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 // Convert previous opcode values into the current value and/or construct
566 // the instruction. This function handles all *abnormal* cases for instruction
567 // generation based on obsolete opcode values. The normal cases are handled
568 // in ParseInstruction below. Generally this function just produces a new
569 // Opcode value (first argument). In a few cases (VAArg, VANext) the upgrade
570 // path requies that the instruction (sequence) be generated differently from
571 // the normal case in order to preserve the original semantics. In these
572 // cases the result of the function will be a non-zero Instruction pointer. In
573 // all other cases, zero will be returned indicating that the *normal*
574 // instruction generation should be used, but with the new Opcode value.
577 BytecodeReader::handleObsoleteOpcodes(
578 unsigned &Opcode, ///< The old opcode, possibly updated by this function
579 std::vector<unsigned> &Oprnds, ///< The operands to the instruction
580 unsigned &iType, ///< The type code from the bytecode file
581 const Type* InstTy, ///< The type of the instruction
582 BasicBlock* BB ///< The basic block to insert into, if we need to
585 // First, short circuit this if no conversion is required. When signless
586 // instructions were implemented the entire opcode sequence was revised so
587 // we key on this first which means that the opcode value read is the one
589 if (!hasSignlessInstructions)
590 return 0; // The opcode is fine the way it is.
592 // Declare the resulting instruction we might build. In general we just
593 // change the Opcode argument but in a few cases we need to generate the
594 // Instruction here because the upgrade case is significantly different from
596 Instruction *Result = 0;
598 // If this is a bytecode format that did not include the unreachable
599 // instruction, bump up the opcode number to adjust it.
600 if (hasNoUnreachableInst)
601 if (Opcode >= Instruction::Unreachable && Opcode < 62)
604 // We're dealing with an upgrade situation. For each of the opcode values,
605 // perform the necessary conversion.
608 // This switch statement provides cases for all known opcodes prior to
609 // version 6 bytecode format. We know we're in an upgrade situation so
610 // if there isn't a match in this switch, then something is horribly
612 error("Unknown obsolete opcode encountered.");
615 Opcode = Instruction::Ret;
618 Opcode = Instruction::Br;
621 Opcode = Instruction::Switch;
624 Opcode = Instruction::Invoke;
627 Opcode = Instruction::Unwind;
629 case 6: // Unreachable
630 Opcode = Instruction::Unreachable;
633 Opcode = Instruction::Add;
636 Opcode = Instruction::Sub;
639 Opcode = Instruction::Mul;
642 // The type of the instruction is based on the operands. We need to select
643 // fdiv, udiv or sdiv based on that type. The iType values are hardcoded
644 // to the values used in bytecode version 5 (and prior) because it is
645 // likely these codes will change in future versions of LLVM.
646 if (iType == 10 || iType == 11 )
647 Opcode = Instruction::FDiv;
648 else if (iType >= 2 && iType <= 9 && iType % 2 != 0)
649 Opcode = Instruction::SDiv;
651 Opcode = Instruction::UDiv;
655 // As with "Div", make the signed/unsigned or floating point Rem
656 // instruction choice based on the type of the operands.
657 if (iType == 10 || iType == 11)
658 Opcode = Instruction::FRem;
659 else if (iType >= 2 && iType <= 9 && iType % 2 != 0)
660 Opcode = Instruction::SRem;
662 Opcode = Instruction::URem;
665 Opcode = Instruction::And;
668 Opcode = Instruction::Or;
671 Opcode = Instruction::Xor;
674 Opcode = Instruction::SetEQ;
677 Opcode = Instruction::SetNE;
680 Opcode = Instruction::SetLE;
683 Opcode = Instruction::SetGE;
686 Opcode = Instruction::SetLT;
689 Opcode = Instruction::SetGT;
692 Opcode = Instruction::Malloc;
695 Opcode = Instruction::Free;
698 Opcode = Instruction::Alloca;
701 Opcode = Instruction::Load;
704 Opcode = Instruction::Store;
706 case 26: // GetElementPtr
707 Opcode = Instruction::GetElementPtr;
710 Opcode = Instruction::PHI;
713 Opcode = Instruction::Cast;
716 Opcode = Instruction::Call;
719 Opcode = Instruction::Shl;
722 Opcode = Instruction::Shr;
724 case 32: { //VANext_old ( <= llvm 1.5 )
725 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
726 Function* NF = TheModule->getOrInsertFunction(
727 "llvm.va_copy", ArgTy, ArgTy, (Type *)0);
729 // In llvm 1.6 the VANext instruction was dropped because it was only
730 // necessary to have a VAArg instruction. The code below transforms an
731 // old vanext instruction into the equivalent code given only the
732 // availability of the new vaarg instruction. Essentially, the transform
734 // b = vanext a, t ->
735 // foo = alloca 1 of t
738 // tmp = vaarg foo, t
740 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vanext.fix");
741 BB->getInstList().push_back(foo);
742 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
743 BB->getInstList().push_back(bar);
744 BB->getInstList().push_back(new StoreInst(bar, foo));
745 Instruction* tmp = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
746 BB->getInstList().push_back(tmp);
747 Result = new LoadInst(foo);
750 case 33: { //VAArg_old
751 const Type* ArgTy = getValue(iType, Oprnds[0])->getType();
752 Function* NF = TheModule->getOrInsertFunction(
753 "llvm.va_copy", ArgTy, ArgTy, (Type *)0);
755 // In llvm 1.6 the VAArg's instruction semantics were changed. The code
756 // below transforms an old vaarg instruction into the equivalent code
757 // given only the availability of the new vaarg instruction. Essentially,
758 // the transform is as follows:
760 // foo = alloca 1 of t
764 AllocaInst* foo = new AllocaInst(ArgTy, 0, "vaarg.fix");
765 BB->getInstList().push_back(foo);
766 CallInst* bar = new CallInst(NF, getValue(iType, Oprnds[0]));
767 BB->getInstList().push_back(bar);
768 BB->getInstList().push_back(new StoreInst(bar, foo));
769 Result = new VAArgInst(foo, getSanitizedType(Oprnds[1]));
773 Opcode = Instruction::Select;
776 Opcode = Instruction::UserOp1;
779 Opcode = Instruction::UserOp2;
782 Opcode = Instruction::VAArg;
784 case 38: // ExtractElement
785 Opcode = Instruction::ExtractElement;
787 case 39: // InsertElement
788 Opcode = Instruction::InsertElement;
790 case 40: // ShuffleVector
791 Opcode = Instruction::ShuffleVector;
793 case 56: // Invoke with encoded CC
794 case 57: // Invoke Fast CC
795 case 58: // Call with extra operand for calling conv
796 case 59: // tail call, Fast CC
797 case 60: // normal call, Fast CC
798 case 61: // tail call, C Calling Conv
799 case 62: // volatile load
800 case 63: // volatile store
801 // In all these cases, we pass the opcode through. The new version uses
802 // the same code (for now, this might change in 2.0). These are listed
803 // here to document the opcodes in use in vers 5 bytecode and to make it
804 // easier to migrate these opcodes in the future.
810 //===----------------------------------------------------------------------===//
811 // Bytecode Parsing Methods
812 //===----------------------------------------------------------------------===//
814 /// This method parses a single instruction. The instruction is
815 /// inserted at the end of the \p BB provided. The arguments of
816 /// the instruction are provided in the \p Oprnds vector.
817 void BytecodeReader::ParseInstruction(std::vector<unsigned> &Oprnds,
821 // Clear instruction data
825 unsigned Op = read_uint();
827 // bits Instruction format: Common to all formats
828 // --------------------------
829 // 01-00: Opcode type, fixed to 1.
831 Opcode = (Op >> 2) & 63;
832 Oprnds.resize((Op >> 0) & 03);
834 // Extract the operands
835 switch (Oprnds.size()) {
837 // bits Instruction format:
838 // --------------------------
839 // 19-08: Resulting type plane
840 // 31-20: Operand #1 (if set to (2^12-1), then zero operands)
842 iType = (Op >> 8) & 4095;
843 Oprnds[0] = (Op >> 20) & 4095;
844 if (Oprnds[0] == 4095) // Handle special encoding for 0 operands...
848 // bits Instruction format:
849 // --------------------------
850 // 15-08: Resulting type plane
854 iType = (Op >> 8) & 255;
855 Oprnds[0] = (Op >> 16) & 255;
856 Oprnds[1] = (Op >> 24) & 255;
859 // bits Instruction format:
860 // --------------------------
861 // 13-08: Resulting type plane
866 iType = (Op >> 8) & 63;
867 Oprnds[0] = (Op >> 14) & 63;
868 Oprnds[1] = (Op >> 20) & 63;
869 Oprnds[2] = (Op >> 26) & 63;
872 At -= 4; // Hrm, try this again...
873 Opcode = read_vbr_uint();
875 iType = read_vbr_uint();
877 unsigned NumOprnds = read_vbr_uint();
878 Oprnds.resize(NumOprnds);
881 error("Zero-argument instruction found; this is invalid.");
883 for (unsigned i = 0; i != NumOprnds; ++i)
884 Oprnds[i] = read_vbr_uint();
889 const Type *InstTy = getSanitizedType(iType);
891 // Make the necessary adjustments for dealing with backwards compatibility
893 Instruction* Result =
894 handleObsoleteOpcodes(Opcode, Oprnds, iType, InstTy, BB);
896 // We have enough info to inform the handler now.
898 Handler->handleInstruction(Opcode, InstTy, Oprnds, At-SaveAt);
900 // If the backwards compatibility code didn't produce an instruction then
901 // we do the *normal* thing ..
903 // First, handle the easy binary operators case
904 if (Opcode >= Instruction::BinaryOpsBegin &&
905 Opcode < Instruction::BinaryOpsEnd && Oprnds.size() == 2)
906 Result = BinaryOperator::create(Instruction::BinaryOps(Opcode),
907 getValue(iType, Oprnds[0]),
908 getValue(iType, Oprnds[1]));
910 // Indicate that we don't think this is a call instruction (yet).
911 // Process based on the Opcode read
913 default: // There was an error, this shouldn't happen.
915 error("Illegal instruction read!");
917 case Instruction::VAArg:
918 if (Oprnds.size() != 2)
919 error("Invalid VAArg instruction!");
920 Result = new VAArgInst(getValue(iType, Oprnds[0]),
921 getSanitizedType(Oprnds[1]));
923 case Instruction::ExtractElement: {
924 if (Oprnds.size() != 2)
925 error("Invalid extractelement instruction!");
926 Value *V1 = getValue(iType, Oprnds[0]);
927 Value *V2 = getValue(Type::UIntTyID, Oprnds[1]);
929 if (!ExtractElementInst::isValidOperands(V1, V2))
930 error("Invalid extractelement instruction!");
932 Result = new ExtractElementInst(V1, V2);
935 case Instruction::InsertElement: {
936 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy);
937 if (!PackedTy || Oprnds.size() != 3)
938 error("Invalid insertelement instruction!");
940 Value *V1 = getValue(iType, Oprnds[0]);
941 Value *V2 = getValue(getTypeSlot(PackedTy->getElementType()),Oprnds[1]);
942 Value *V3 = getValue(Type::UIntTyID, Oprnds[2]);
944 if (!InsertElementInst::isValidOperands(V1, V2, V3))
945 error("Invalid insertelement instruction!");
946 Result = new InsertElementInst(V1, V2, V3);
949 case Instruction::ShuffleVector: {
950 const PackedType *PackedTy = dyn_cast<PackedType>(InstTy);
951 if (!PackedTy || Oprnds.size() != 3)
952 error("Invalid shufflevector instruction!");
953 Value *V1 = getValue(iType, Oprnds[0]);
954 Value *V2 = getValue(iType, Oprnds[1]);
955 const PackedType *EltTy =
956 PackedType::get(Type::UIntTy, PackedTy->getNumElements());
957 Value *V3 = getValue(getTypeSlot(EltTy), Oprnds[2]);
958 if (!ShuffleVectorInst::isValidOperands(V1, V2, V3))
959 error("Invalid shufflevector instruction!");
960 Result = new ShuffleVectorInst(V1, V2, V3);
963 case Instruction::Cast:
964 if (Oprnds.size() != 2)
965 error("Invalid Cast instruction!");
966 Result = new CastInst(getValue(iType, Oprnds[0]),
967 getSanitizedType(Oprnds[1]));
969 case Instruction::Select:
970 if (Oprnds.size() != 3)
971 error("Invalid Select instruction!");
972 Result = new SelectInst(getValue(Type::BoolTyID, Oprnds[0]),
973 getValue(iType, Oprnds[1]),
974 getValue(iType, Oprnds[2]));
976 case Instruction::PHI: {
977 if (Oprnds.size() == 0 || (Oprnds.size() & 1))
978 error("Invalid phi node encountered!");
980 PHINode *PN = new PHINode(InstTy);
981 PN->reserveOperandSpace(Oprnds.size());
982 for (unsigned i = 0, e = Oprnds.size(); i != e; i += 2)
984 getValue(iType, Oprnds[i]), getBasicBlock(Oprnds[i+1]));
989 case Instruction::Shl:
990 case Instruction::Shr:
991 Result = new ShiftInst(Instruction::OtherOps(Opcode),
992 getValue(iType, Oprnds[0]),
993 getValue(Type::UByteTyID, Oprnds[1]));
995 case Instruction::Ret:
996 if (Oprnds.size() == 0)
997 Result = new ReturnInst();
998 else if (Oprnds.size() == 1)
999 Result = new ReturnInst(getValue(iType, Oprnds[0]));
1001 error("Unrecognized instruction!");
1004 case Instruction::Br:
1005 if (Oprnds.size() == 1)
1006 Result = new BranchInst(getBasicBlock(Oprnds[0]));
1007 else if (Oprnds.size() == 3)
1008 Result = new BranchInst(getBasicBlock(Oprnds[0]),
1009 getBasicBlock(Oprnds[1]), getValue(Type::BoolTyID , Oprnds[2]));
1011 error("Invalid number of operands for a 'br' instruction!");
1013 case Instruction::Switch: {
1014 if (Oprnds.size() & 1)
1015 error("Switch statement with odd number of arguments!");
1017 SwitchInst *I = new SwitchInst(getValue(iType, Oprnds[0]),
1018 getBasicBlock(Oprnds[1]),
1020 for (unsigned i = 2, e = Oprnds.size(); i != e; i += 2)
1021 I->addCase(cast<ConstantInt>(getValue(iType, Oprnds[i])),
1022 getBasicBlock(Oprnds[i+1]));
1026 case 58: // Call with extra operand for calling conv
1027 case 59: // tail call, Fast CC
1028 case 60: // normal call, Fast CC
1029 case 61: // tail call, C Calling Conv
1030 case Instruction::Call: { // Normal Call, C Calling Convention
1031 if (Oprnds.size() == 0)
1032 error("Invalid call instruction encountered!");
1034 Value *F = getValue(iType, Oprnds[0]);
1036 unsigned CallingConv = CallingConv::C;
1037 bool isTailCall = false;
1039 if (Opcode == 61 || Opcode == 59)
1043 isTailCall = Oprnds.back() & 1;
1044 CallingConv = Oprnds.back() >> 1;
1046 } else if (Opcode == 59 || Opcode == 60) {
1047 CallingConv = CallingConv::Fast;
1050 // Check to make sure we have a pointer to function type
1051 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
1052 if (PTy == 0) error("Call to non function pointer value!");
1053 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
1054 if (FTy == 0) error("Call to non function pointer value!");
1056 std::vector<Value *> Params;
1057 if (!FTy->isVarArg()) {
1058 FunctionType::param_iterator It = FTy->param_begin();
1060 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
1061 if (It == FTy->param_end())
1062 error("Invalid call instruction!");
1063 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
1065 if (It != FTy->param_end())
1066 error("Invalid call instruction!");
1068 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
1070 unsigned FirstVariableOperand;
1071 if (Oprnds.size() < FTy->getNumParams())
1072 error("Call instruction missing operands!");
1074 // Read all of the fixed arguments
1075 for (unsigned i = 0, e = FTy->getNumParams(); i != e; ++i)
1077 getValue(getTypeSlot(FTy->getParamType(i)),Oprnds[i]));
1079 FirstVariableOperand = FTy->getNumParams();
1081 if ((Oprnds.size()-FirstVariableOperand) & 1)
1082 error("Invalid call instruction!"); // Must be pairs of type/value
1084 for (unsigned i = FirstVariableOperand, e = Oprnds.size();
1086 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
1089 Result = new CallInst(F, Params);
1090 if (isTailCall) cast<CallInst>(Result)->setTailCall();
1091 if (CallingConv) cast<CallInst>(Result)->setCallingConv(CallingConv);
1094 case 56: // Invoke with encoded CC
1095 case 57: // Invoke Fast CC
1096 case Instruction::Invoke: { // Invoke C CC
1097 if (Oprnds.size() < 3)
1098 error("Invalid invoke instruction!");
1099 Value *F = getValue(iType, Oprnds[0]);
1101 // Check to make sure we have a pointer to function type
1102 const PointerType *PTy = dyn_cast<PointerType>(F->getType());
1104 error("Invoke to non function pointer value!");
1105 const FunctionType *FTy = dyn_cast<FunctionType>(PTy->getElementType());
1107 error("Invoke to non function pointer value!");
1109 std::vector<Value *> Params;
1110 BasicBlock *Normal, *Except;
1111 unsigned CallingConv = CallingConv::C;
1114 CallingConv = CallingConv::Fast;
1115 else if (Opcode == 56) {
1116 CallingConv = Oprnds.back();
1120 if (!FTy->isVarArg()) {
1121 Normal = getBasicBlock(Oprnds[1]);
1122 Except = getBasicBlock(Oprnds[2]);
1124 FunctionType::param_iterator It = FTy->param_begin();
1125 for (unsigned i = 3, e = Oprnds.size(); i != e; ++i) {
1126 if (It == FTy->param_end())
1127 error("Invalid invoke instruction!");
1128 Params.push_back(getValue(getTypeSlot(*It++), Oprnds[i]));
1130 if (It != FTy->param_end())
1131 error("Invalid invoke instruction!");
1133 Oprnds.erase(Oprnds.begin(), Oprnds.begin()+1);
1135 Normal = getBasicBlock(Oprnds[0]);
1136 Except = getBasicBlock(Oprnds[1]);
1138 unsigned FirstVariableArgument = FTy->getNumParams()+2;
1139 for (unsigned i = 2; i != FirstVariableArgument; ++i)
1140 Params.push_back(getValue(getTypeSlot(FTy->getParamType(i-2)),
1143 // Must be type/value pairs. If not, error out.
1144 if (Oprnds.size()-FirstVariableArgument & 1)
1145 error("Invalid invoke instruction!");
1147 for (unsigned i = FirstVariableArgument; i < Oprnds.size(); i += 2)
1148 Params.push_back(getValue(Oprnds[i], Oprnds[i+1]));
1151 Result = new InvokeInst(F, Normal, Except, Params);
1152 if (CallingConv) cast<InvokeInst>(Result)->setCallingConv(CallingConv);
1155 case Instruction::Malloc: {
1157 if (Oprnds.size() == 2)
1158 Align = (1 << Oprnds[1]) >> 1;
1159 else if (Oprnds.size() > 2)
1160 error("Invalid malloc instruction!");
1161 if (!isa<PointerType>(InstTy))
1162 error("Invalid malloc instruction!");
1164 Result = new MallocInst(cast<PointerType>(InstTy)->getElementType(),
1165 getValue(Type::UIntTyID, Oprnds[0]), Align);
1168 case Instruction::Alloca: {
1170 if (Oprnds.size() == 2)
1171 Align = (1 << Oprnds[1]) >> 1;
1172 else if (Oprnds.size() > 2)
1173 error("Invalid alloca instruction!");
1174 if (!isa<PointerType>(InstTy))
1175 error("Invalid alloca instruction!");
1177 Result = new AllocaInst(cast<PointerType>(InstTy)->getElementType(),
1178 getValue(Type::UIntTyID, Oprnds[0]), Align);
1181 case Instruction::Free:
1182 if (!isa<PointerType>(InstTy))
1183 error("Invalid free instruction!");
1184 Result = new FreeInst(getValue(iType, Oprnds[0]));
1186 case Instruction::GetElementPtr: {
1187 if (Oprnds.size() == 0 || !isa<PointerType>(InstTy))
1188 error("Invalid getelementptr instruction!");
1190 std::vector<Value*> Idx;
1192 const Type *NextTy = InstTy;
1193 for (unsigned i = 1, e = Oprnds.size(); i != e; ++i) {
1194 const CompositeType *TopTy = dyn_cast_or_null<CompositeType>(NextTy);
1196 error("Invalid getelementptr instruction!");
1198 unsigned ValIdx = Oprnds[i];
1200 if (!hasRestrictedGEPTypes) {
1201 // Struct indices are always uints, sequential type indices can be
1202 // any of the 32 or 64-bit integer types. The actual choice of
1203 // type is encoded in the low two bits of the slot number.
1204 if (isa<StructType>(TopTy))
1205 IdxTy = Type::UIntTyID;
1207 switch (ValIdx & 3) {
1209 case 0: IdxTy = Type::UIntTyID; break;
1210 case 1: IdxTy = Type::IntTyID; break;
1211 case 2: IdxTy = Type::ULongTyID; break;
1212 case 3: IdxTy = Type::LongTyID; break;
1217 IdxTy = isa<StructType>(TopTy) ? Type::UByteTyID : Type::LongTyID;
1220 Idx.push_back(getValue(IdxTy, ValIdx));
1222 // Convert ubyte struct indices into uint struct indices.
1223 if (isa<StructType>(TopTy) && hasRestrictedGEPTypes)
1224 if (ConstantInt *C = dyn_cast<ConstantInt>(Idx.back()))
1225 if (C->getType() == Type::UByteTy)
1226 Idx[Idx.size()-1] = ConstantExpr::getCast(C, Type::UIntTy);
1228 NextTy = GetElementPtrInst::getIndexedType(InstTy, Idx, true);
1231 Result = new GetElementPtrInst(getValue(iType, Oprnds[0]), Idx);
1234 case 62: // volatile load
1235 case Instruction::Load:
1236 if (Oprnds.size() != 1 || !isa<PointerType>(InstTy))
1237 error("Invalid load instruction!");
1238 Result = new LoadInst(getValue(iType, Oprnds[0]), "", Opcode == 62);
1240 case 63: // volatile store
1241 case Instruction::Store: {
1242 if (!isa<PointerType>(InstTy) || Oprnds.size() != 2)
1243 error("Invalid store instruction!");
1245 Value *Ptr = getValue(iType, Oprnds[1]);
1246 const Type *ValTy = cast<PointerType>(Ptr->getType())->getElementType();
1247 Result = new StoreInst(getValue(getTypeSlot(ValTy), Oprnds[0]), Ptr,
1251 case Instruction::Unwind:
1252 if (Oprnds.size() != 0) error("Invalid unwind instruction!");
1253 Result = new UnwindInst();
1255 case Instruction::Unreachable:
1256 if (Oprnds.size() != 0) error("Invalid unreachable instruction!");
1257 Result = new UnreachableInst();
1259 } // end switch(Opcode)
1260 } // end if *normal*
1262 BB->getInstList().push_back(Result);
1265 if (Result->getType() == InstTy)
1268 TypeSlot = getTypeSlot(Result->getType());
1270 insertValue(Result, TypeSlot, FunctionValues);
1273 /// Get a particular numbered basic block, which might be a forward reference.
1274 /// This works together with ParseBasicBlock to handle these forward references
1275 /// in a clean manner. This function is used when constructing phi, br, switch,
1276 /// and other instructions that reference basic blocks. Blocks are numbered
1277 /// sequentially as they appear in the function.
1278 BasicBlock *BytecodeReader::getBasicBlock(unsigned ID) {
1279 // Make sure there is room in the table...
1280 if (ParsedBasicBlocks.size() <= ID) ParsedBasicBlocks.resize(ID+1);
1282 // First check to see if this is a backwards reference, i.e., ParseBasicBlock
1283 // has already created this block, or if the forward reference has already
1285 if (ParsedBasicBlocks[ID])
1286 return ParsedBasicBlocks[ID];
1288 // Otherwise, the basic block has not yet been created. Do so and add it to
1289 // the ParsedBasicBlocks list.
1290 return ParsedBasicBlocks[ID] = new BasicBlock();
1293 /// In LLVM 1.0 bytecode files, we used to output one basicblock at a time.
1294 /// This method reads in one of the basicblock packets. This method is not used
1295 /// for bytecode files after LLVM 1.0
1296 /// @returns The basic block constructed.
1297 BasicBlock *BytecodeReader::ParseBasicBlock(unsigned BlockNo) {
1298 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1302 if (ParsedBasicBlocks.size() == BlockNo)
1303 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1304 else if (ParsedBasicBlocks[BlockNo] == 0)
1305 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1307 BB = ParsedBasicBlocks[BlockNo];
1309 std::vector<unsigned> Operands;
1310 while (moreInBlock())
1311 ParseInstruction(Operands, BB);
1313 if (Handler) Handler->handleBasicBlockEnd(BlockNo);
1317 /// Parse all of the BasicBlock's & Instruction's in the body of a function.
1318 /// In post 1.0 bytecode files, we no longer emit basic block individually,
1319 /// in order to avoid per-basic-block overhead.
1320 /// @returns Rhe number of basic blocks encountered.
1321 unsigned BytecodeReader::ParseInstructionList(Function* F) {
1322 unsigned BlockNo = 0;
1323 std::vector<unsigned> Args;
1325 while (moreInBlock()) {
1326 if (Handler) Handler->handleBasicBlockBegin(BlockNo);
1328 if (ParsedBasicBlocks.size() == BlockNo)
1329 ParsedBasicBlocks.push_back(BB = new BasicBlock());
1330 else if (ParsedBasicBlocks[BlockNo] == 0)
1331 BB = ParsedBasicBlocks[BlockNo] = new BasicBlock();
1333 BB = ParsedBasicBlocks[BlockNo];
1335 F->getBasicBlockList().push_back(BB);
1337 // Read instructions into this basic block until we get to a terminator
1338 while (moreInBlock() && !BB->getTerminator())
1339 ParseInstruction(Args, BB);
1341 if (!BB->getTerminator())
1342 error("Non-terminated basic block found!");
1344 if (Handler) Handler->handleBasicBlockEnd(BlockNo-1);
1350 /// Parse a symbol table. This works for both module level and function
1351 /// level symbol tables. For function level symbol tables, the CurrentFunction
1352 /// parameter must be non-zero and the ST parameter must correspond to
1353 /// CurrentFunction's symbol table. For Module level symbol tables, the
1354 /// CurrentFunction argument must be zero.
1355 void BytecodeReader::ParseSymbolTable(Function *CurrentFunction,
1357 if (Handler) Handler->handleSymbolTableBegin(CurrentFunction,ST);
1359 // Allow efficient basic block lookup by number.
1360 std::vector<BasicBlock*> BBMap;
1361 if (CurrentFunction)
1362 for (Function::iterator I = CurrentFunction->begin(),
1363 E = CurrentFunction->end(); I != E; ++I)
1366 /// In LLVM 1.3 we write types separately from values so
1367 /// The types are always first in the symbol table. This is
1368 /// because Type no longer derives from Value.
1369 if (!hasTypeDerivedFromValue) {
1370 // Symtab block header: [num entries]
1371 unsigned NumEntries = read_vbr_uint();
1372 for (unsigned i = 0; i < NumEntries; ++i) {
1373 // Symtab entry: [def slot #][name]
1374 unsigned slot = read_vbr_uint();
1375 std::string Name = read_str();
1376 const Type* T = getType(slot);
1377 ST->insert(Name, T);
1381 while (moreInBlock()) {
1382 // Symtab block header: [num entries][type id number]
1383 unsigned NumEntries = read_vbr_uint();
1385 bool isTypeType = read_typeid(Typ);
1386 const Type *Ty = getType(Typ);
1388 for (unsigned i = 0; i != NumEntries; ++i) {
1389 // Symtab entry: [def slot #][name]
1390 unsigned slot = read_vbr_uint();
1391 std::string Name = read_str();
1393 // if we're reading a pre 1.3 bytecode file and the type plane
1394 // is the "type type", handle it here
1396 const Type* T = getType(slot);
1398 error("Failed type look-up for name '" + Name + "'");
1399 ST->insert(Name, T);
1400 continue; // code below must be short circuited
1403 if (Typ == Type::LabelTyID) {
1404 if (slot < BBMap.size())
1407 V = getValue(Typ, slot, false); // Find mapping...
1410 error("Failed value look-up for name '" + Name + "'");
1415 checkPastBlockEnd("Symbol Table");
1416 if (Handler) Handler->handleSymbolTableEnd();
1419 /// Read in the types portion of a compaction table.
1420 void BytecodeReader::ParseCompactionTypes(unsigned NumEntries) {
1421 for (unsigned i = 0; i != NumEntries; ++i) {
1422 unsigned TypeSlot = 0;
1423 if (read_typeid(TypeSlot))
1424 error("Invalid type in compaction table: type type");
1425 const Type *Typ = getGlobalTableType(TypeSlot);
1426 CompactionTypes.push_back(std::make_pair(Typ, TypeSlot));
1427 if (Handler) Handler->handleCompactionTableType(i, TypeSlot, Typ);
1431 /// Parse a compaction table.
1432 void BytecodeReader::ParseCompactionTable() {
1434 // Notify handler that we're beginning a compaction table.
1435 if (Handler) Handler->handleCompactionTableBegin();
1437 // In LLVM 1.3 Type no longer derives from Value. So,
1438 // we always write them first in the compaction table
1439 // because they can't occupy a "type plane" where the
1441 if (! hasTypeDerivedFromValue) {
1442 unsigned NumEntries = read_vbr_uint();
1443 ParseCompactionTypes(NumEntries);
1446 // Compaction tables live in separate blocks so we have to loop
1447 // until we've read the whole thing.
1448 while (moreInBlock()) {
1449 // Read the number of Value* entries in the compaction table
1450 unsigned NumEntries = read_vbr_uint();
1452 unsigned isTypeType = false;
1454 // Decode the type from value read in. Most compaction table
1455 // planes will have one or two entries in them. If that's the
1456 // case then the length is encoded in the bottom two bits and
1457 // the higher bits encode the type. This saves another VBR value.
1458 if ((NumEntries & 3) == 3) {
1459 // In this case, both low-order bits are set (value 3). This
1460 // is a signal that the typeid follows.
1462 isTypeType = read_typeid(Ty);
1464 // In this case, the low-order bits specify the number of entries
1465 // and the high order bits specify the type.
1466 Ty = NumEntries >> 2;
1467 isTypeType = sanitizeTypeId(Ty);
1471 // if we're reading a pre 1.3 bytecode file and the type plane
1472 // is the "type type", handle it here
1474 ParseCompactionTypes(NumEntries);
1476 // Make sure we have enough room for the plane.
1477 if (Ty >= CompactionValues.size())
1478 CompactionValues.resize(Ty+1);
1480 // Make sure the plane is empty or we have some kind of error.
1481 if (!CompactionValues[Ty].empty())
1482 error("Compaction table plane contains multiple entries!");
1484 // Notify handler about the plane.
1485 if (Handler) Handler->handleCompactionTablePlane(Ty, NumEntries);
1487 // Push the implicit zero.
1488 CompactionValues[Ty].push_back(Constant::getNullValue(getType(Ty)));
1490 // Read in each of the entries, put them in the compaction table
1491 // and notify the handler that we have a new compaction table value.
1492 for (unsigned i = 0; i != NumEntries; ++i) {
1493 unsigned ValSlot = read_vbr_uint();
1494 Value *V = getGlobalTableValue(Ty, ValSlot);
1495 CompactionValues[Ty].push_back(V);
1496 if (Handler) Handler->handleCompactionTableValue(i, Ty, ValSlot);
1500 // Notify handler that the compaction table is done.
1501 if (Handler) Handler->handleCompactionTableEnd();
1504 // Parse a single type. The typeid is read in first. If its a primitive type
1505 // then nothing else needs to be read, we know how to instantiate it. If its
1506 // a derived type, then additional data is read to fill out the type
1508 const Type *BytecodeReader::ParseType() {
1509 unsigned PrimType = 0;
1510 if (read_typeid(PrimType))
1511 error("Invalid type (type type) in type constants!");
1513 const Type *Result = 0;
1514 if ((Result = Type::getPrimitiveType((Type::TypeID)PrimType)))
1518 case Type::FunctionTyID: {
1519 const Type *RetType = readSanitizedType();
1521 unsigned NumParams = read_vbr_uint();
1523 std::vector<const Type*> Params;
1525 Params.push_back(readSanitizedType());
1527 bool isVarArg = Params.size() && Params.back() == Type::VoidTy;
1528 if (isVarArg) Params.pop_back();
1530 Result = FunctionType::get(RetType, Params, isVarArg);
1533 case Type::ArrayTyID: {
1534 const Type *ElementType = readSanitizedType();
1535 unsigned NumElements = read_vbr_uint();
1536 Result = ArrayType::get(ElementType, NumElements);
1539 case Type::PackedTyID: {
1540 const Type *ElementType = readSanitizedType();
1541 unsigned NumElements = read_vbr_uint();
1542 Result = PackedType::get(ElementType, NumElements);
1545 case Type::StructTyID: {
1546 std::vector<const Type*> Elements;
1548 if (read_typeid(Typ))
1549 error("Invalid element type (type type) for structure!");
1551 while (Typ) { // List is terminated by void/0 typeid
1552 Elements.push_back(getType(Typ));
1553 if (read_typeid(Typ))
1554 error("Invalid element type (type type) for structure!");
1557 Result = StructType::get(Elements);
1560 case Type::PointerTyID: {
1561 Result = PointerType::get(readSanitizedType());
1565 case Type::OpaqueTyID: {
1566 Result = OpaqueType::get();
1571 error("Don't know how to deserialize primitive type " + utostr(PrimType));
1574 if (Handler) Handler->handleType(Result);
1578 // ParseTypes - We have to use this weird code to handle recursive
1579 // types. We know that recursive types will only reference the current slab of
1580 // values in the type plane, but they can forward reference types before they
1581 // have been read. For example, Type #0 might be '{ Ty#1 }' and Type #1 might
1582 // be 'Ty#0*'. When reading Type #0, type number one doesn't exist. To fix
1583 // this ugly problem, we pessimistically insert an opaque type for each type we
1584 // are about to read. This means that forward references will resolve to
1585 // something and when we reread the type later, we can replace the opaque type
1586 // with a new resolved concrete type.
1588 void BytecodeReader::ParseTypes(TypeListTy &Tab, unsigned NumEntries){
1589 assert(Tab.size() == 0 && "should not have read type constants in before!");
1591 // Insert a bunch of opaque types to be resolved later...
1592 Tab.reserve(NumEntries);
1593 for (unsigned i = 0; i != NumEntries; ++i)
1594 Tab.push_back(OpaqueType::get());
1597 Handler->handleTypeList(NumEntries);
1599 // If we are about to resolve types, make sure the type cache is clear.
1601 ModuleTypeIDCache.clear();
1603 // Loop through reading all of the types. Forward types will make use of the
1604 // opaque types just inserted.
1606 for (unsigned i = 0; i != NumEntries; ++i) {
1607 const Type* NewTy = ParseType();
1608 const Type* OldTy = Tab[i].get();
1610 error("Couldn't parse type!");
1612 // Don't directly push the new type on the Tab. Instead we want to replace
1613 // the opaque type we previously inserted with the new concrete value. This
1614 // approach helps with forward references to types. The refinement from the
1615 // abstract (opaque) type to the new type causes all uses of the abstract
1616 // type to use the concrete type (NewTy). This will also cause the opaque
1617 // type to be deleted.
1618 cast<DerivedType>(const_cast<Type*>(OldTy))->refineAbstractTypeTo(NewTy);
1620 // This should have replaced the old opaque type with the new type in the
1621 // value table... or with a preexisting type that was already in the system.
1622 // Let's just make sure it did.
1623 assert(Tab[i] != OldTy && "refineAbstractType didn't work!");
1627 // Upgrade obsolete constant expression opcodes (ver. 5 and prior) to the new
1628 // values used after ver 6. bytecode format. The operands are provided to the
1629 // function so that decisions based on the operand type can be made when
1630 // auto-upgrading obsolete opcodes to the new ones.
1631 // NOTE: This code needs to be kept synchronized with handleObsoleteOpcodes.
1632 // We can't use that function because of that functions argument requirements.
1633 // This function only deals with the subset of opcodes that are applicable to
1634 // constant expressions and is therefore simpler than handleObsoleteOpcodes.
1635 inline unsigned fixCEOpcodes(
1636 unsigned Opcode, const std::vector<Constant*> &ArgVec
1639 default: // Pass Through
1640 // If we don't match any of the cases here then the opcode is fine the
1644 Opcode = Instruction::Add;
1647 Opcode = Instruction::Sub;
1650 Opcode = Instruction::Mul;
1653 // The type of the instruction is based on the operands. We need to select
1654 // either udiv or sdiv based on that type. This expression selects the
1655 // cases where the type is floating point or signed in which case we
1656 // generated an sdiv instruction.
1657 if (ArgVec[0]->getType()->isFloatingPoint())
1658 Opcode = Instruction::FDiv;
1659 else if (ArgVec[0]->getType()->isSigned())
1660 Opcode = Instruction::SDiv;
1662 Opcode = Instruction::UDiv;
1665 // As with "Div", make the signed/unsigned or floating point Rem
1666 // instruction choice based on the type of the operands.
1667 if (ArgVec[0]->getType()->isFloatingPoint())
1668 Opcode = Instruction::FRem;
1669 else if (ArgVec[0]->getType()->isSigned())
1670 Opcode = Instruction::SRem;
1672 Opcode = Instruction::URem;
1675 Opcode = Instruction::And;
1678 Opcode = Instruction::Or;
1681 Opcode = Instruction::Xor;
1684 Opcode = Instruction::SetEQ;
1687 Opcode = Instruction::SetNE;
1690 Opcode = Instruction::SetLE;
1693 Opcode = Instruction::SetGE;
1696 Opcode = Instruction::SetLT;
1699 Opcode = Instruction::SetGT;
1701 case 26: // GetElementPtr
1702 Opcode = Instruction::GetElementPtr;
1705 Opcode = Instruction::Cast;
1708 Opcode = Instruction::Shl;
1711 Opcode = Instruction::Shr;
1714 Opcode = Instruction::Select;
1716 case 38: // ExtractElement
1717 Opcode = Instruction::ExtractElement;
1719 case 39: // InsertElement
1720 Opcode = Instruction::InsertElement;
1722 case 40: // ShuffleVector
1723 Opcode = Instruction::ShuffleVector;
1729 /// Parse a single constant value
1730 Value *BytecodeReader::ParseConstantPoolValue(unsigned TypeID) {
1731 // We must check for a ConstantExpr before switching by type because
1732 // a ConstantExpr can be of any type, and has no explicit value.
1734 // 0 if not expr; numArgs if is expr
1735 unsigned isExprNumArgs = read_vbr_uint();
1737 if (isExprNumArgs) {
1738 if (!hasNoUndefValue) {
1739 // 'undef' is encoded with 'exprnumargs' == 1.
1740 if (isExprNumArgs == 1)
1741 return UndefValue::get(getType(TypeID));
1743 // Inline asm is encoded with exprnumargs == ~0U.
1744 if (isExprNumArgs == ~0U) {
1745 std::string AsmStr = read_str();
1746 std::string ConstraintStr = read_str();
1747 unsigned Flags = read_vbr_uint();
1749 const PointerType *PTy = dyn_cast<PointerType>(getType(TypeID));
1750 const FunctionType *FTy =
1751 PTy ? dyn_cast<FunctionType>(PTy->getElementType()) : 0;
1753 if (!FTy || !InlineAsm::Verify(FTy, ConstraintStr))
1754 error("Invalid constraints for inline asm");
1756 error("Invalid flags for inline asm");
1757 bool HasSideEffects = Flags & 1;
1758 return InlineAsm::get(FTy, AsmStr, ConstraintStr, HasSideEffects);
1764 // FIXME: Encoding of constant exprs could be much more compact!
1765 std::vector<Constant*> ArgVec;
1766 ArgVec.reserve(isExprNumArgs);
1767 unsigned Opcode = read_vbr_uint();
1769 // Bytecode files before LLVM 1.4 need have a missing terminator inst.
1770 if (hasNoUnreachableInst) Opcode++;
1772 // Read the slot number and types of each of the arguments
1773 for (unsigned i = 0; i != isExprNumArgs; ++i) {
1774 unsigned ArgValSlot = read_vbr_uint();
1775 unsigned ArgTypeSlot = 0;
1776 if (read_typeid(ArgTypeSlot))
1777 error("Invalid argument type (type type) for constant value");
1779 // Get the arg value from its slot if it exists, otherwise a placeholder
1780 ArgVec.push_back(getConstantValue(ArgTypeSlot, ArgValSlot));
1783 // Handle backwards compatibility for the opcode numbers
1784 if (hasSignlessInstructions)
1785 Opcode = fixCEOpcodes(Opcode, ArgVec);
1787 // Construct a ConstantExpr of the appropriate kind
1788 if (isExprNumArgs == 1) { // All one-operand expressions
1789 if (Opcode != Instruction::Cast)
1790 error("Only cast instruction has one argument for ConstantExpr");
1792 Constant* Result = ConstantExpr::getCast(ArgVec[0], getType(TypeID));
1793 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1795 } else if (Opcode == Instruction::GetElementPtr) { // GetElementPtr
1796 std::vector<Constant*> IdxList(ArgVec.begin()+1, ArgVec.end());
1798 if (hasRestrictedGEPTypes) {
1799 const Type *BaseTy = ArgVec[0]->getType();
1800 generic_gep_type_iterator<std::vector<Constant*>::iterator>
1801 GTI = gep_type_begin(BaseTy, IdxList.begin(), IdxList.end()),
1802 E = gep_type_end(BaseTy, IdxList.begin(), IdxList.end());
1803 for (unsigned i = 0; GTI != E; ++GTI, ++i)
1804 if (isa<StructType>(*GTI)) {
1805 if (IdxList[i]->getType() != Type::UByteTy)
1806 error("Invalid index for getelementptr!");
1807 IdxList[i] = ConstantExpr::getCast(IdxList[i], Type::UIntTy);
1811 Constant* Result = ConstantExpr::getGetElementPtr(ArgVec[0], IdxList);
1812 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1814 } else if (Opcode == Instruction::Select) {
1815 if (ArgVec.size() != 3)
1816 error("Select instruction must have three arguments.");
1817 Constant* Result = ConstantExpr::getSelect(ArgVec[0], ArgVec[1],
1819 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1821 } else if (Opcode == Instruction::ExtractElement) {
1822 if (ArgVec.size() != 2 ||
1823 !ExtractElementInst::isValidOperands(ArgVec[0], ArgVec[1]))
1824 error("Invalid extractelement constand expr arguments");
1825 Constant* Result = ConstantExpr::getExtractElement(ArgVec[0], ArgVec[1]);
1826 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1828 } else if (Opcode == Instruction::InsertElement) {
1829 if (ArgVec.size() != 3 ||
1830 !InsertElementInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2]))
1831 error("Invalid insertelement constand expr arguments");
1834 ConstantExpr::getInsertElement(ArgVec[0], ArgVec[1], ArgVec[2]);
1835 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1837 } else if (Opcode == Instruction::ShuffleVector) {
1838 if (ArgVec.size() != 3 ||
1839 !ShuffleVectorInst::isValidOperands(ArgVec[0], ArgVec[1], ArgVec[2]))
1840 error("Invalid shufflevector constant expr arguments.");
1842 ConstantExpr::getShuffleVector(ArgVec[0], ArgVec[1], ArgVec[2]);
1843 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1845 } else { // All other 2-operand expressions
1846 Constant* Result = ConstantExpr::get(Opcode, ArgVec[0], ArgVec[1]);
1847 if (Handler) Handler->handleConstantExpression(Opcode, ArgVec, Result);
1852 // Ok, not an ConstantExpr. We now know how to read the given type...
1853 const Type *Ty = getType(TypeID);
1854 Constant *Result = 0;
1855 switch (Ty->getTypeID()) {
1856 case Type::BoolTyID: {
1857 unsigned Val = read_vbr_uint();
1858 if (Val != 0 && Val != 1)
1859 error("Invalid boolean value read.");
1860 Result = ConstantBool::get(Val == 1);
1861 if (Handler) Handler->handleConstantValue(Result);
1865 case Type::UByteTyID: // Unsigned integer types...
1866 case Type::UShortTyID:
1867 case Type::UIntTyID: {
1868 unsigned Val = read_vbr_uint();
1869 if (!ConstantInt::isValueValidForType(Ty, uint64_t(Val)))
1870 error("Invalid unsigned byte/short/int read.");
1871 Result = ConstantInt::get(Ty, Val);
1872 if (Handler) Handler->handleConstantValue(Result);
1876 case Type::ULongTyID:
1877 Result = ConstantInt::get(Ty, read_vbr_uint64());
1878 if (Handler) Handler->handleConstantValue(Result);
1881 case Type::SByteTyID: // Signed integer types...
1882 case Type::ShortTyID:
1884 case Type::LongTyID: {
1885 int64_t Val = read_vbr_int64();
1886 if (!ConstantInt::isValueValidForType(Ty, Val))
1887 error("Invalid signed byte/short/int/long read.");
1888 Result = ConstantInt::get(Ty, Val);
1889 if (Handler) Handler->handleConstantValue(Result);
1893 case Type::FloatTyID: {
1896 Result = ConstantFP::get(Ty, Val);
1897 if (Handler) Handler->handleConstantValue(Result);
1901 case Type::DoubleTyID: {
1904 Result = ConstantFP::get(Ty, Val);
1905 if (Handler) Handler->handleConstantValue(Result);
1909 case Type::ArrayTyID: {
1910 const ArrayType *AT = cast<ArrayType>(Ty);
1911 unsigned NumElements = AT->getNumElements();
1912 unsigned TypeSlot = getTypeSlot(AT->getElementType());
1913 std::vector<Constant*> Elements;
1914 Elements.reserve(NumElements);
1915 while (NumElements--) // Read all of the elements of the constant.
1916 Elements.push_back(getConstantValue(TypeSlot,
1918 Result = ConstantArray::get(AT, Elements);
1919 if (Handler) Handler->handleConstantArray(AT, Elements, TypeSlot, Result);
1923 case Type::StructTyID: {
1924 const StructType *ST = cast<StructType>(Ty);
1926 std::vector<Constant *> Elements;
1927 Elements.reserve(ST->getNumElements());
1928 for (unsigned i = 0; i != ST->getNumElements(); ++i)
1929 Elements.push_back(getConstantValue(ST->getElementType(i),
1932 Result = ConstantStruct::get(ST, Elements);
1933 if (Handler) Handler->handleConstantStruct(ST, Elements, Result);
1937 case Type::PackedTyID: {
1938 const PackedType *PT = cast<PackedType>(Ty);
1939 unsigned NumElements = PT->getNumElements();
1940 unsigned TypeSlot = getTypeSlot(PT->getElementType());
1941 std::vector<Constant*> Elements;
1942 Elements.reserve(NumElements);
1943 while (NumElements--) // Read all of the elements of the constant.
1944 Elements.push_back(getConstantValue(TypeSlot,
1946 Result = ConstantPacked::get(PT, Elements);
1947 if (Handler) Handler->handleConstantPacked(PT, Elements, TypeSlot, Result);
1951 case Type::PointerTyID: { // ConstantPointerRef value (backwards compat).
1952 const PointerType *PT = cast<PointerType>(Ty);
1953 unsigned Slot = read_vbr_uint();
1955 // Check to see if we have already read this global variable...
1956 Value *Val = getValue(TypeID, Slot, false);
1958 GlobalValue *GV = dyn_cast<GlobalValue>(Val);
1959 if (!GV) error("GlobalValue not in ValueTable!");
1960 if (Handler) Handler->handleConstantPointer(PT, Slot, GV);
1963 error("Forward references are not allowed here.");
1968 error("Don't know how to deserialize constant value of type '" +
1969 Ty->getDescription());
1973 // Check that we didn't read a null constant if they are implicit for this
1974 // type plane. Do not do this check for constantexprs, as they may be folded
1975 // to a null value in a way that isn't predicted when a .bc file is initially
1977 assert((!isa<Constant>(Result) || !cast<Constant>(Result)->isNullValue()) ||
1978 !hasImplicitNull(TypeID) &&
1979 "Cannot read null values from bytecode!");
1983 /// Resolve references for constants. This function resolves the forward
1984 /// referenced constants in the ConstantFwdRefs map. It uses the
1985 /// replaceAllUsesWith method of Value class to substitute the placeholder
1986 /// instance with the actual instance.
1987 void BytecodeReader::ResolveReferencesToConstant(Constant *NewV, unsigned Typ,
1989 ConstantRefsType::iterator I =
1990 ConstantFwdRefs.find(std::make_pair(Typ, Slot));
1991 if (I == ConstantFwdRefs.end()) return; // Never forward referenced?
1993 Value *PH = I->second; // Get the placeholder...
1994 PH->replaceAllUsesWith(NewV);
1995 delete PH; // Delete the old placeholder
1996 ConstantFwdRefs.erase(I); // Remove the map entry for it
1999 /// Parse the constant strings section.
2000 void BytecodeReader::ParseStringConstants(unsigned NumEntries, ValueTable &Tab){
2001 for (; NumEntries; --NumEntries) {
2003 if (read_typeid(Typ))
2004 error("Invalid type (type type) for string constant");
2005 const Type *Ty = getType(Typ);
2006 if (!isa<ArrayType>(Ty))
2007 error("String constant data invalid!");
2009 const ArrayType *ATy = cast<ArrayType>(Ty);
2010 if (ATy->getElementType() != Type::SByteTy &&
2011 ATy->getElementType() != Type::UByteTy)
2012 error("String constant data invalid!");
2014 // Read character data. The type tells us how long the string is.
2015 char *Data = reinterpret_cast<char *>(alloca(ATy->getNumElements()));
2016 read_data(Data, Data+ATy->getNumElements());
2018 std::vector<Constant*> Elements(ATy->getNumElements());
2019 const Type* ElemType = ATy->getElementType();
2020 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
2021 Elements[i] = ConstantInt::get(ElemType, (unsigned char)Data[i]);
2023 // Create the constant, inserting it as needed.
2024 Constant *C = ConstantArray::get(ATy, Elements);
2025 unsigned Slot = insertValue(C, Typ, Tab);
2026 ResolveReferencesToConstant(C, Typ, Slot);
2027 if (Handler) Handler->handleConstantString(cast<ConstantArray>(C));
2031 /// Parse the constant pool.
2032 void BytecodeReader::ParseConstantPool(ValueTable &Tab,
2033 TypeListTy &TypeTab,
2035 if (Handler) Handler->handleGlobalConstantsBegin();
2037 /// In LLVM 1.3 Type does not derive from Value so the types
2038 /// do not occupy a plane. Consequently, we read the types
2039 /// first in the constant pool.
2040 if (isFunction && !hasTypeDerivedFromValue) {
2041 unsigned NumEntries = read_vbr_uint();
2042 ParseTypes(TypeTab, NumEntries);
2045 while (moreInBlock()) {
2046 unsigned NumEntries = read_vbr_uint();
2048 bool isTypeType = read_typeid(Typ);
2050 /// In LLVM 1.2 and before, Types were written to the
2051 /// bytecode file in the "Type Type" plane (#12).
2052 /// In 1.3 plane 12 is now the label plane. Handle this here.
2054 ParseTypes(TypeTab, NumEntries);
2055 } else if (Typ == Type::VoidTyID) {
2056 /// Use of Type::VoidTyID is a misnomer. It actually means
2057 /// that the following plane is constant strings
2058 assert(&Tab == &ModuleValues && "Cannot read strings in functions!");
2059 ParseStringConstants(NumEntries, Tab);
2061 for (unsigned i = 0; i < NumEntries; ++i) {
2062 Value *V = ParseConstantPoolValue(Typ);
2063 assert(V && "ParseConstantPoolValue returned NULL!");
2064 unsigned Slot = insertValue(V, Typ, Tab);
2066 // If we are reading a function constant table, make sure that we adjust
2067 // the slot number to be the real global constant number.
2069 if (&Tab != &ModuleValues && Typ < ModuleValues.size() &&
2071 Slot += ModuleValues[Typ]->size();
2072 if (Constant *C = dyn_cast<Constant>(V))
2073 ResolveReferencesToConstant(C, Typ, Slot);
2078 // After we have finished parsing the constant pool, we had better not have
2079 // any dangling references left.
2080 if (!ConstantFwdRefs.empty()) {
2081 ConstantRefsType::const_iterator I = ConstantFwdRefs.begin();
2082 Constant* missingConst = I->second;
2083 error(utostr(ConstantFwdRefs.size()) +
2084 " unresolved constant reference exist. First one is '" +
2085 missingConst->getName() + "' of type '" +
2086 missingConst->getType()->getDescription() + "'.");
2089 checkPastBlockEnd("Constant Pool");
2090 if (Handler) Handler->handleGlobalConstantsEnd();
2093 /// Parse the contents of a function. Note that this function can be
2094 /// called lazily by materializeFunction
2095 /// @see materializeFunction
2096 void BytecodeReader::ParseFunctionBody(Function* F) {
2098 unsigned FuncSize = BlockEnd - At;
2099 GlobalValue::LinkageTypes Linkage = GlobalValue::ExternalLinkage;
2101 unsigned LinkageType = read_vbr_uint();
2102 switch (LinkageType) {
2103 case 0: Linkage = GlobalValue::ExternalLinkage; break;
2104 case 1: Linkage = GlobalValue::WeakLinkage; break;
2105 case 2: Linkage = GlobalValue::AppendingLinkage; break;
2106 case 3: Linkage = GlobalValue::InternalLinkage; break;
2107 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
2108 case 5: Linkage = GlobalValue::DLLImportLinkage; break;
2109 case 6: Linkage = GlobalValue::DLLExportLinkage; break;
2110 case 7: Linkage = GlobalValue::ExternalWeakLinkage; break;
2112 error("Invalid linkage type for Function.");
2113 Linkage = GlobalValue::InternalLinkage;
2117 F->setLinkage(Linkage);
2118 if (Handler) Handler->handleFunctionBegin(F,FuncSize);
2120 // Keep track of how many basic blocks we have read in...
2121 unsigned BlockNum = 0;
2122 bool InsertedArguments = false;
2124 BufPtr MyEnd = BlockEnd;
2125 while (At < MyEnd) {
2126 unsigned Type, Size;
2128 read_block(Type, Size);
2131 case BytecodeFormat::ConstantPoolBlockID:
2132 if (!InsertedArguments) {
2133 // Insert arguments into the value table before we parse the first basic
2134 // block in the function, but after we potentially read in the
2135 // compaction table.
2137 InsertedArguments = true;
2140 ParseConstantPool(FunctionValues, FunctionTypes, true);
2143 case BytecodeFormat::CompactionTableBlockID:
2144 ParseCompactionTable();
2147 case BytecodeFormat::BasicBlock: {
2148 if (!InsertedArguments) {
2149 // Insert arguments into the value table before we parse the first basic
2150 // block in the function, but after we potentially read in the
2151 // compaction table.
2153 InsertedArguments = true;
2156 BasicBlock *BB = ParseBasicBlock(BlockNum++);
2157 F->getBasicBlockList().push_back(BB);
2161 case BytecodeFormat::InstructionListBlockID: {
2162 // Insert arguments into the value table before we parse the instruction
2163 // list for the function, but after we potentially read in the compaction
2165 if (!InsertedArguments) {
2167 InsertedArguments = true;
2171 error("Already parsed basic blocks!");
2172 BlockNum = ParseInstructionList(F);
2176 case BytecodeFormat::SymbolTableBlockID:
2177 ParseSymbolTable(F, &F->getSymbolTable());
2183 error("Wrapped around reading bytecode.");
2188 // Malformed bc file if read past end of block.
2192 // Make sure there were no references to non-existant basic blocks.
2193 if (BlockNum != ParsedBasicBlocks.size())
2194 error("Illegal basic block operand reference");
2196 ParsedBasicBlocks.clear();
2198 // Resolve forward references. Replace any uses of a forward reference value
2199 // with the real value.
2200 while (!ForwardReferences.empty()) {
2201 std::map<std::pair<unsigned,unsigned>, Value*>::iterator
2202 I = ForwardReferences.begin();
2203 Value *V = getValue(I->first.first, I->first.second, false);
2204 Value *PlaceHolder = I->second;
2205 PlaceHolder->replaceAllUsesWith(V);
2206 ForwardReferences.erase(I);
2210 // If upgraded intrinsic functions were detected during reading of the
2211 // module information, then we need to look for instructions that need to
2212 // be upgraded. This can't be done while the instructions are read in because
2213 // additional instructions inserted mess up the slot numbering.
2214 if (!upgradedFunctions.empty()) {
2215 for (Function::iterator BI = F->begin(), BE = F->end(); BI != BE; ++BI)
2216 for (BasicBlock::iterator II = BI->begin(), IE = BI->end();
2218 if (CallInst* CI = dyn_cast<CallInst>(II++)) {
2219 std::map<Function*,Function*>::iterator FI =
2220 upgradedFunctions.find(CI->getCalledFunction());
2221 if (FI != upgradedFunctions.end())
2222 UpgradeIntrinsicCall(CI, FI->second);
2226 // Clear out function-level types...
2227 FunctionTypes.clear();
2228 CompactionTypes.clear();
2229 CompactionValues.clear();
2230 freeTable(FunctionValues);
2232 if (Handler) Handler->handleFunctionEnd(F);
2235 /// This function parses LLVM functions lazily. It obtains the type of the
2236 /// function and records where the body of the function is in the bytecode
2237 /// buffer. The caller can then use the ParseNextFunction and
2238 /// ParseAllFunctionBodies to get handler events for the functions.
2239 void BytecodeReader::ParseFunctionLazily() {
2240 if (FunctionSignatureList.empty())
2241 error("FunctionSignatureList empty!");
2243 Function *Func = FunctionSignatureList.back();
2244 FunctionSignatureList.pop_back();
2246 // Save the information for future reading of the function
2247 LazyFunctionLoadMap[Func] = LazyFunctionInfo(BlockStart, BlockEnd);
2249 // This function has a body but it's not loaded so it appears `External'.
2250 // Mark it as a `Ghost' instead to notify the users that it has a body.
2251 Func->setLinkage(GlobalValue::GhostLinkage);
2253 // Pretend we've `parsed' this function
2257 /// The ParserFunction method lazily parses one function. Use this method to
2258 /// casue the parser to parse a specific function in the module. Note that
2259 /// this will remove the function from what is to be included by
2260 /// ParseAllFunctionBodies.
2261 /// @see ParseAllFunctionBodies
2262 /// @see ParseBytecode
2263 bool BytecodeReader::ParseFunction(Function* Func, std::string* ErrMsg) {
2265 if (setjmp(context))
2268 // Find {start, end} pointers and slot in the map. If not there, we're done.
2269 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(Func);
2271 // Make sure we found it
2272 if (Fi == LazyFunctionLoadMap.end()) {
2273 error("Unrecognized function of type " + Func->getType()->getDescription());
2277 BlockStart = At = Fi->second.Buf;
2278 BlockEnd = Fi->second.EndBuf;
2279 assert(Fi->first == Func && "Found wrong function?");
2281 LazyFunctionLoadMap.erase(Fi);
2283 this->ParseFunctionBody(Func);
2287 /// The ParseAllFunctionBodies method parses through all the previously
2288 /// unparsed functions in the bytecode file. If you want to completely parse
2289 /// a bytecode file, this method should be called after Parsebytecode because
2290 /// Parsebytecode only records the locations in the bytecode file of where
2291 /// the function definitions are located. This function uses that information
2292 /// to materialize the functions.
2293 /// @see ParseBytecode
2294 bool BytecodeReader::ParseAllFunctionBodies(std::string* ErrMsg) {
2295 if (setjmp(context))
2298 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.begin();
2299 LazyFunctionMap::iterator Fe = LazyFunctionLoadMap.end();
2302 Function* Func = Fi->first;
2303 BlockStart = At = Fi->second.Buf;
2304 BlockEnd = Fi->second.EndBuf;
2305 ParseFunctionBody(Func);
2308 LazyFunctionLoadMap.clear();
2312 /// Parse the global type list
2313 void BytecodeReader::ParseGlobalTypes() {
2314 // Read the number of types
2315 unsigned NumEntries = read_vbr_uint();
2317 // Ignore the type plane identifier for types if the bc file is pre 1.3
2318 if (hasTypeDerivedFromValue)
2321 ParseTypes(ModuleTypes, NumEntries);
2324 /// Parse the Global info (types, global vars, constants)
2325 void BytecodeReader::ParseModuleGlobalInfo() {
2327 if (Handler) Handler->handleModuleGlobalsBegin();
2329 // SectionID - If a global has an explicit section specified, this map
2330 // remembers the ID until we can translate it into a string.
2331 std::map<GlobalValue*, unsigned> SectionID;
2333 // Read global variables...
2334 unsigned VarType = read_vbr_uint();
2335 while (VarType != Type::VoidTyID) { // List is terminated by Void
2336 // VarType Fields: bit0 = isConstant, bit1 = hasInitializer, bit2,3,4 =
2337 // Linkage, bit4+ = slot#
2338 unsigned SlotNo = VarType >> 5;
2339 if (sanitizeTypeId(SlotNo))
2340 error("Invalid type (type type) for global var!");
2341 unsigned LinkageID = (VarType >> 2) & 7;
2342 bool isConstant = VarType & 1;
2343 bool hasInitializer = (VarType & 2) != 0;
2344 unsigned Alignment = 0;
2345 unsigned GlobalSectionID = 0;
2347 // An extension word is present when linkage = 3 (internal) and hasinit = 0.
2348 if (LinkageID == 3 && !hasInitializer) {
2349 unsigned ExtWord = read_vbr_uint();
2350 // The extension word has this format: bit 0 = has initializer, bit 1-3 =
2351 // linkage, bit 4-8 = alignment (log2), bits 10+ = future use.
2352 hasInitializer = ExtWord & 1;
2353 LinkageID = (ExtWord >> 1) & 7;
2354 Alignment = (1 << ((ExtWord >> 4) & 31)) >> 1;
2356 if (ExtWord & (1 << 9)) // Has a section ID.
2357 GlobalSectionID = read_vbr_uint();
2360 GlobalValue::LinkageTypes Linkage;
2361 switch (LinkageID) {
2362 case 0: Linkage = GlobalValue::ExternalLinkage; break;
2363 case 1: Linkage = GlobalValue::WeakLinkage; break;
2364 case 2: Linkage = GlobalValue::AppendingLinkage; break;
2365 case 3: Linkage = GlobalValue::InternalLinkage; break;
2366 case 4: Linkage = GlobalValue::LinkOnceLinkage; break;
2367 case 5: Linkage = GlobalValue::DLLImportLinkage; break;
2368 case 6: Linkage = GlobalValue::DLLExportLinkage; break;
2369 case 7: Linkage = GlobalValue::ExternalWeakLinkage; break;
2371 error("Unknown linkage type: " + utostr(LinkageID));
2372 Linkage = GlobalValue::InternalLinkage;
2376 const Type *Ty = getType(SlotNo);
2378 error("Global has no type! SlotNo=" + utostr(SlotNo));
2380 if (!isa<PointerType>(Ty))
2381 error("Global not a pointer type! Ty= " + Ty->getDescription());
2383 const Type *ElTy = cast<PointerType>(Ty)->getElementType();
2385 // Create the global variable...
2386 GlobalVariable *GV = new GlobalVariable(ElTy, isConstant, Linkage,
2388 GV->setAlignment(Alignment);
2389 insertValue(GV, SlotNo, ModuleValues);
2391 if (GlobalSectionID != 0)
2392 SectionID[GV] = GlobalSectionID;
2394 unsigned initSlot = 0;
2395 if (hasInitializer) {
2396 initSlot = read_vbr_uint();
2397 GlobalInits.push_back(std::make_pair(GV, initSlot));
2400 // Notify handler about the global value.
2402 Handler->handleGlobalVariable(ElTy, isConstant, Linkage, SlotNo,initSlot);
2405 VarType = read_vbr_uint();
2408 // Read the function objects for all of the functions that are coming
2409 unsigned FnSignature = read_vbr_uint();
2411 if (hasNoFlagsForFunctions)
2412 FnSignature = (FnSignature << 5) + 1;
2414 // List is terminated by VoidTy.
2415 while (((FnSignature & (~0U >> 1)) >> 5) != Type::VoidTyID) {
2416 const Type *Ty = getType((FnSignature & (~0U >> 1)) >> 5);
2417 if (!isa<PointerType>(Ty) ||
2418 !isa<FunctionType>(cast<PointerType>(Ty)->getElementType())) {
2419 error("Function not a pointer to function type! Ty = " +
2420 Ty->getDescription());
2423 // We create functions by passing the underlying FunctionType to create...
2424 const FunctionType* FTy =
2425 cast<FunctionType>(cast<PointerType>(Ty)->getElementType());
2427 // Insert the place holder.
2428 Function *Func = new Function(FTy, GlobalValue::ExternalLinkage,
2431 insertValue(Func, (FnSignature & (~0U >> 1)) >> 5, ModuleValues);
2433 // Flags are not used yet.
2434 unsigned Flags = FnSignature & 31;
2436 // Save this for later so we know type of lazily instantiated functions.
2437 // Note that known-external functions do not have FunctionInfo blocks, so we
2438 // do not add them to the FunctionSignatureList.
2439 if ((Flags & (1 << 4)) == 0)
2440 FunctionSignatureList.push_back(Func);
2442 // Get the calling convention from the low bits.
2443 unsigned CC = Flags & 15;
2444 unsigned Alignment = 0;
2445 if (FnSignature & (1 << 31)) { // Has extension word?
2446 unsigned ExtWord = read_vbr_uint();
2447 Alignment = (1 << (ExtWord & 31)) >> 1;
2448 CC |= ((ExtWord >> 5) & 15) << 4;
2450 if (ExtWord & (1 << 10)) // Has a section ID.
2451 SectionID[Func] = read_vbr_uint();
2453 // Parse external declaration linkage
2454 switch ((ExtWord >> 11) & 3) {
2456 case 1: Func->setLinkage(Function::DLLImportLinkage); break;
2457 case 2: Func->setLinkage(Function::ExternalWeakLinkage); break;
2458 default: assert(0 && "Unsupported external linkage");
2462 Func->setCallingConv(CC-1);
2463 Func->setAlignment(Alignment);
2465 if (Handler) Handler->handleFunctionDeclaration(Func);
2467 // Get the next function signature.
2468 FnSignature = read_vbr_uint();
2469 if (hasNoFlagsForFunctions)
2470 FnSignature = (FnSignature << 5) + 1;
2473 // Now that the function signature list is set up, reverse it so that we can
2474 // remove elements efficiently from the back of the vector.
2475 std::reverse(FunctionSignatureList.begin(), FunctionSignatureList.end());
2477 /// SectionNames - This contains the list of section names encoded in the
2478 /// moduleinfoblock. Functions and globals with an explicit section index
2479 /// into this to get their section name.
2480 std::vector<std::string> SectionNames;
2482 if (hasInconsistentModuleGlobalInfo) {
2484 } else if (!hasNoDependentLibraries) {
2485 // If this bytecode format has dependent library information in it, read in
2486 // the number of dependent library items that follow.
2487 unsigned num_dep_libs = read_vbr_uint();
2488 std::string dep_lib;
2489 while (num_dep_libs--) {
2490 dep_lib = read_str();
2491 TheModule->addLibrary(dep_lib);
2493 Handler->handleDependentLibrary(dep_lib);
2496 // Read target triple and place into the module.
2497 std::string triple = read_str();
2498 TheModule->setTargetTriple(triple);
2500 Handler->handleTargetTriple(triple);
2502 if (!hasAlignment && At != BlockEnd) {
2503 // If the file has section info in it, read the section names now.
2504 unsigned NumSections = read_vbr_uint();
2505 while (NumSections--)
2506 SectionNames.push_back(read_str());
2509 // If the file has module-level inline asm, read it now.
2510 if (!hasAlignment && At != BlockEnd)
2511 TheModule->setModuleInlineAsm(read_str());
2514 // If any globals are in specified sections, assign them now.
2515 for (std::map<GlobalValue*, unsigned>::iterator I = SectionID.begin(), E =
2516 SectionID.end(); I != E; ++I)
2518 if (I->second > SectionID.size())
2519 error("SectionID out of range for global!");
2520 I->first->setSection(SectionNames[I->second-1]);
2523 // This is for future proofing... in the future extra fields may be added that
2524 // we don't understand, so we transparently ignore them.
2528 if (Handler) Handler->handleModuleGlobalsEnd();
2531 /// Parse the version information and decode it by setting flags on the
2532 /// Reader that enable backward compatibility of the reader.
2533 void BytecodeReader::ParseVersionInfo() {
2534 unsigned Version = read_vbr_uint();
2536 // Unpack version number: low four bits are for flags, top bits = version
2537 Module::Endianness Endianness;
2538 Module::PointerSize PointerSize;
2539 Endianness = (Version & 1) ? Module::BigEndian : Module::LittleEndian;
2540 PointerSize = (Version & 2) ? Module::Pointer64 : Module::Pointer32;
2542 bool hasNoEndianness = Version & 4;
2543 bool hasNoPointerSize = Version & 8;
2545 RevisionNum = Version >> 4;
2547 // Default values for the current bytecode version
2548 hasInconsistentModuleGlobalInfo = false;
2549 hasExplicitPrimitiveZeros = false;
2550 hasRestrictedGEPTypes = false;
2551 hasTypeDerivedFromValue = false;
2552 hasLongBlockHeaders = false;
2553 has32BitTypes = false;
2554 hasNoDependentLibraries = false;
2555 hasAlignment = false;
2556 hasNoUndefValue = false;
2557 hasNoFlagsForFunctions = false;
2558 hasNoUnreachableInst = false;
2559 hasSignlessInstructions = false;
2561 // Determine which backwards compatibility flags to set based on the
2562 // bytecode file's version number
2563 switch (RevisionNum) {
2564 case 0: // LLVM 1.0, 1.1 (Released)
2565 // Base LLVM 1.0 bytecode format.
2566 hasInconsistentModuleGlobalInfo = true;
2567 hasExplicitPrimitiveZeros = true;
2571 case 1: // LLVM 1.2 (Released)
2572 // LLVM 1.2 added explicit support for emitting strings efficiently.
2574 // Also, it fixed the problem where the size of the ModuleGlobalInfo block
2575 // included the size for the alignment at the end, where the rest of the
2578 // LLVM 1.2 and before required that GEP indices be ubyte constants for
2579 // structures and longs for sequential types.
2580 hasRestrictedGEPTypes = true;
2582 // LLVM 1.2 and before had the Type class derive from Value class. This
2583 // changed in release 1.3 and consequently LLVM 1.3 bytecode files are
2584 // written differently because Types can no longer be part of the
2585 // type planes for Values.
2586 hasTypeDerivedFromValue = true;
2590 case 2: // 1.2.5 (Not Released)
2592 // LLVM 1.2 and earlier had two-word block headers. This is a bit wasteful,
2593 // especially for small files where the 8 bytes per block is a large
2594 // fraction of the total block size. In LLVM 1.3, the block type and length
2595 // are compressed into a single 32-bit unsigned integer. 27 bits for length,
2596 // 5 bits for block type.
2597 hasLongBlockHeaders = true;
2599 // LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
2600 // this has been reduced to vbr_uint24. It shouldn't make much difference
2601 // since we haven't run into a module with > 24 million types, but for
2602 // safety the 24-bit restriction has been enforced in 1.3 to free some bits
2603 // in various places and to ensure consistency.
2604 has32BitTypes = true;
2606 // LLVM 1.2 and earlier did not provide a target triple nor a list of
2607 // libraries on which the bytecode is dependent. LLVM 1.3 provides these
2608 // features, for use in future versions of LLVM.
2609 hasNoDependentLibraries = true;
2613 case 3: // LLVM 1.3 (Released)
2614 // LLVM 1.3 and earlier caused alignment bytes to be written on some block
2615 // boundaries and at the end of some strings. In extreme cases (e.g. lots
2616 // of GEP references to a constant array), this can increase the file size
2617 // by 30% or more. In version 1.4 alignment is done away with completely.
2618 hasAlignment = true;
2622 case 4: // 1.3.1 (Not Released)
2623 // In version 4, we did not support the 'undef' constant.
2624 hasNoUndefValue = true;
2626 // In version 4 and above, we did not include space for flags for functions
2627 // in the module info block.
2628 hasNoFlagsForFunctions = true;
2630 // In version 4 and above, we did not include the 'unreachable' instruction
2631 // in the opcode numbering in the bytecode file.
2632 hasNoUnreachableInst = true;
2636 case 5: // 1.4 (Released)
2637 // In version 5 and prior, instructions were signless while integer types
2638 // were signed. In version 6, instructions became signed and types became
2639 // signless. For example in version 5 we have the DIV instruction but in
2640 // version 6 we have FDIV, SDIV and UDIV to replace it. This caused a
2641 // renumbering of the instruction codes in version 6 that must be dealt with
2642 // when reading old bytecode files.
2643 hasSignlessInstructions = true;
2647 case 6: // SignlessTypes Implementation (1.9 release)
2651 error("Unknown bytecode version number: " + itostr(RevisionNum));
2654 if (hasNoEndianness) Endianness = Module::AnyEndianness;
2655 if (hasNoPointerSize) PointerSize = Module::AnyPointerSize;
2657 TheModule->setEndianness(Endianness);
2658 TheModule->setPointerSize(PointerSize);
2660 if (Handler) Handler->handleVersionInfo(RevisionNum, Endianness, PointerSize);
2663 /// Parse a whole module.
2664 void BytecodeReader::ParseModule() {
2665 unsigned Type, Size;
2667 FunctionSignatureList.clear(); // Just in case...
2669 // Read into instance variables...
2673 bool SeenModuleGlobalInfo = false;
2674 bool SeenGlobalTypePlane = false;
2675 BufPtr MyEnd = BlockEnd;
2676 while (At < MyEnd) {
2678 read_block(Type, Size);
2682 case BytecodeFormat::GlobalTypePlaneBlockID:
2683 if (SeenGlobalTypePlane)
2684 error("Two GlobalTypePlane Blocks Encountered!");
2688 SeenGlobalTypePlane = true;
2691 case BytecodeFormat::ModuleGlobalInfoBlockID:
2692 if (SeenModuleGlobalInfo)
2693 error("Two ModuleGlobalInfo Blocks Encountered!");
2694 ParseModuleGlobalInfo();
2695 SeenModuleGlobalInfo = true;
2698 case BytecodeFormat::ConstantPoolBlockID:
2699 ParseConstantPool(ModuleValues, ModuleTypes,false);
2702 case BytecodeFormat::FunctionBlockID:
2703 ParseFunctionLazily();
2706 case BytecodeFormat::SymbolTableBlockID:
2707 ParseSymbolTable(0, &TheModule->getSymbolTable());
2713 error("Unexpected Block of Type #" + utostr(Type) + " encountered!");
2721 // After the module constant pool has been read, we can safely initialize
2722 // global variables...
2723 while (!GlobalInits.empty()) {
2724 GlobalVariable *GV = GlobalInits.back().first;
2725 unsigned Slot = GlobalInits.back().second;
2726 GlobalInits.pop_back();
2728 // Look up the initializer value...
2729 // FIXME: Preserve this type ID!
2731 const llvm::PointerType* GVType = GV->getType();
2732 unsigned TypeSlot = getTypeSlot(GVType->getElementType());
2733 if (Constant *CV = getConstantValue(TypeSlot, Slot)) {
2734 if (GV->hasInitializer())
2735 error("Global *already* has an initializer?!");
2736 if (Handler) Handler->handleGlobalInitializer(GV,CV);
2737 GV->setInitializer(CV);
2739 error("Cannot find initializer value.");
2742 if (!ConstantFwdRefs.empty())
2743 error("Use of undefined constants in a module");
2745 /// Make sure we pulled them all out. If we didn't then there's a declaration
2746 /// but a missing body. That's not allowed.
2747 if (!FunctionSignatureList.empty())
2748 error("Function declared, but bytecode stream ended before definition");
2751 /// This function completely parses a bytecode buffer given by the \p Buf
2752 /// and \p Length parameters.
2753 bool BytecodeReader::ParseBytecode(volatile BufPtr Buf, unsigned Length,
2754 const std::string &ModuleID,
2755 std::string* ErrMsg) {
2757 /// We handle errors by
2758 if (setjmp(context)) {
2759 // Cleanup after error
2760 if (Handler) Handler->handleError(ErrorMsg);
2764 if (decompressedBlock != 0 ) {
2765 ::free(decompressedBlock);
2766 decompressedBlock = 0;
2768 // Set caller's error message, if requested
2771 // Indicate an error occurred
2776 At = MemStart = BlockStart = Buf;
2777 MemEnd = BlockEnd = Buf + Length;
2779 // Create the module
2780 TheModule = new Module(ModuleID);
2782 if (Handler) Handler->handleStart(TheModule, Length);
2784 // Read the four bytes of the signature.
2785 unsigned Sig = read_uint();
2787 // If this is a compressed file
2788 if (Sig == ('l' | ('l' << 8) | ('v' << 16) | ('c' << 24))) {
2790 // Invoke the decompression of the bytecode. Note that we have to skip the
2791 // file's magic number which is not part of the compressed block. Hence,
2792 // the Buf+4 and Length-4. The result goes into decompressedBlock, a data
2793 // member for retention until BytecodeReader is destructed.
2794 unsigned decompressedLength = Compressor::decompressToNewBuffer(
2795 (char*)Buf+4,Length-4,decompressedBlock);
2797 // We must adjust the buffer pointers used by the bytecode reader to point
2798 // into the new decompressed block. After decompression, the
2799 // decompressedBlock will point to a contiguous memory area that has
2800 // the decompressed data.
2801 At = MemStart = BlockStart = Buf = (BufPtr) decompressedBlock;
2802 MemEnd = BlockEnd = Buf + decompressedLength;
2804 // else if this isn't a regular (uncompressed) bytecode file, then its
2805 // and error, generate that now.
2806 } else if (Sig != ('l' | ('l' << 8) | ('v' << 16) | ('m' << 24))) {
2807 error("Invalid bytecode signature: " + utohexstr(Sig));
2810 // Tell the handler we're starting a module
2811 if (Handler) Handler->handleModuleBegin(ModuleID);
2813 // Get the module block and size and verify. This is handled specially
2814 // because the module block/size is always written in long format. Other
2815 // blocks are written in short format so the read_block method is used.
2816 unsigned Type, Size;
2819 if (Type != BytecodeFormat::ModuleBlockID) {
2820 error("Expected Module Block! Type:" + utostr(Type) + ", Size:"
2824 // It looks like the darwin ranlib program is broken, and adds trailing
2825 // garbage to the end of some bytecode files. This hack allows the bc
2826 // reader to ignore trailing garbage on bytecode files.
2827 if (At + Size < MemEnd)
2828 MemEnd = BlockEnd = At+Size;
2830 if (At + Size != MemEnd)
2831 error("Invalid Top Level Block Length! Type:" + utostr(Type)
2832 + ", Size:" + utostr(Size));
2834 // Parse the module contents
2835 this->ParseModule();
2837 // Check for missing functions
2839 error("Function expected, but bytecode stream ended!");
2841 // Look for intrinsic functions to upgrade, upgrade them, and save the
2842 // mapping from old function to new for use later when instructions are
2844 for (Module::iterator FI = TheModule->begin(), FE = TheModule->end();
2846 if (Function* newF = UpgradeIntrinsicFunction(FI)) {
2847 upgradedFunctions.insert(std::make_pair(FI, newF));
2851 // Tell the handler we're done with the module
2853 Handler->handleModuleEnd(ModuleID);
2855 // Tell the handler we're finished the parse
2856 if (Handler) Handler->handleFinish();
2862 //===----------------------------------------------------------------------===//
2863 //=== Default Implementations of Handler Methods
2864 //===----------------------------------------------------------------------===//
2866 BytecodeHandler::~BytecodeHandler() {}