1 //===-- Reader.h - Interface To Bytecode Reading ----------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Reid Spencer and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This header file defines the interface to the Bytecode Reader which is
11 // responsible for correctly interpreting bytecode files (backwards compatible)
12 // and materializing a module from the bytecode read.
14 //===----------------------------------------------------------------------===//
16 #ifndef BYTECODE_PARSER_H
17 #define BYTECODE_PARSER_H
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/GlobalValue.h"
22 #include "llvm/Function.h"
23 #include "llvm/ModuleProvider.h"
24 #include "llvm/Bytecode/Analyzer.h"
30 class BytecodeHandler; ///< Forward declare the handler interface
32 /// This class defines the interface for parsing a buffer of bytecode. The
33 /// parser itself takes no action except to call the various functions of
34 /// the handler interface. The parser's sole responsibility is the correct
35 /// interpretation of the bytecode buffer. The handler is responsible for
36 /// instantiating and keeping track of all values. As a convenience, the parser
37 /// is responsible for materializing types and will pass them through the
38 /// handler interface as necessary.
39 /// @see BytecodeHandler
40 /// @brief Bytecode Reader interface
41 class BytecodeReader : public ModuleProvider {
43 /// @name Constructors
46 /// @brief Default constructor. By default, no handler is used.
47 BytecodeReader(BytecodeHandler* h = 0) {
48 decompressedBlock = 0;
54 if (decompressedBlock) {
55 ::free(decompressedBlock);
56 decompressedBlock = 0;
65 /// @brief A convenience type for the buffer pointer
66 typedef const unsigned char* BufPtr;
68 /// @brief The type used for a vector of potentially abstract types
69 typedef std::vector<PATypeHolder> TypeListTy;
71 /// This type provides a vector of Value* via the User class for
72 /// storage of Values that have been constructed when reading the
73 /// bytecode. Because of forward referencing, constant replacement
74 /// can occur so we ensure that our list of Value* is updated
75 /// properly through those transitions. This ensures that the
76 /// correct Value* is in our list when it comes time to associate
77 /// constants with global variables at the end of reading the
79 /// @brief A list of values as a User of those Values.
80 class ValueList : public User {
81 std::vector<Use> Uses;
83 ValueList() : User(Type::VoidTy, Value::ArgumentVal, 0, 0) {}
85 // vector compatibility methods
86 unsigned size() const { return getNumOperands(); }
87 void push_back(Value *V) {
88 Uses.push_back(Use(V, this));
89 OperandList = &Uses[0];
92 Value *back() const { return Uses.back(); }
93 void pop_back() { Uses.pop_back(); --NumOperands; }
94 bool empty() const { return NumOperands == 0; }
95 virtual void print(std::ostream& os) const {
96 for (unsigned i = 0; i < size(); ++i) {
98 getOperand(i)->print(os);
104 /// @brief A 2 dimensional table of values
105 typedef std::vector<ValueList*> ValueTable;
107 /// This map is needed so that forward references to constants can be looked
108 /// up by Type and slot number when resolving those references.
109 /// @brief A mapping of a Type/slot pair to a Constant*.
110 typedef std::map<std::pair<unsigned,unsigned>, Constant*> ConstantRefsType;
112 /// For lazy read-in of functions, we need to save the location in the
113 /// data stream where the function is located. This structure provides that
114 /// information. Lazy read-in is used mostly by the JIT which only wants to
115 /// resolve functions as it needs them.
116 /// @brief Keeps pointers to function contents for later use.
117 struct LazyFunctionInfo {
118 const unsigned char *Buf, *EndBuf;
119 LazyFunctionInfo(const unsigned char *B = 0, const unsigned char *EB = 0)
120 : Buf(B), EndBuf(EB) {}
123 /// @brief A mapping of functions to their LazyFunctionInfo for lazy reading.
124 typedef std::map<Function*, LazyFunctionInfo> LazyFunctionMap;
126 /// @brief A list of global variables and the slot number that initializes
128 typedef std::vector<std::pair<GlobalVariable*, unsigned> > GlobalInitsList;
130 /// This type maps a typeslot/valueslot pair to the corresponding Value*.
131 /// It is used for dealing with forward references as values are read in.
132 /// @brief A map for dealing with forward references of values.
133 typedef std::map<std::pair<unsigned,unsigned>,Value*> ForwardReferenceMap;
139 /// @brief Main interface to parsing a bytecode buffer.
141 const unsigned char *Buf, ///< Beginning of the bytecode buffer
142 unsigned Length, ///< Length of the bytecode buffer
143 const std::string &ModuleID ///< An identifier for the module constructed.
146 /// @brief Parse all function bodies
147 void ParseAllFunctionBodies();
149 /// @brief Parse the next function of specific type
150 void ParseFunction(Function* Func) ;
152 /// This method is abstract in the parent ModuleProvider class. Its
153 /// implementation is identical to the ParseFunction method.
154 /// @see ParseFunction
155 /// @brief Make a specific function materialize.
156 virtual void materializeFunction(Function *F) {
157 LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(F);
158 if (Fi == LazyFunctionLoadMap.end()) return;
162 /// This method is abstract in the parent ModuleProvider class. Its
163 /// implementation is identical to ParseAllFunctionBodies.
164 /// @see ParseAllFunctionBodies
165 /// @brief Make the whole module materialize
166 virtual Module* materializeModule() {
167 ParseAllFunctionBodies();
171 /// This method is provided by the parent ModuleProvde class and overriden
172 /// here. It simply releases the module from its provided and frees up our
174 /// @brief Release our hold on the generated module
175 Module* releaseModule() {
176 // Since we're losing control of this Module, we must hand it back complete
177 Module *M = ModuleProvider::releaseModule();
183 /// @name Parsing Units For Subclasses
186 /// @brief Parse whole module scope
189 /// @brief Parse the version information block
190 void ParseVersionInfo();
192 /// @brief Parse the ModuleGlobalInfo block
193 void ParseModuleGlobalInfo();
195 /// @brief Parse a symbol table
196 void ParseSymbolTable( Function* Func, SymbolTable *ST);
198 /// @brief Parse functions lazily.
199 void ParseFunctionLazily();
201 /// @brief Parse a function body
202 void ParseFunctionBody(Function* Func);
204 /// @brief Parse the type list portion of a compaction table
205 void ParseCompactionTypes(unsigned NumEntries);
207 /// @brief Parse a compaction table
208 void ParseCompactionTable();
210 /// @brief Parse global types
211 void ParseGlobalTypes();
213 /// @brief Parse a basic block (for LLVM 1.0 basic block blocks)
214 BasicBlock* ParseBasicBlock(unsigned BlockNo);
216 /// @brief parse an instruction list (for post LLVM 1.0 instruction lists
217 /// with blocks differentiated by terminating instructions.
218 unsigned ParseInstructionList(
219 Function* F ///< The function into which BBs will be inserted
222 /// @brief Parse a single instruction.
223 void ParseInstruction(
224 std::vector<unsigned>& Args, ///< The arguments to be filled in
225 BasicBlock* BB ///< The BB the instruction goes in
228 /// @brief Parse the whole constant pool
229 void ParseConstantPool(ValueTable& Values, TypeListTy& Types,
232 /// @brief Parse a single constant value
233 Constant* ParseConstantValue(unsigned TypeID);
235 /// @brief Parse a block of types constants
236 void ParseTypes(TypeListTy &Tab, unsigned NumEntries);
238 /// @brief Parse a single type constant
239 const Type *ParseType();
241 /// @brief Parse a string constants block
242 void ParseStringConstants(unsigned NumEntries, ValueTable &Tab);
248 char* decompressedBlock; ///< Result of decompression
249 BufPtr MemStart; ///< Start of the memory buffer
250 BufPtr MemEnd; ///< End of the memory buffer
251 BufPtr BlockStart; ///< Start of current block being parsed
252 BufPtr BlockEnd; ///< End of current block being parsed
253 BufPtr At; ///< Where we're currently parsing at
255 /// Information about the module, extracted from the bytecode revision number.
257 unsigned char RevisionNum; // The rev # itself
259 /// Flags to distinguish LLVM 1.0 & 1.1 bytecode formats (revision #0)
261 /// Revision #0 had an explicit alignment of data only for the
262 /// ModuleGlobalInfo block. This was fixed to be like all other blocks in 1.2
263 bool hasInconsistentModuleGlobalInfo;
265 /// Revision #0 also explicitly encoded zero values for primitive types like
267 bool hasExplicitPrimitiveZeros;
269 // Flags to control features specific the LLVM 1.2 and before (revision #1)
271 /// LLVM 1.2 and earlier required that getelementptr structure indices were
272 /// ubyte constants and that sequential type indices were longs.
273 bool hasRestrictedGEPTypes;
275 /// LLVM 1.2 and earlier had class Type deriving from Value and the Type
276 /// objects were located in the "Type Type" plane of various lists in read
277 /// by the bytecode reader. In LLVM 1.3 this is no longer the case. Types are
278 /// completely distinct from Values. Consequently, Types are written in fixed
279 /// locations in LLVM 1.3. This flag indicates that the older Type derived
280 /// from Value style of bytecode file is being read.
281 bool hasTypeDerivedFromValue;
283 /// LLVM 1.2 and earlier encoded block headers as two uint (8 bytes), one for
284 /// the size and one for the type. This is a bit wasteful, especially for
285 /// small files where the 8 bytes per block is a large fraction of the total
286 /// block size. In LLVM 1.3, the block type and length are encoded into a
287 /// single uint32 by restricting the number of block types (limit 31) and the
288 /// maximum size of a block (limit 2^27-1=134,217,727). Note that the module
289 /// block still uses the 8-byte format so the maximum size of a file can be
290 /// 2^32-1 bytes long.
291 bool hasLongBlockHeaders;
293 /// LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
294 /// this has been reduced to vbr_uint24. It shouldn't make much difference
295 /// since we haven't run into a module with > 24 million types, but for safety
296 /// the 24-bit restriction has been enforced in 1.3 to free some bits in
297 /// various places and to ensure consistency. In particular, global vars are
298 /// restricted to 24-bits.
301 /// LLVM 1.2 and earlier did not provide a target triple nor a list of
302 /// libraries on which the bytecode is dependent. LLVM 1.3 provides these
303 /// features, for use in future versions of LLVM.
304 bool hasNoDependentLibraries;
306 /// LLVM 1.3 and earlier caused blocks and other fields to start on 32-bit
307 /// aligned boundaries. This can lead to as much as 30% bytecode size overhead
308 /// in various corner cases (lots of long instructions). In LLVM 1.4,
309 /// alignment of bytecode fields was done away with completely.
312 // In version 4 and earlier, the bytecode format did not support the 'undef'
314 bool hasNoUndefValue;
316 // In version 4 and earlier, the bytecode format did not save space for flags
317 // in the global info block for functions.
318 bool hasNoFlagsForFunctions;
320 // In version 4 and earlier, there was no opcode space reserved for the
321 // unreachable instruction.
322 bool hasNoUnreachableInst;
324 /// CompactionTypes - If a compaction table is active in the current function,
325 /// this is the mapping that it contains. We keep track of what resolved type
326 /// it is as well as what global type entry it is.
327 std::vector<std::pair<const Type*, unsigned> > CompactionTypes;
329 /// @brief If a compaction table is active in the current function,
330 /// this is the mapping that it contains.
331 std::vector<std::vector<Value*> > CompactionValues;
333 /// @brief This vector is used to deal with forward references to types in
335 TypeListTy ModuleTypes;
337 /// @brief This is an inverse mapping of ModuleTypes from the type to an
338 /// index. Because refining types causes the index of this map to be
339 /// invalidated, any time we refine a type, we clear this cache and recompute
340 /// it next time we need it. These entries are ordered by the pointer value.
341 std::vector<std::pair<const Type*, unsigned> > ModuleTypeIDCache;
343 /// @brief This vector is used to deal with forward references to types in
345 TypeListTy FunctionTypes;
347 /// When the ModuleGlobalInfo section is read, we create a Function object
348 /// for each function in the module. When the function is loaded, after the
349 /// module global info is read, this Function is populated. Until then, the
350 /// functions in this vector just hold the function signature.
351 std::vector<Function*> FunctionSignatureList;
353 /// @brief This is the table of values belonging to the current function
354 ValueTable FunctionValues;
356 /// @brief This is the table of values belonging to the module (global)
357 ValueTable ModuleValues;
359 /// @brief This keeps track of function level forward references.
360 ForwardReferenceMap ForwardReferences;
362 /// @brief The basic blocks we've parsed, while parsing a function.
363 std::vector<BasicBlock*> ParsedBasicBlocks;
365 /// This maintains a mapping between <Type, Slot #>'s and forward references
366 /// to constants. Such values may be referenced before they are defined, and
367 /// if so, the temporary object that they represent is held here. @brief
368 /// Temporary place for forward references to constants.
369 ConstantRefsType ConstantFwdRefs;
371 /// Constant values are read in after global variables. Because of this, we
372 /// must defer setting the initializers on global variables until after module
373 /// level constants have been read. In the mean time, this list keeps track
374 /// of what we must do.
375 GlobalInitsList GlobalInits;
377 // For lazy reading-in of functions, we need to save away several pieces of
378 // information about each function: its begin and end pointer in the buffer
379 // and its FunctionSlot.
380 LazyFunctionMap LazyFunctionLoadMap;
382 /// This stores the parser's handler which is used for handling tasks other
383 /// just than reading bytecode into the IR. If this is non-null, calls on
384 /// the (polymorphic) BytecodeHandler interface (see llvm/Bytecode/Handler.h)
385 /// will be made to report the logical structure of the bytecode file. What
386 /// the handler does with the events it receives is completely orthogonal to
387 /// the business of parsing the bytecode and building the IR. This is used,
388 /// for example, by the llvm-abcd tool for analysis of byte code.
389 /// @brief Handler for parsing events.
390 BytecodeHandler* Handler;
393 /// @name Implementation Details
396 /// @brief Determines if this module has a function or not.
397 bool hasFunctions() { return ! FunctionSignatureList.empty(); }
399 /// @brief Determines if the type id has an implicit null value.
400 bool hasImplicitNull(unsigned TyID );
402 /// @brief Converts a type slot number to its Type*
403 const Type *getType(unsigned ID);
405 /// @brief Converts a pre-sanitized type slot number to its Type* and
406 /// sanitizes the type id.
407 inline const Type* getSanitizedType(unsigned& ID );
409 /// @brief Read in and get a sanitized type id
410 inline const Type* readSanitizedType();
412 /// @brief Converts a Type* to its type slot number
413 unsigned getTypeSlot(const Type *Ty);
415 /// @brief Converts a normal type slot number to a compacted type slot num.
416 unsigned getCompactionTypeSlot(unsigned type);
418 /// @brief Gets the global type corresponding to the TypeId
419 const Type *getGlobalTableType(unsigned TypeId);
421 /// This is just like getTypeSlot, but when a compaction table is in use,
423 unsigned getGlobalTableTypeSlot(const Type *Ty);
425 /// @brief Get a value from its typeid and slot number
426 Value* getValue(unsigned TypeID, unsigned num, bool Create = true);
428 /// @brief Get a value from its type and slot number, ignoring compaction
430 Value *getGlobalTableValue(unsigned TyID, unsigned SlotNo);
432 /// @brief Get a basic block for current function
433 BasicBlock *getBasicBlock(unsigned ID);
435 /// @brief Get a constant value from its typeid and value slot.
436 Constant* getConstantValue(unsigned typeSlot, unsigned valSlot);
438 /// @brief Convenience function for getting a constant value when
439 /// the Type has already been resolved.
440 Constant* getConstantValue(const Type *Ty, unsigned valSlot) {
441 return getConstantValue(getTypeSlot(Ty), valSlot);
444 /// @brief Insert a newly created value
445 unsigned insertValue(Value *V, unsigned Type, ValueTable &Table);
447 /// @brief Insert the arguments of a function.
448 void insertArguments(Function* F );
450 /// @brief Resolve all references to the placeholder (if any) for the
452 void ResolveReferencesToConstant(Constant *C, unsigned Typ, unsigned Slot);
454 /// @brief Release our memory.
456 freeTable(FunctionValues);
457 freeTable(ModuleValues);
460 /// @brief Free a table, making sure to free the ValueList in the table.
461 void freeTable(ValueTable &Tab) {
462 while (!Tab.empty()) {
468 inline void error(std::string errmsg);
470 BytecodeReader(const BytecodeReader &); // DO NOT IMPLEMENT
471 void operator=(const BytecodeReader &); // DO NOT IMPLEMENT
474 /// @name Reader Primitives
478 /// @brief Is there more to parse in the current block?
479 inline bool moreInBlock();
481 /// @brief Have we read past the end of the block
482 inline void checkPastBlockEnd(const char * block_name);
484 /// @brief Align to 32 bits
485 inline void align32();
487 /// @brief Read an unsigned integer as 32-bits
488 inline unsigned read_uint();
490 /// @brief Read an unsigned integer with variable bit rate encoding
491 inline unsigned read_vbr_uint();
493 /// @brief Read an unsigned integer of no more than 24-bits with variable
494 /// bit rate encoding.
495 inline unsigned read_vbr_uint24();
497 /// @brief Read an unsigned 64-bit integer with variable bit rate encoding.
498 inline uint64_t read_vbr_uint64();
500 /// @brief Read a signed 64-bit integer with variable bit rate encoding.
501 inline int64_t read_vbr_int64();
503 /// @brief Read a string
504 inline std::string read_str();
506 /// @brief Read a float value
507 inline void read_float(float& FloatVal);
509 /// @brief Read a double value
510 inline void read_double(double& DoubleVal);
512 /// @brief Read an arbitrary data chunk of fixed length
513 inline void read_data(void *Ptr, void *End);
515 /// @brief Read a bytecode block header
516 inline void read_block(unsigned &Type, unsigned &Size);
518 /// @brief Read a type identifier and sanitize it.
519 inline bool read_typeid(unsigned &TypeId);
521 /// @brief Recalculate type ID for pre 1.3 bytecode files.
522 inline bool sanitizeTypeId(unsigned &TypeId );
526 /// @brief A function for creating a BytecodeAnalzer as a handler
527 /// for the Bytecode reader.
528 BytecodeHandler* createBytecodeAnalyzerHandler(BytecodeAnalysis& bca,
529 std::ostream* output );
532 } // End llvm namespace