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.
- bool processFunctions=false ///< Process all function bodies fully.
+ const std::string &ModuleID ///< An identifier for the module constructed.
);
/// @brief Parse all function bodies
void ParseFunctionBody(Function* Func);
/// @brief Parse the type list portion of a compaction table
- void BytecodeReader::ParseCompactionTypes( unsigned NumEntries );
+ void ParseCompactionTypes(unsigned NumEntries);
/// @brief Parse a compaction table
void ParseCompactionTable();
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
+ /// 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
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
+ /// 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 floating point values in a platform specific
- /// bit ordering. This was fixed in LLVM 1.3
- bool hasPlatformSpecificFloatingPoint;
-
/// 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
/// features, for use in future versions of LLVM.
bool hasNoDependentLibraries;
- /// LLVM 1.2 and earlier encoded the file version as part of the module block
- /// but this information may be needed to
+ /// 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;
- /// CompactionTable - If a compaction table is active in the current function,
- /// this is the mapping that it contains.
- std::vector<const Type*> CompactionTypes;
+ // 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.
/// @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(const Type *Ty, unsigned SlotNo);
+ /// @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 A function for creating a BytecodeAnalzer as a handler
/// for the Bytecode reader.
-BytecodeHandler* createBytecodeAnalyzerHandler(BytecodeAnalysis& bca );
+BytecodeHandler* createBytecodeAnalyzerHandler(BytecodeAnalysis& bca,
+ std::ostream* output );
} // End llvm namespace