llvm-6502/lib/Bytecode/Reader/Reader.h
Reid Spencer 572c256519 Add boolean file format flags in preparation for version 5 bytecode.
Remove the "processFunctions" boolean from ParseBytecode as it is no
longer needed. This is part of avoiding double reading of functions
when analyzing bytecode.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@15982 91177308-0d34-0410-b5e6-96231b3b80d8
2004-08-21 20:50:49 +00:00

523 lines
19 KiB
C++

//===-- Reader.h - Interface To Bytecode Reading ----------------*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by Reid Spencer and is distributed under the
// University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This header file defines the interface to the Bytecode Reader which is
// responsible for correctly interpreting bytecode files (backwards compatible)
// and materializing a module from the bytecode read.
//
//===----------------------------------------------------------------------===//
#ifndef BYTECODE_PARSER_H
#define BYTECODE_PARSER_H
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalValue.h"
#include "llvm/Function.h"
#include "llvm/ModuleProvider.h"
#include "llvm/Bytecode/Analyzer.h"
#include <utility>
#include <map>
namespace llvm {
class BytecodeHandler; ///< Forward declare the handler interface
/// This class defines the interface for parsing a buffer of bytecode. The
/// parser itself takes no action except to call the various functions of
/// the handler interface. The parser's sole responsibility is the correct
/// interpretation of the bytecode buffer. The handler is responsible for
/// instantiating and keeping track of all values. As a convenience, the parser
/// is responsible for materializing types and will pass them through the
/// handler interface as necessary.
/// @see BytecodeHandler
/// @brief Bytecode Reader interface
class BytecodeReader : public ModuleProvider {
/// @name Constructors
/// @{
public:
/// @brief Default constructor. By default, no handler is used.
BytecodeReader(
BytecodeHandler* h = 0
) {
Handler = h;
}
~BytecodeReader() { freeState(); }
/// @}
/// @name Types
/// @{
public:
/// @brief A convenience type for the buffer pointer
typedef const unsigned char* BufPtr;
/// @brief The type used for a vector of potentially abstract types
typedef std::vector<PATypeHolder> TypeListTy;
/// This type provides a vector of Value* via the User class for
/// storage of Values that have been constructed when reading the
/// bytecode. Because of forward referencing, constant replacement
/// can occur so we ensure that our list of Value* is updated
/// properly through those transitions. This ensures that the
/// correct Value* is in our list when it comes time to associate
/// constants with global variables at the end of reading the
/// globals section.
/// @brief A list of values as a User of those Values.
struct ValueList : public User {
ValueList() : User(Type::VoidTy, Value::ValueListVal) {}
// vector compatibility methods
unsigned size() const { return getNumOperands(); }
void push_back(Value *V) { Operands.push_back(Use(V, this)); }
Value *back() const { return Operands.back(); }
void pop_back() { Operands.pop_back(); }
bool empty() const { return Operands.empty(); }
// must override this
virtual void print(std::ostream& os) const {
for ( unsigned i = 0; i < size(); i++ ) {
os << i << " ";
getOperand(i)->print(os);
os << "\n";
}
}
};
/// @brief A 2 dimensional table of values
typedef std::vector<ValueList*> ValueTable;
/// This map is needed so that forward references to constants can be looked
/// up by Type and slot number when resolving those references.
/// @brief A mapping of a Type/slot pair to a Constant*.
typedef std::map<std::pair<const Type*,unsigned>, Constant*> ConstantRefsType;
/// For lazy read-in of functions, we need to save the location in the
/// data stream where the function is located. This structure provides that
/// information. Lazy read-in is used mostly by the JIT which only wants to
/// resolve functions as it needs them.
/// @brief Keeps pointers to function contents for later use.
struct LazyFunctionInfo {
const unsigned char *Buf, *EndBuf;
LazyFunctionInfo(const unsigned char *B = 0, const unsigned char *EB = 0)
: Buf(B), EndBuf(EB) {}
};
/// @brief A mapping of functions to their LazyFunctionInfo for lazy reading.
typedef std::map<Function*, LazyFunctionInfo> LazyFunctionMap;
/// @brief A list of global variables and the slot number that initializes
/// them.
typedef std::vector<std::pair<GlobalVariable*, unsigned> > GlobalInitsList;
/// This type maps a typeslot/valueslot pair to the corresponding Value*.
/// It is used for dealing with forward references as values are read in.
/// @brief A map for dealing with forward references of values.
typedef std::map<std::pair<unsigned,unsigned>,Value*> ForwardReferenceMap;
/// @}
/// @name Methods
/// @{
public:
/// @brief Main interface to parsing a bytecode buffer.
void ParseBytecode(
const unsigned char *Buf, ///< Beginning of the bytecode buffer
unsigned Length, ///< Length of the bytecode buffer
const std::string &ModuleID ///< An identifier for the module constructed.
);
/// @brief Parse all function bodies
void ParseAllFunctionBodies();
/// @brief Parse the next function of specific type
void ParseFunction(Function* Func) ;
/// This method is abstract in the parent ModuleProvider class. Its
/// implementation is identical to the ParseFunction method.
/// @see ParseFunction
/// @brief Make a specific function materialize.
virtual void materializeFunction(Function *F) {
LazyFunctionMap::iterator Fi = LazyFunctionLoadMap.find(F);
if (Fi == LazyFunctionLoadMap.end()) return;
ParseFunction(F);
}
/// This method is abstract in the parent ModuleProvider class. Its
/// implementation is identical to ParseAllFunctionBodies.
/// @see ParseAllFunctionBodies
/// @brief Make the whole module materialize
virtual Module* materializeModule() {
ParseAllFunctionBodies();
return TheModule;
}
/// This method is provided by the parent ModuleProvde class and overriden
/// here. It simply releases the module from its provided and frees up our
/// state.
/// @brief Release our hold on the generated module
Module* releaseModule() {
// Since we're losing control of this Module, we must hand it back complete
Module *M = ModuleProvider::releaseModule();
freeState();
return M;
}
/// @}
/// @name Parsing Units For Subclasses
/// @{
protected:
/// @brief Parse whole module scope
void ParseModule();
/// @brief Parse the version information block
void ParseVersionInfo();
/// @brief Parse the ModuleGlobalInfo block
void ParseModuleGlobalInfo();
/// @brief Parse a symbol table
void ParseSymbolTable( Function* Func, SymbolTable *ST);
/// @brief Parse functions lazily.
void ParseFunctionLazily();
/// @brief Parse a function body
void ParseFunctionBody(Function* Func);
/// @brief Parse the type list portion of a compaction table
void ParseCompactionTypes(unsigned NumEntries);
/// @brief Parse a compaction table
void ParseCompactionTable();
/// @brief Parse global types
void ParseGlobalTypes();
/// @brief Parse a basic block (for LLVM 1.0 basic block blocks)
BasicBlock* ParseBasicBlock(unsigned BlockNo);
/// @brief parse an instruction list (for post LLVM 1.0 instruction lists
/// with blocks differentiated by terminating instructions.
unsigned ParseInstructionList(
Function* F ///< The function into which BBs will be inserted
);
/// @brief Parse a single instruction.
void ParseInstruction(
std::vector<unsigned>& Args, ///< The arguments to be filled in
BasicBlock* BB ///< The BB the instruction goes in
);
/// @brief Parse the whole constant pool
void ParseConstantPool(ValueTable& Values, TypeListTy& Types,
bool isFunction);
/// @brief Parse a single constant value
Constant* ParseConstantValue(unsigned TypeID);
/// @brief Parse a block of types constants
void ParseTypes(TypeListTy &Tab, unsigned NumEntries);
/// @brief Parse a single type constant
const Type *ParseType();
/// @brief Parse a string constants block
void ParseStringConstants(unsigned NumEntries, ValueTable &Tab);
/// @}
/// @name Data
/// @{
private:
BufPtr MemStart; ///< Start of the memory buffer
BufPtr MemEnd; ///< End of the memory buffer
BufPtr BlockStart; ///< Start of current block being parsed
BufPtr BlockEnd; ///< End of current block being parsed
BufPtr At; ///< Where we're currently parsing at
/// Information about the module, extracted from the bytecode revision number.
///
unsigned char RevisionNum; // The rev # itself
/// Flags to distinguish LLVM 1.0 & 1.1 bytecode formats (revision #0)
/// Revision #0 had an explicit alignment of data only for the
/// ModuleGlobalInfo block. This was fixed to be like all other blocks in 1.2
bool hasInconsistentModuleGlobalInfo;
/// Revision #0 also explicitly encoded zero values for primitive types like
/// int/sbyte/etc.
bool hasExplicitPrimitiveZeros;
// Flags to control features specific the LLVM 1.2 and before (revision #1)
/// LLVM 1.2 and earlier required that getelementptr structure indices were
/// ubyte constants and that sequential type indices were longs.
bool hasRestrictedGEPTypes;
/// LLVM 1.2 and earlier had class Type deriving from Value and the Type
/// objects were located in the "Type Type" plane of various lists in read
/// by the bytecode reader. In LLVM 1.3 this is no longer the case. Types are
/// completely distinct from Values. Consequently, Types are written in fixed
/// locations in LLVM 1.3. This flag indicates that the older Type derived
/// from Value style of bytecode file is being read.
bool hasTypeDerivedFromValue;
/// LLVM 1.2 and earlier encoded block headers as two uint (8 bytes), one for
/// the size and one for the type. This is a bit wasteful, especially for
/// small files where the 8 bytes per block is a large fraction of the total
/// block size. In LLVM 1.3, the block type and length are encoded into a
/// single uint32 by restricting the number of block types (limit 31) and the
/// maximum size of a block (limit 2^27-1=134,217,727). Note that the module
/// block still uses the 8-byte format so the maximum size of a file can be
/// 2^32-1 bytes long.
bool hasLongBlockHeaders;
/// LLVM 1.2 and earlier wrote type slot numbers as vbr_uint32. In LLVM 1.3
/// this has been reduced to vbr_uint24. It shouldn't make much difference
/// since we haven't run into a module with > 24 million types, but for safety
/// the 24-bit restriction has been enforced in 1.3 to free some bits in
/// various places and to ensure consistency. In particular, global vars are
/// restricted to 24-bits.
bool has32BitTypes;
/// LLVM 1.2 and earlier did not provide a target triple nor a list of
/// libraries on which the bytecode is dependent. LLVM 1.3 provides these
/// features, for use in future versions of LLVM.
bool hasNoDependentLibraries;
/// LLVM 1.3 and earlier caused blocks and other fields to start on 32-bit
/// aligned boundaries. This can lead to as much as 30% bytecode size overhead
/// in various corner cases (lots of long instructions). In LLVM 1.4,
/// alignment of bytecode fields was done away with completely.
bool hasAlignment;
// In version 4, basic blocks have a minimum index of 0 whereas all the
// other primitives have a minimum index of 1 (because 0 is the "null"
// value. In version 5, we made this consistent.
bool hasInconsistentBBSlotNums;
// In version 4, the types SByte and UByte were encoded as vbr_uint so that
// signed values > 63 and unsigned values >127 would be encoded as two
// bytes. In version 5, they are encoded directly in a single byte.
bool hasVBRByteTypes;
// In version 4, modules begin with a "Module Block" which encodes a 4-byte
// integer value 0x01 to identify the module block. This is unnecessary and
// removed in version 5.
bool hasUnnecessaryModuleBlockId;
/// CompactionTypes - If a compaction table is active in the current function,
/// this is the mapping that it contains. We keep track of what resolved type
/// it is as well as what global type entry it is.
std::vector<std::pair<const Type*, unsigned> > CompactionTypes;
/// @brief If a compaction table is active in the current function,
/// this is the mapping that it contains.
std::vector<std::vector<Value*> > CompactionValues;
/// @brief This vector is used to deal with forward references to types in
/// a module.
TypeListTy ModuleTypes;
/// @brief This vector is used to deal with forward references to types in
/// a function.
TypeListTy FunctionTypes;
/// When the ModuleGlobalInfo section is read, we create a Function object
/// for each function in the module. When the function is loaded, after the
/// module global info is read, this Function is populated. Until then, the
/// functions in this vector just hold the function signature.
std::vector<Function*> FunctionSignatureList;
/// @brief This is the table of values belonging to the current function
ValueTable FunctionValues;
/// @brief This is the table of values belonging to the module (global)
ValueTable ModuleValues;
/// @brief This keeps track of function level forward references.
ForwardReferenceMap ForwardReferences;
/// @brief The basic blocks we've parsed, while parsing a function.
std::vector<BasicBlock*> ParsedBasicBlocks;
/// This maintains a mapping between <Type, Slot #>'s and
/// forward references to constants. Such values may be referenced before they
/// are defined, and if so, the temporary object that they represent is held
/// here.
/// @brief Temporary place for forward references to constants.
ConstantRefsType ConstantFwdRefs;
/// Constant values are read in after global variables. Because of this, we
/// must defer setting the initializers on global variables until after module
/// level constants have been read. In the mean time, this list keeps track of
/// what we must do.
GlobalInitsList GlobalInits;
// For lazy reading-in of functions, we need to save away several pieces of
// information about each function: its begin and end pointer in the buffer
// and its FunctionSlot.
LazyFunctionMap LazyFunctionLoadMap;
/// This stores the parser's handler which is used for handling tasks other
/// just than reading bytecode into the IR. If this is non-null, calls on
/// the (polymorphic) BytecodeHandler interface (see llvm/Bytecode/Handler.h)
/// will be made to report the logical structure of the bytecode file. What
/// the handler does with the events it receives is completely orthogonal to
/// the business of parsing the bytecode and building the IR. This is used,
/// for example, by the llvm-abcd tool for analysis of byte code.
/// @brief Handler for parsing events.
BytecodeHandler* Handler;
/// @}
/// @name Implementation Details
/// @{
private:
/// @brief Determines if this module has a function or not.
bool hasFunctions() { return ! FunctionSignatureList.empty(); }
/// @brief Determines if the type id has an implicit null value.
bool hasImplicitNull(unsigned TyID );
/// @brief Converts a type slot number to its Type*
const Type *getType(unsigned ID);
/// @brief Converts a pre-sanitized type slot number to its Type* and
/// sanitizes the type id.
inline const Type* getSanitizedType(unsigned& ID );
/// @brief Read in and get a sanitized type id
inline const Type* BytecodeReader::readSanitizedType();
/// @brief Converts a Type* to its type slot number
unsigned getTypeSlot(const Type *Ty);
/// @brief Converts a normal type slot number to a compacted type slot num.
unsigned getCompactionTypeSlot(unsigned type);
/// @brief Gets the global type corresponding to the TypeId
const Type *getGlobalTableType(unsigned TypeId);
/// This is just like getTypeSlot, but when a compaction table is in use,
/// it is ignored.
unsigned getGlobalTableTypeSlot(const Type *Ty);
/// @brief Get a value from its typeid and slot number
Value* getValue(unsigned TypeID, unsigned num, bool Create = true);
/// @brief Get a value from its type and slot number, ignoring compaction
/// tables.
Value *getGlobalTableValue(unsigned TyID, unsigned SlotNo);
/// @brief Get a basic block for current function
BasicBlock *getBasicBlock(unsigned ID);
/// @brief Get a constant value from its typeid and value slot.
Constant* getConstantValue(unsigned typeSlot, unsigned valSlot);
/// @brief Convenience function for getting a constant value when
/// the Type has already been resolved.
Constant* getConstantValue(const Type *Ty, unsigned valSlot) {
return getConstantValue(getTypeSlot(Ty), valSlot);
}
/// @brief Insert a newly created value
unsigned insertValue(Value *V, unsigned Type, ValueTable &Table);
/// @brief Insert the arguments of a function.
void insertArguments(Function* F );
/// @brief Resolve all references to the placeholder (if any) for the
/// given constant.
void ResolveReferencesToConstant(Constant *C, unsigned Slot);
/// @brief Release our memory.
void freeState() {
freeTable(FunctionValues);
freeTable(ModuleValues);
}
/// @brief Free a table, making sure to free the ValueList in the table.
void freeTable(ValueTable &Tab) {
while (!Tab.empty()) {
delete Tab.back();
Tab.pop_back();
}
}
inline void error(std::string errmsg);
BytecodeReader(const BytecodeReader &); // DO NOT IMPLEMENT
void operator=(const BytecodeReader &); // DO NOT IMPLEMENT
/// @}
/// @name Reader Primitives
/// @{
private:
/// @brief Is there more to parse in the current block?
inline bool moreInBlock();
/// @brief Have we read past the end of the block
inline void checkPastBlockEnd(const char * block_name);
/// @brief Align to 32 bits
inline void align32();
/// @brief Read an unsigned integer as 32-bits
inline unsigned read_uint();
/// @brief Read an unsigned integer with variable bit rate encoding
inline unsigned read_vbr_uint();
/// @brief Read an unsigned integer of no more than 24-bits with variable
/// bit rate encoding.
inline unsigned read_vbr_uint24();
/// @brief Read an unsigned 64-bit integer with variable bit rate encoding.
inline uint64_t read_vbr_uint64();
/// @brief Read a signed 64-bit integer with variable bit rate encoding.
inline int64_t read_vbr_int64();
/// @brief Read a string
inline std::string read_str();
/// @brief Read a float value
inline void read_float(float& FloatVal);
/// @brief Read a double value
inline void read_double(double& DoubleVal);
/// @brief Read an arbitrary data chunk of fixed length
inline void read_data(void *Ptr, void *End);
/// @brief Read a bytecode block header
inline void read_block(unsigned &Type, unsigned &Size);
/// @brief Read a type identifier and sanitize it.
inline bool read_typeid(unsigned &TypeId);
/// @brief Recalculate type ID for pre 1.3 bytecode files.
inline bool sanitizeTypeId(unsigned &TypeId );
/// @}
};
/// @brief A function for creating a BytecodeAnalzer as a handler
/// for the Bytecode reader.
BytecodeHandler* createBytecodeAnalyzerHandler(BytecodeAnalysis& bca,
std::ostream* output );
} // End llvm namespace
// vim: sw=2
#endif