/short/int read.");
1555 Result = ConstantInt::get(Ty, Val);
1556 if (Handler) Handler->handleConstantValue(Result);
1560 case Type::ULongTyID:
1561 Result = ConstantInt::get(Ty, read_vbr_uint64());
1562 if (Handler) Handler->handleConstantValue(Result);
1565 case Type::SByteTyID: // Signed integer types...
1566 case Type::ShortTyID:
1568 case Type::LongTyID: {
1569 int64_t Val = read_vbr_int64();
1570 if (!ConstantInt::isValueValidForType(Ty, Val))
1571 error("Invalid signed byte/short/int/long read.");
1572 Result = ConstantInt::get(Ty, Val);
1573 if (Handler) Handler->handleConstantValue(Result);
1577 case Type::FloatTyID: {
1580 Result = ConstantFP::get(Ty, Val);
1581 if (Handler) Handler->handleConstantValue(Result);
1585 case Type::DoubleTyID: {
1588 Result = ConstantFP::get(Ty, Val);
1589 if (Handler) Handler->handleConstantValue(Result);
1593 case Type::ArrayTyID: {
1594 const ArrayType *AT = cast<ArrayType>(Ty);
1595 unsigned NumElements = AT->getNumElements();
1596 unsigned TypeSlot = getTypeSlot(AT->getElementType());
1597 std::vector<Constant*> Elements;
1598 Elements.reserve(NumElements);
1599 while (NumElements--) // Read all of the elements of the constant.
1600 Elements.push_back(getConstantValue(TypeSlot,
1602 Result = ConstantArray::get(AT, Elements);
1603 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result);
1607 case Type::StructTyID: {
1608 const StructType *ST = cast<StructType>(Ty);
1610 std::vector<Constant *> Elements;
1611 Elements.reserve(ST->getNumElements());
1612 for (unsigned i = 0; i != ST->getNumElements(); ++i)
1613 Elements.push_back(getConstantValue(ST->getElementType(i),
1616 Result = ConstantStruct::get(ST, Elements);
1617 if (Handler) Handler->handleConstantStruct(ST, Elements, Result);
1621 case Type::PackedTyID: {
1622 const PackedType *PT = cast<PackedType>(Ty);
1623 unsigned NumElements = PT->getNumElements();
1624 unsigned TypeSlot = getTypeSlot(PT->getElementType());
1625 std::vector<Constant*> Elements;
1626 Elements.reserve(NumElements);
1627 while (NumElements--) // Read all of the elements of the constant.
1628 Elements.push_back(getConstantValue(TypeSlot,
1630 Result = ConstantPacked::get(PT, Elements);
1631 if (Handler) Handler->handleConstantPacked(PT, Elements, TypeSlot, Result);
1635 case Type::PointerTyID: { // ConstantPointerRef value (backwards compat).
1636 const PointerType *PT = cast<PointerType>(Ty);
1637 unsigned Slot = read_vbr_uint();
1639 // Check to see if we have already read this global variable...
1640 Value *Val = getValue(TypeID, Slot, false);
1642 GlobalValue *GV = dyn_cast<GlobalValue>(Val);
1643 if (!GV) error("GlobalValue not in ValueTable!");
1644 if (Handler) Handler->handleConstantPointer(PT, Slot, GV);
1647 error("Forward references are not allowed here.");
1652 error("Don't know how to deserialize constant value of type '" +
1653 Ty->getDescription());
1657 // Check that we didn't read a null constant if they are implicit for this
1658 // type plane. Do not do this check for constantexprs, as they may be folded
1659 // to a null value in a way that isn't predicted when a .bc file is initially
1661 assert((!isa<Constant>(Result) || !cast<Constant>(Result)->isNullValue()) ||
1662 !hasImplicitNull(TypeID) &&
1663 "Cannot read null values from bytecode!");
1667 /// Resolve references for constants. This function resolves the forward
1668 /// referenced constants in the ConstantFwdRefs map. It uses the
1669 /// replaceAllUsesWith method of Value class to substitute the placeholder
1670 /// instance with the actual instance.
1671 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Typ,
1673 ConstantRefsType::iterator I =
1674 ConstantFwdRefs.find(std::make_pair(Typ, Slot));
1675 if (I == ConstantFwdRefs.end()) return; // Never forward referenced?
1677 Value *PH = I->second; // Get the placeholder...
1678 PH->replaceAllUsesWith(NewV);
1679 delete PH; // Delete the old placeholder
1680 ConstantFwdRefs.erase(I); // Remove the map entry for it
1683 /// Parse the constant strings section.
1684 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){
1685 for (; NumEntries; --NumEntries) {
1687 if (read_typeid(Typ))
1688 error("Invalid type (type type) for string constant");
1689 const Type *Ty = getType(Typ);
1690 if (!isa<ArrayType>(Ty))
1691 error("String constant data invalid!");
1693 const ArrayType *ATy = cast<ArrayType>(Ty);
1694 if (ATy->getElementType() != Type::SByteTy &&
1695 ATy->getElementType() != Type::UByteTy)
1696 error("String constant data invalid!");
1698 // Read character data. The type tells us how long the string is.
1699 char *Data = reinterpret_cast<char *>(alloca(ATy->getNumElements()));
1700 read_data(Data, Data+ATy->getNumElements());
1702 std::vector<Constant*> Elements(ATy->getNumElements());
1703 const Type* ElemType = ATy->getElementType();
1704 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
1705 Elements[i] = ConstantInt::get(ElemType, (unsigned char)Data[i]);
1707 // Create the constant, inserting it as needed.
1708 Constant *C = ConstantArray::get(ATy, Elements);
1709 unsigned Slot = insertValue(C, Typ, Tab);
1710 ResolveReferencesToConstant(C, Typ, Slot);
1711 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C));
1715 /// Parse the constant pool.
1716 void BytecodeReader::ParseConstantPool(ValueTable &Tab,
1717 TypeListTy &TypeTab,
1719 if (Handler) Handler->handleGlobalConstantsBegin();
1721 /// In LLVM 1.3 Type does not derive from Value so the types
1722 /// do not occupy a plane. Consequently, we read the types
1723 /// first in the constant pool.
1724 if (isFunction && !hasTypeDerivedFromValue) {
1725 unsigned NumEntries = read_vbr_uint();
1726 ParseTypes(TypeTab, NumEntries);
1729 while (moreInBlock()) {
1730 unsigned NumEntries = read_vbr_uint();
1732 bool isTypeType = read_typeid(Typ);
1734 /// In LLVM 1.2 and before, Types were written to the
1735 /// bytecode file in the "Type Type" plane (#12).
1736 /// In 1.3 plane 12 is now the label plane. Handle this here.
1738 ParseTypes(TypeTab, NumEntries);
1739 } else if (Typ == Type::VoidTyID) {
1740 /// Use of Type::VoidTyID is a misnomer. It actually means
1741 /// that the following plane is constant strings
1742 assert(&Tab == &ModuleValues && "Cannot read strings in functions!");
1743 ParseStringConstants(NumEntries, Tab);
1745 for (unsigned i = 0; i < NumEntries; ++i) {
1746 Value *V = ParseConstantPoolValue(Typ);
1747 assert(V && "ParseConstantPoolValue returned NULL!");
1748 unsigned Slot = insertValue(V, Typ, Tab);
1750 // If we are reading a function constant table, make sure that we adjust
1751 // the slot number to be the real global constant number.
1753 if (&Tab != &ModuleValues && Typ < ModuleValues.size() &&
1755 Slot += ModuleValues[Typ]->size();
1756 if (Constant *C = dyn_cast<Constant>(V))
1757 ResolveReferencesToConstant(C, Typ, Slot);
1762 // After we have finished parsing the constant pool, we had better not have
1763 // any dangling references left.
1764 if (!ConstantFwdRefs.empty()) {
1765 ConstantRefsType::const_iterator I = ConstantFwdRefs.begin();
1766 Constant* missingConst = I->second;
1767 error(utostr(ConstantFwdRefs.size()) +
1768 " unresolved constant reference exist. First one is '" +
1769 missingConst->getName() + "' of type '" +
1770 missingConst->getType()->getDescription() + "'.");
1773 checkPastBlockEnd("Constant Pool");
1774 if (Handler) Handler->handleGlobalConstantsEnd();
1777 /// Parse the contents of a function. Note that this function can be
1778 /// called lazily by materializeFunction
1779 /// @see materializeFunction
1780 void BytecodeReader::ParseFunctionBody(Function* F) {
1782 unsigned FuncSize = BlockEnd - At;
1783 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage;
1785 unsigned LinkageType = read_vbr_uint();
1786 switch (LinkageType) {
1787 case 0: Linkage = GlobalValue::ExternalLinkage; break;
1788 case 1: Linkage = GlobalValue::WeakLinkage; break;
1789 case 2: Linkage = GlobalValue::AppendingLinkage; break;
1790 case 3: Linkage = GlobalValue::InternalLinkage; break;
1791 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
1792 case 5: Linkage = GlobalValue::DLLImportLinkage; break;
1793 case 6: Linkage = GlobalValue::DLLExportLinkage; break;
1794 case 7: Linkage = GlobalValue::ExternalWeakLinkage; break;
1796 error("Invalid linkage type for Function.");
1797 Linkage = GlobalValue::InternalLinkage;
1801 F->setLinkage(Linkage);
1802 if (Handler) Handler->handleFunctionBegin(F,FuncSize);
1804 // Keep track of how many basic blocks we have read in...
1805 unsigned BlockNum = 0;
1806 bool InsertedArguments = false;
1808 BufPtr MyEnd = BlockEnd;
1809 while (At < MyEnd) {
1810 unsigned Type, Size;
1812 read_block(Type, Size);
1815 case BytecodeFormat::ConstantPoolBlockID:
1816 if (!InsertedArguments) {
1817 // Insert arguments into the value table before we parse the first basic
1818 // block in the function, but after we potentially read in the
1819 // compaction table.
1821 InsertedArguments = true;
1824 ParseConstantPool(FunctionValues, FunctionTypes, true);
1827 case BytecodeFormat::CompactionTableBlockID:
1828 ParseCompactionTable();
1831 case BytecodeFormat::BasicBlock: {
1832 if (!InsertedArguments) {
1833 // Insert arguments into the value table before we parse the first basic
1834 // block in the function, but after we potentially read in the
1835 // compaction table.
1837 InsertedArguments = true;
1840 BasicBlock *BB = ParseBasicBlock(BlockNum++);
1841 F->getBasicBlockList().push_back(BB);
1845 case BytecodeFormat::InstructionListBlockID: {
1846 // Insert arguments into the value table before we parse the instruction
1847 // list for the function, but after we potentially read in the compaction
1849 if (!InsertedArguments) {
1851 InsertedArguments = true;
1855 error("Already parsed basic blocks!");
1856 BlockNum = ParseInstructionList(F);
1860 case BytecodeFormat::SymbolTableBlockID:
1861 ParseSymbolTable(F, &F->getSymbolTable());
1867 error("Wrapped around reading bytecode.");
1872 // Malformed bc file if read past end of block.
1876 // Make sure there were no references to non-existant basic blocks.
1877 if (BlockNum != ParsedBasicBlocks.size())
1878 error("Illegal basic block operand reference");
1880 ParsedBasicBlocks.clear();
1882 // Resolve forward references. Replace any uses of a forward reference value
1883 // with the real value.
1884 while (!ForwardReferences.empty()) {
1885 std::map<std::pair<unsigned,unsigned>, Value*>::iterator
1886 I = ForwardReferences.begin();
1887 Value *V = getValue(I->first.first, I->first.second, false);
1888 Value *PlaceHolder = I->second;
1889 PlaceHolder->replaceAllUsesWith(V);
1890 ForwardReferences.erase(I);
1894 // If upgraded intrinsic functions were detected during reading of the
1895 // module information, then we need to look for instructions that need to
1896 // be upgraded. This can't be done while the instructions are read in because
1897 // additional instructions inserted mess up the slot numbering.
1898 if (!upgradedFunctions.empty()) {
1899 for (Function::iterator BI = F->begin(), BE = F->end(); BI != BE; ++BI)
1900 for (BasicBlock::iterator II = BI->begin(), IE = BI->end();
1902 if (CallInst* CI = dyn_cast<CallInst>(II++)) {
1903 std::map<Function*,Function*>::iterator FI =
1904 upgradedFunctions.find(CI->getCalledFunction());
1905 if (FI != upgradedFunctions.end())
1906 UpgradeIntrinsicCall(CI, FI->second);
1910 // Clear out function-level types...
1911 FunctionTypes.clear();
1912 CompactionTypes.clear();
1913 CompactionValues.clear();
1914 freeTable(FunctionValues);
1916 if (Handler) Handler->handleFunctionEnd(F);
1919 /// This function parses LLVM functions lazily. It obtains the type of the
1920 /// function and records where the body of the function is in the bytecode
1921 /// buffer. The caller can then use the ParseNextFunction and
1922 /// ParseAllFunctionBodies to get handler events for the functions.
1923 void BytecodeReader::ParseFunctionLazily() {
1924 if (FunctionSignatureList.empty())
1925 error("FunctionSignatureList empty!");
1927 Function *Func = FunctionSignatureList.back();
1928 FunctionSignatureList.pop_back();
1930 // Save the information for future reading of the function
1931 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd);
1933 // This function has a body but it's not loaded so it appears `External'.
1934 // Mark it as a `Ghost' instead to notify the users that it has a body.
1935 Func->setLinkage(GlobalValue::GhostLinkage);
1937 // Pretend we've `parsed' this function
1941 /// The ParserFunction method lazily parses one function. Use this method to
1942 /// casue the parser to parse a specific function in the module. Note that
1943 /// this will remove the function from what is to be included by
1944 /// ParseAllFunctionBodies.
1945 /// @see ParseAllFunctionBodies
1946 /// @see ParseBytecode
1947 bool BytecodeReader::ParseFunction(Function* Func, std::string* ErrMsg) {
1949 if (setjmp(context))
1952 // Find {start, end} pointers and slot in the map. If not there, we're done.
1953 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func);
1955 // Make sure we found it
1956 if (Fi == LazyFunctionLoadMap.end()) {
1957 error("Unrecognized function of type " + Func->getType()->getDescription());
1961 BlockStart = At = Fi->second.Buf;
1962 BlockEnd = Fi->second.EndBuf;
1963 assert(Fi->first == Func && "Found wrong function?");
1965 LazyFunctionLoadMap.erase(Fi);
1967 this->ParseFunctionBody(Func);
1971 /// The ParseAllFunctionBodies method parses through all the previously
1972 /// unparsed functions in the bytecode file. If you want to completely parse
1973 /// a bytecode file, this method should be called after Parsebytecode because
1974 /// Parsebytecode only records the locations in the bytecode file of where
1975 /// the function definitions are located. This function uses that information
1976 /// to materialize the functions.
1977 /// @see ParseBytecode
1978 bool BytecodeReader::ParseAllFunctionBodies(std::string* ErrMsg) {
1979 if (setjmp(context))
1982 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin();
1983 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end();
1986 Function* Func = Fi->first;
1987 BlockStart = At = Fi->second.Buf;
1988 BlockEnd = Fi->second.EndBuf;
1989 ParseFunctionBody(Func);
1992 LazyFunctionLoadMap.clear();
1996 /// Parse the global type list
1997 void BytecodeReader::ParseGlobalTypes() {
1998 // Read the number of types
1999 unsigned NumEntries = read_vbr_uint();
2001 // Ignore the type plane identifier for types if the bc file is pre 1.3
2002 if (hasTypeDerivedFromValue)
2005 ParseTypes(ModuleTypes, NumEntries);
2008 /// Parse the Global info (types, global vars, constants)
2009 void BytecodeReader::ParseModuleGlobalInfo() {
2011 if (Handler) Handler->handleModuleGlobalsBegin();
2013 // SectionID - If a global has an explicit section specified, this map
2014 // remembers the ID until we can translate it into a string.
2015 std::map<GlobalValue*, unsigned> SectionID;
2017 // Read global variables...
2018 unsigned VarType = read_vbr_uint();
2019 while (VarType != Type::VoidTyID) { // List is terminated by Void
2020 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 =
2021 // Linkage, bit4+ = slot#
2022 unsigned SlotNo = VarType >> 5;
2023 if (sanitizeTypeId(SlotNo))
2024 error("Invalid type (type type) for global var!");
2025 unsigned LinkageID = (VarType >> 2) & 7;
2026 bool isConstant = VarType & 1;
2027 bool hasInitializer = (VarType & 2) != 0;
2028 unsigned Alignment = 0;
2029 unsigned GlobalSectionID = 0;
2031 // An extension word is present when linkage = 3 (internal) and hasinit = 0.
2032 if (LinkageID == 3 && !hasInitializer) {
2033 unsigned ExtWord = read_vbr_uint();
2034 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
2035 // linkage, bit 4-8 = alignment (log2), bits 10+ = future use.
2036 hasInitializer = ExtWord & 1;
2037 LinkageID = (ExtWord >> 1) & 7;
2038 Alignment = (1 << ((ExtWord >> 4) & 31)) >> 1;
2040 if (ExtWord & (1 << 9)) // Has a section ID.
2041 GlobalSectionID = read_vbr_uint();
2044 GlobalValue::LinkageTypes Linkage;
2045 switch (LinkageID) {
2046 case 0: Linkage = GlobalValue::ExternalLinkage; break;
2047 case 1: Linkage = GlobalValue::WeakLinkage; break;
2048 case 2: Linkage = GlobalValue::AppendingLinkage; break;
2049 case 3: Linkage = GlobalValue::InternalLinkage; break;
2050 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
2051 case 5: Linkage = GlobalValue::DLLImportLinkage; break;
2052 case 6: Linkage = GlobalValue::DLLExportLinkage; break;
2053 case 7: Linkage = GlobalValue::ExternalWeakLinkage; break;
2055 error("Unknown linkage type: " + utostr(LinkageID));
2056 Linkage = GlobalValue::InternalLinkage;
2060 const Type *Ty = getType(SlotNo);
2062 error("Global has no type! SlotNo=" + utostr(SlotNo));
2064 if (!isa<PointerType>(Ty))
2065 error("Global not a pointer type! Ty= " + Ty->getDescription());
2067 const Type *ElTy = cast<PointerType>(Ty)->getElementType();
2069 // Create the global variable...
2070 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage,
2072 GV->setAlignment(Alignment);
2073 insertValue(GV, SlotNo, ModuleValues);
2075 if (GlobalSectionID != 0)
2076 SectionID[GV] = GlobalSectionID;
2078 unsigned initSlot = 0;
2079 if (hasInitializer) {
2080 initSlot = read_vbr_uint();
2081 GlobalInits.push_back(std::make_pair(GV, initSlot));
2084 // Notify handler about the global value.
2086 Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo,initSlot);
2089 VarType = read_vbr_uint();
2092 // Read the function objects for all of the functions that are coming
2093 unsigned FnSignature = read_vbr_uint();
2095 if (hasNoFlagsForFunctions)
2096 FnSignature = (FnSignature << 5) + 1;
2098 // List is terminated by VoidTy.
2099 while (((FnSignature & (~0U >> 1)) >> 5) != Type::VoidTyID) {
2100 const Type *Ty = getType((FnSignature & (~0U >> 1)) >> 5);
2101 if (!isa<PointerType>(Ty) ||
2102 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) {
2103 error("Function not a pointer to function type! Ty = " +
2104 Ty->getDescription());
2107 // We create functions by passing the underlying FunctionType to create...
2108 const FunctionType* FTy =
2109 cast<FunctionType>(cast<PointerType>(Ty)->getElementType());
2111 // Insert the place holder.
2112 Function *Func = new Function(FTy, GlobalValue::ExternalLinkage,
2115 insertValue(Func, (FnSignature & (~0U >> 1)) >> 5, ModuleValues);
2117 // Flags are not used yet.
2118 unsigned Flags = FnSignature & 31;
2120 // Save this for later so we know type of lazily instantiated functions.
2121 // Note that known-external functions do not have FunctionInfo blocks, so we
2122 // do not add them to the FunctionSignatureList.
2123 if ((Flags & (1 << 4)) == 0)
2124 FunctionSignatureList.push_back(Func);
2126 // Get the calling convention from the low bits.
2127 unsigned CC = Flags & 15;
2128 unsigned Alignment = 0;
2129 if (FnSignature & (1 << 31)) { // Has extension word?
2130 unsigned ExtWord = read_vbr_uint();
2131 Alignment = (1 << (ExtWord & 31)) >> 1;
2132 CC |= ((ExtWord >> 5) & 15) << 4;
2134 if (ExtWord & (1 << 10)) // Has a section ID.
2135 SectionID[Func] = read_vbr_uint();
2137 // Parse external declaration linkage
2138 switch ((ExtWord >> 11) & 3) {
2140 case 1: Func->setLinkage(Function::DLLImportLinkage); break;
2141 case 2: Func->setLinkage(Function::ExternalWeakLinkage); break;
2142 default: assert(0 && "Unsupported external linkage");
2146 Func->setCallingConv(CC-1);
2147 Func->setAlignment(Alignment);
2149 if (Handler) Handler->handleFunctionDeclaration(Func);
2151 // Get the next function signature.
2152 FnSignature = read_vbr_uint();
2153 if (hasNoFlagsForFunctions)
2154 FnSignature = (FnSignature << 5) + 1;
2157 // Now that the function signature list is set up, reverse it so that we can
2158 // remove elements efficiently from the back of the vector.
2159 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end());
2161 /// SectionNames - This contains the list of section names encoded in the
2162 /// moduleinfoblock. Functions and globals with an explicit section index
2163 /// into this to get their section name.
2164 std::vector<std::string> SectionNames;
2166 if (hasInconsistentModuleGlobalInfo) {
2168 } else if (!hasNoDependentLibraries) {
2169 // If this bytecode format has dependent library information in it, read in
2170 // the number of dependent library items that follow.
2171 unsigned num_dep_libs = read_vbr_uint();
2172 std::string dep_lib;
2173 while (num_dep_libs--) {
2174 dep_lib = read_str();
2175 TheModule->addLibrary(dep_lib);
2177 Handler->handleDependentLibrary(dep_lib);
2180 // Read target triple and place into the module.
2181 std::string triple = read_str();
2182 TheModule->setTargetTriple(triple);
2184 Handler->handleTargetTriple(triple);
2186 if (!hasAlignment && At != BlockEnd) {
2187 // If the file has section info in it, read the section names now.
2188 unsigned NumSections = read_vbr_uint();
2189 while (NumSections--)
2190 SectionNames.push_back(read_str());
2193 // If the file has module-level inline asm, read it now.
2194 if (!hasAlignment && At != BlockEnd)
2195 TheModule->setModuleInlineAsm(read_str());
2198 // If any globals are in specified sections, assign them now.
2199 for (std::map<GlobalValue*, unsigned>::iterator I = SectionID.begin(), E =
2200 SectionID.end(); I != E; ++I)
2202 if (I->second > SectionID.size())
2203 error("SectionID out of range for global!");
2204 I->first->setSection(SectionNames[I->second-1]);
2207 // This is for future proofing... in the future extra fields may be added that
2208 // we don't understand, so we transparently ignore them.
2212 if (Handler) Handler->handleModuleGlobalsEnd();
2215 /// Parse the version information and decode it by setting flags on the
2216 /// Reader that enable backward compatibility of the reader.
2217 void BytecodeReader::ParseVersionInfo() {
2218 unsigned Version = read_vbr_uint();
2220 // Unpack version number: low four bits are for flags, top bits = version
2221 Module::Endianness Endianness;
2222 Module::PointerSize PointerSize;
2223 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian;
2224 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32;
2226 bool hasNoEndianness = Version & 4;
2227 bool hasNoPointerSize = Version & 8;
2229 RevisionNum = Version >> 4;
2231 // Default values for the current bytecode version
2232 hasInconsistentModuleGlobalInfo = false;
2233 hasExplicitPrimitiveZeros = false;
2234 hasRestrictedGEPTypes = false;
2235 hasTypeDerivedFromValue = false;
2236 hasLongBlockHeaders = false;
2237 has32BitTypes = false;
2238 hasNoDependentLibraries = false;
2239 hasAlignment = false;
2240 hasNoUndefValue = false;
2241 hasNoFlagsForFunctions = false;
2242 hasNoUnreachableInst = false;
2244 switch (RevisionNum) {
2245 case 0: // LLVM 1.0, 1.1 (Released)
2246 // Base LLVM 1.0 bytecode format.
2247 hasInconsistentModuleGlobalInfo = true;
2248 hasExplicitPrimitiveZeros = true;
2252 case 1: // LLVM 1.2 (Released)
2253 // LLVM 1.2 added explicit support for emitting strings efficiently.
2255 // Also, it fixed the problem where the size of the ModuleGlobalInfo block
2256 // included the size for the alignment at the end, where the rest of the
2259 // LLVM 1.2 and before required that GEP indices be ubyte constants for
2260 // structures and longs for sequential types.
2261 hasRestrictedGEPTypes = true;
2263 // LLVM 1.2 and before had the Type class derive from Value class. This
2264 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are
2265 // written differently because Types can no longer be part of the
2266 // type planes for Values.
2267 hasTypeDerivedFromValue = true;
2271 case 2: // 1.2.5 (Not Released)
2273 // LLVM 1.2 and earlier had two-word block headers. This is a bit wasteful,
2274 // especially for small files where the 8 bytes per block is a large
2275 // fraction of the total block size. In LLVM 1.3, the block type and length
2276 // are compressed into a single 32-bit unsigned integer. 27 bits for length,
2277 // 5 bits for block type.
2278 hasLongBlockHeaders = true;
2280 // LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
2281 // this has been reduced to vbr_uint24. It shouldn't make much difference
2282 // since we haven't run into a module with > 24 million types, but for
2283 // safety the 24-bit restriction has been enforced in 1.3 to free some bits
2284 // in various places and to ensure consistency.
2285 has32BitTypes = true;
2287 // LLVM 1.2 and earlier did not provide a target triple nor a list of
2288 // libraries on which the bytecode is dependent. LLVM 1.3 provides these
2289 // features, for use in future versions of LLVM.
2290 hasNoDependentLibraries = true;
2294 case 3: // LLVM 1.3 (Released)
2295 // LLVM 1.3 and earlier caused alignment bytes to be written on some block
2296 // boundaries and at the end of some strings. In extreme cases (e.g. lots
2297 // of GEP references to a constant array), this can increase the file size
2298 // by 30% or more. In version 1.4 alignment is done away with completely.
2299 hasAlignment = true;
2303 case 4: // 1.3.1 (Not Released)
2304 // In version 4, we did not support the 'undef' constant.
2305 hasNoUndefValue = true;
2307 // In version 4 and above, we did not include space for flags for functions
2308 // in the module info block.
2309 hasNoFlagsForFunctions = true;
2311 // In version 4 and above, we did not include the 'unreachable' instruction
2312 // in the opcode numbering in the bytecode file.
2313 hasNoUnreachableInst = true;
2318 case 5: // 1.4 (Released)
2322 error("Unknown bytecode version number: " + itostr(RevisionNum));
2325 if (hasNoEndianness) Endianness = Module::AnyEndianness;
2326 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize;
2328 TheModule->setEndianness(Endianness);
2329 TheModule->setPointerSize(PointerSize);
2331 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize);
2334 /// Parse a whole module.
2335 void BytecodeReader::ParseModule() {
2336 unsigned Type, Size;
2338 FunctionSignatureList.clear(); // Just in case...
2340 // Read into instance variables...
2344 bool SeenModuleGlobalInfo = false;
2345 bool SeenGlobalTypePlane = false;
2346 BufPtr MyEnd = BlockEnd;
2347 while (At < MyEnd) {
2349 read_block(Type, Size);
2353 case BytecodeFormat::GlobalTypePlaneBlockID:
2354 if (SeenGlobalTypePlane)
2355 error("Two GlobalTypePlane Blocks Encountered!");
2359 SeenGlobalTypePlane = true;
2362 case BytecodeFormat::ModuleGlobalInfoBlockID:
2363 if (SeenModuleGlobalInfo)
2364 error("Two ModuleGlobalInfo Blocks Encountered!");
2365 ParseModuleGlobalInfo();
2366 SeenModuleGlobalInfo = true;
2369 case BytecodeFormat::ConstantPoolBlockID:
2370 ParseConstantPool(ModuleValues, ModuleTypes,false);
2373 case BytecodeFormat::FunctionBlockID:
2374 ParseFunctionLazily();
2377 case BytecodeFormat::SymbolTableBlockID:
2378 ParseSymbolTable(0, &TheModule->getSymbolTable());
2384 error("Unexpected Block of Type #" + utostr(Type) + " encountered!");
2392 // After the module constant pool has been read, we can safely initialize
2393 // global variables...
2394 while (!GlobalInits.empty()) {
2395 GlobalVariable *GV = GlobalInits.back().first;
2396 unsigned Slot = GlobalInits.back().second;
2397 GlobalInits.pop_back();
2399 // Look up the initializer value...
2400 // FIXME: Preserve this type ID!
2402 const llvm::PointerType* GVType = GV->getType();
2403 unsigned TypeSlot = getTypeSlot(GVType->getElementType());
2404 if (Constant *CV = getConstantValue(TypeSlot, Slot)) {
2405 if (GV->hasInitializer())
2406 error("Global *already* has an initializer?!");
2407 if (Handler) Handler->handleGlobalInitializer(GV,CV);
2408 GV->setInitializer(CV);
2410 error("Cannot find initializer value.");
2413 if (!ConstantFwdRefs.empty())
2414 error("Use of undefined constants in a module");
2416 /// Make sure we pulled them all out. If we didn't then there's a declaration
2417 /// but a missing body. That's not allowed.
2418 if (!FunctionSignatureList.empty())
2419 error("Function declared, but bytecode stream ended before definition");
2422 /// This function completely parses a bytecode buffer given by the \p Buf
2423 /// and \p Length parameters.
2424 bool BytecodeReader::ParseBytecode(volatile BufPtr Buf, unsigned Length,
2425 const std::string &ModuleID,
2426 std::string* ErrMsg) {
2428 /// We handle errors by
2429 if (setjmp(context)) {
2430 // Cleanup after error
2431 if (Handler) Handler->handleError(ErrorMsg);
2435 if (decompressedBlock != 0 ) {
2436 ::free(decompressedBlock);
2437 decompressedBlock = 0;
2439 // Set caller's error message, if requested
2442 // Indicate an error occurred
2447 At = MemStart = BlockStart = Buf;
2448 MemEnd = BlockEnd = Buf + Length;
2450 // Create the module
2451 TheModule = new Module(ModuleID);
2453 if (Handler) Handler->handleStart(TheModule, Length);
2455 // Read the four bytes of the signature.
2456 unsigned Sig = read_uint();
2458 // If this is a compressed file
2459 if (Sig == ('l' | ('l' << 8) | ('v' << 16) | ('c' << 24))) {
2461 // Invoke the decompression of the bytecode. Note that we have to skip the
2462 // file's magic number which is not part of the compressed block. Hence,
2463 // the Buf+4 and Length-4. The result goes into decompressedBlock, a data
2464 // member for retention until BytecodeReader is destructed.
2465 unsigned decompressedLength = Compressor::decompressToNewBuffer(
2466 (char*)Buf+4,Length-4,decompressedBlock);
2468 // We must adjust the buffer pointers used by the bytecode reader to point
2469 // into the new decompressed block. After decompression, the
2470 // decompressedBlock will point to a contiguous memory area that has
2471 // the decompressed data.
2472 At = MemStart = BlockStart = Buf = (BufPtr) decompressedBlock;
2473 MemEnd = BlockEnd = Buf + decompressedLength;
2475 // else if this isn't a regular (uncompressed) bytecode file, then its
2476 // and error, generate that now.
2477 } else if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) {
2478 error("Invalid bytecode signature: " + utohexstr(Sig));
2481 // Tell the handler we're starting a module
2482 if (Handler) Handler->handleModuleBegin(ModuleID);
2484 // Get the module block and size and verify. This is handled specially
2485 // because the module block/size is always written in long format. Other
2486 // blocks are written in short format so the read_block method is used.
2487 unsigned Type, Size;
2490 if (Type != BytecodeFormat::ModuleBlockID) {
2491 error("Expected Module Block! Type:" + utostr(Type) + ", Size:"
2495 // It looks like the darwin ranlib program is broken, and adds trailing
2496 // garbage to the end of some bytecode files. This hack allows the bc
2497 // reader to ignore trailing garbage on bytecode files.
2498 if (At + Size < MemEnd)
2499 MemEnd = BlockEnd = At+Size;
2501 if (At + Size != MemEnd)
2502 error("Invalid Top Level Block Length! Type:" + utostr(Type)
2503 + ", Size:" + utostr(Size));
2505 // Parse the module contents
2506 this->ParseModule();
2508 // Check for missing functions
2510 error("Function expected, but bytecode stream ended!");
2512 // Look for intrinsic functions to upgrade, upgrade them, and save the
2513 // mapping from old function to new for use later when instructions are
2515 for (Module::iterator FI = TheModule->begin(), FE = TheModule->end();
2517 if (Function* newF = UpgradeIntrinsicFunction(FI)) {
2518 upgradedFunctions.insert(std::make_pair(FI, newF));
2522 // Tell the handler we're done with the module
2524 Handler->handleModuleEnd(ModuleID);
2526 // Tell the handler we're finished the parse
2527 if (Handler) Handler->handleFinish();
2533 //===----------------------------------------------------------------------===//
2534 //=== Default Implementations of Handler Methods
2535 //===----------------------------------------------------------------------===//
2537 BytecodeHandler::~BytecodeHandler() {}