This document describes the LLVM bitstream file format and the encoding of the LLVM IR into it.
What is commonly known as the LLVM bitcode file format (also, sometimes anachronistically known as bytecode) is actually two things: a bitstream container format and an encoding of LLVM IR into the container format.
The bitstream format is an abstract encoding of structured data, very similar to XML in some ways. Like XML, bitstream files contain tags, and nested structures, and you can parse the file without having to understand the tags. Unlike XML, the bitstream format is a binary encoding, and unlike XML it provides a mechanism for the file to self-describe "abbreviations", which are effectively size optimizations for the content.
LLVM IR files may be optionally embedded into a wrapper structure that makes it easy to embed extra data along with LLVM IR files.
This document first describes the LLVM bitstream format, describes the wrapper format, then describes the record structure used by LLVM IR files.
The bitstream format is literally a stream of bits, with a very simple structure. This structure consists of the following concepts:
Note that the llvm-bcanalyzer tool can be used to dump and inspect arbitrary bitstreams, which is very useful for understanding the encoding.
The first two bytes of a bitcode file are 'BC' (0x42, 0x43). The second two bytes are an application-specific magic number. Generic bitcode tools can look at only the first two bytes to verify the file is bitcode, while application-specific programs will want to look at all four.
A bitstream literally consists of a stream of bits, which are read in order starting with the least significant bit of each byte. The stream is made up of a number of primitive values that encode a stream of unsigned integer values. These integers are encoded in two ways: either as Fixed Width Integers or as Variable Width Integers.
Fixed-width integer values have their low bits emitted directly to the file. For example, a 3-bit integer value encodes 1 as 001. Fixed width integers are used when there are a well-known number of options for a field. For example, boolean values are usually encoded with a 1-bit wide integer.
Variable-width integer (VBR) values encode values of arbitrary size, optimizing for the case where the values are small. Given a 4-bit VBR field, any 3-bit value (0 through 7) is encoded directly, with the high bit set to zero. Values larger than N-1 bits emit their bits in a series of N-1 bit chunks, where all but the last set the high bit.
For example, the value 27 (0x1B) is encoded as 1011 0011 when emitted as a vbr4 value. The first set of four bits indicates the value 3 (011) with a continuation piece (indicated by a high bit of 1). The next word indicates a value of 24 (011 << 3) with no continuation. The sum (3+24) yields the value 27.
6-bit characters encode common characters into a fixed 6-bit field. They represent the following characters with the following 6-bit values:
'a' .. 'z' — 0 .. 25 'A' .. 'Z' — 26 .. 51 '0' .. '9' — 52 .. 61 '.' — 62 '_' — 63
This encoding is only suitable for encoding characters and strings that consist only of the above characters. It is completely incapable of encoding characters not in the set.
Occasionally, it is useful to emit zero bits until the bitstream is a multiple of 32 bits. This ensures that the bit position in the stream can be represented as a multiple of 32-bit words.
A bitstream is a sequential series of Blocks and Data Records. Both of these start with an abbreviation ID encoded as a fixed-bitwidth field. The width is specified by the current block, as described below. The value of the abbreviation ID specifies either a builtin ID (which have special meanings, defined below) or one of the abbreviation IDs defined for the current block by the stream itself.
The set of builtin abbrev IDs is:
Abbreviation IDs 4 and above are defined by the stream itself, and specify an abbreviated record encoding.
Blocks in a bitstream denote nested regions of the stream, and are identified by a content-specific id number (for example, LLVM IR uses an ID of 12 to represent function bodies). Block IDs 0-7 are reserved for standard blocks whose meaning is defined by Bitcode; block IDs 8 and greater are application specific. Nested blocks capture the hierarchical structure of the data encoded in it, and various properties are associated with blocks as the file is parsed. Block definitions allow the reader to efficiently skip blocks in constant time if the reader wants a summary of blocks, or if it wants to efficiently skip data it does not understand. The LLVM IR reader uses this mechanism to skip function bodies, lazily reading them on demand.
When reading and encoding the stream, several properties are maintained for the block. In particular, each block maintains:
As sub blocks are entered, these properties are saved and the new sub-block has its own set of abbreviations, and its own abbrev id width. When a sub-block is popped, the saved values are restored.
[ENTER_SUBBLOCK, blockidvbr8, newabbrevlenvbr4, <align32bits>, blocklen32]
The ENTER_SUBBLOCK abbreviation ID specifies the start of a new block record. The blockid value is encoded as an 8-bit VBR identifier, and indicates the type of block being entered, which can be a standard block or an application-specific block. The newabbrevlen value is a 4-bit VBR, which specifies the abbrev id width for the sub-block. The blocklen value is a 32-bit aligned value that specifies the size of the subblock in 32-bit words. This value allows the reader to skip over the entire block in one jump.
The END_BLOCK abbreviation ID specifies the end of the current block record. Its end is aligned to 32-bits to ensure that the size of the block is an even multiple of 32-bits.
Data records consist of a record code and a number of (up to) 64-bit integer values. The interpretation of the code and values is application specific and may vary between different block types. Records can be encoded either using an unabbrev record, or with an abbreviation. In the LLVM IR format, for example, there is a record which encodes the target triple of a module. The code is MODULE_CODE_TRIPLE, and the values of the record are the ASCII codes for the characters in the string.
[UNABBREV_RECORD, codevbr6, numopsvbr6, op0vbr6, op1vbr6, ...]
An UNABBREV_RECORD provides a default fallback encoding, which is both completely general and extremely inefficient. It can describe an arbitrary record by emitting the code and operands as VBRs.
For example, emitting an LLVM IR target triple as an unabbreviated record requires emitting the UNABBREV_RECORD abbrevid, a vbr6 for the MODULE_CODE_TRIPLE code, a vbr6 for the length of the string, which is equal to the number of operands, and a vbr6 for each character. Because there are no letters with values less than 32, each letter would need to be emitted as at least a two-part VBR, which means that each letter would require at least 12 bits. This is not an efficient encoding, but it is fully general.
An abbreviated record is a abbreviation id followed by a set of fields that are encoded according to the abbreviation definition. This allows records to be encoded significantly more densely than records encoded with the UNABBREV_RECORD type, and allows the abbreviation types to be specified in the stream itself, which allows the files to be completely self describing. The actual encoding of abbreviations is defined below.
The record code, which is the first field of an abbreviated record, may be encoded in the abbreviation definition (as a literal operand) or supplied in the abbreviated record (as a Fixed or VBR operand value).
Abbreviations are an important form of compression for bitstreams. The idea is to specify a dense encoding for a class of records once, then use that encoding to emit many records. It takes space to emit the encoding into the file, but the space is recouped (hopefully plus some) when the records that use it are emitted.
Abbreviations can be determined dynamically per client, per file. Because the abbreviations are stored in the bitstream itself, different streams of the same format can contain different sets of abbreviations according to the needs of the specific stream. As a concrete example, LLVM IR files usually emit an abbreviation for binary operators. If a specific LLVM module contained no or few binary operators, the abbreviation does not need to be emitted.
[DEFINE_ABBREV, numabbrevopsvbr5, abbrevop0, abbrevop1, ...]
A DEFINE_ABBREV record adds an abbreviation to the list of currently defined abbreviations in the scope of this block. This definition only exists inside this immediate block — it is not visible in subblocks or enclosing blocks. Abbreviations are implicitly assigned IDs sequentially starting from 4 (the first application-defined abbreviation ID). Any abbreviations defined in a BLOCKINFO record for the particular block type receive IDs first, in order, followed by any abbreviations defined within the block itself. Abbreviated data records reference this ID to indicate what abbreviation they are invoking.
An abbreviation definition consists of the DEFINE_ABBREV abbrevid followed by a VBR that specifies the number of abbrev operands, then the abbrev operands themselves. Abbreviation operands come in three forms. They all start with a single bit that indicates whether the abbrev operand is a literal operand (when the bit is 1) or an encoding operand (when the bit is 0).
The possible operand encodings are:
For example, target triples in LLVM modules are encoded as a record of the form [TRIPLE, 'a', 'b', 'c', 'd']. Consider if the bitstream emitted the following abbrev entry:
[0, Fixed, 4] [0, Array] [0, Char6]
When emitting a record with this abbreviation, the above entry would be emitted as:
[4abbrevwidth, 24, 4vbr6, 06, 16, 26, 36]
These values are:
With this abbreviation, the triple is emitted with only 37 bits (assuming a abbrev id width of 3). Without the abbreviation, significantly more space would be required to emit the target triple. Also, because the TRIPLE value is not emitted as a literal in the abbreviation, the abbreviation can also be used for any other string value.
In addition to the basic block structure and record encodings, the bitstream also defines specific built-in block types. These block types specify how the stream is to be decoded or other metadata. In the future, new standard blocks may be added. Block IDs 0-7 are reserved for standard blocks.
The BLOCKINFO block allows the description of metadata for other blocks. The currently specified records are:
[SETBID (#1), blockid] [DEFINE_ABBREV, ...] [BLOCKNAME, ...name...] [SETRECORDNAME, RecordID, ...name...]
The SETBID record (code 1) indicates which block ID is being described. SETBID records can occur multiple times throughout the block to change which block ID is being described. There must be a SETBID record prior to any other records.
Standard DEFINE_ABBREV records can occur inside BLOCKINFO blocks, but unlike their occurrence in normal blocks, the abbreviation is defined for blocks matching the block ID we are describing, not the BLOCKINFO block itself. The abbreviations defined in BLOCKINFO blocks receive abbreviation IDs as described in DEFINE_ABBREV.
The BLOCKNAME record (code 2) can optionally occur in this block. The elements of the record are the bytes of the string name of the block. llvm-bcanalyzer can use this to dump out bitcode files symbolically.
The SETRECORDNAME record (code 3) can also optionally occur in this block. The first operand value is a record ID number, and the rest of the elements of the record are the bytes for the string name of the record. llvm-bcanalyzer can use this to dump out bitcode files symbolically.
Note that although the data in BLOCKINFO blocks is described as "metadata," the abbreviations they contain are essential for parsing records from the corresponding blocks. It is not safe to skip them.
Bitcode files for LLVM IR may optionally be wrapped in a simple wrapper structure. This structure contains a simple header that indicates the offset and size of the embedded BC file. This allows additional information to be stored alongside the BC file. The structure of this file header is:
[Magic32, Version32, Offset32, Size32, CPUType32]
Each of the fields are 32-bit fields stored in little endian form (as with the rest of the bitcode file fields). The Magic number is always 0x0B17C0DE and the version is currently always 0. The Offset field is the offset in bytes to the start of the bitcode stream in the file, and the Size field is the size in bytes of the stream. CPUType is a target-specific value that can be used to encode the CPU of the target.
LLVM IR is encoded into a bitstream by defining blocks and records. It uses blocks for things like constant pools, functions, symbol tables, etc. It uses records for things like instructions, global variable descriptors, type descriptions, etc. This document does not describe the set of abbreviations that the writer uses, as these are fully self-described in the file, and the reader is not allowed to build in any knowledge of this.
The magic number for LLVM IR files is:
[0x04, 0xC4, 0xE4, 0xD4]
When combined with the bitcode magic number and viewed as bytes, this is "BC 0xC0DE".
Variable Width Integer encoding is an efficient way to encode arbitrary sized unsigned values, but is an extremely inefficient for encoding signed values, as signed values are otherwise treated as maximally large unsigned values.
As such, signed VBR values of a specific width are emitted as follows:
With this encoding, small positive and small negative values can both be emitted efficiently. Signed VBR encoding is used in CST_CODE_INTEGER and CST_CODE_WIDE_INTEGER records within CONSTANTS_BLOCK blocks.
LLVM IR is defined with the following blocks:
The MODULE_BLOCK block (id 8) is the top-level block for LLVM bitcode files, and each bitcode file must contain exactly one. In addition to records (described below) containing information about the module, a MODULE_BLOCK block may contain the following sub-blocks:
The VERSION record (code 1) contains a single value indicating the format version. Only version 0 is supported at this time.
The TRIPLE record (code 2) contains a variable number of values representing the bytes of the target triple specification string.
The DATALAYOUT record (code 3) contains a variable number of values representing the bytes of the target datalayout specification string.
The ASM record (code 4) contains a variable number of values representing the bytes of module asm strings, with individual assembly blocks separated by newline (ASCII 10) characters.
The SECTIONNAME record (code 5) contains a variable number of values representing the bytes of a single section name string. There should be one SECTIONNAME record for each section name referenced (e.g., in global variable or function section attributes) within the module. These records can be referenced by the 1-based index in the section fields of GLOBALVAR or FUNCTION records.
The DEPLIB record (code 6) contains a variable number of values representing the bytes of a single dependent library name string, one of the libraries mentioned in a deplibs declaration. There should be one DEPLIB record for each library name referenced.
[GLOBALVAR, pointer type, isconst, initid, linkage, alignment, section, visibility, threadlocal]
The GLOBALVAR record (code 7) marks the declaration or definition of a global variable. The operand fields are:
[FUNCTION, type, callingconv, isproto, linkage, paramattr, alignment, section, visibility, gc]
The FUNCTION record (code 8) marks the declaration or definition of a function. The operand fields are:
[ALIAS, alias type, aliasee val#, linkage, visibility]
The ALIAS record (code 9) marks the definition of an alias. The operand fields are
The PURGEVALS record (code 10) resets the module-level value list to the size given by the single operand value. Module-level value list items are added by GLOBALVAR, FUNCTION, and ALIAS records. After a PURGEVALS record is seen, new value indices will start from the given numvals value.
The GCNAME record (code 11) contains a variable number of values representing the bytes of a single garbage collector name string. There should be one GCNAME record for each garbage collector name referenced in function gc attributes within the module. These records can be referenced by 1-based index in the gc fields of FUNCTION records.
The PARAMATTR_BLOCK block (id 9) contains a table of entries describing the attributes of function parameters. These entries are referenced by 1-based index in the paramattr field of module block FUNCTION records, or within the attr field of function block INST_INVOKE and INST_CALL records.
Entries within PARAMATTR_BLOCK are constructed to ensure that each is unique (i.e., no two indicies represent equivalent attribute lists).
[ENTRY, paramidx0, attr0, paramidx1, attr1...]
The ENTRY record (code 1) contains an even number of values describing a unique set of function parameter attributes. Each paramidx value indicates which set of attributes is represented, with 0 representing the return value attributes, 0xFFFFFFFF representing function attributes, and other values representing 1-based function parameters. Each attr value is a bitmap with the following interpretation:
The TYPE_BLOCK block (id 10) contains records which constitute a table of type operator entries used to represent types referenced within an LLVM module. Each record (with the exception of NUMENTRY) generates a single type table entry, which may be referenced by 0-based index from instructions, constants, metadata, type symbol table entries, or other type operator records.
Entries within TYPE_BLOCK are constructed to ensure that each entry is unique (i.e., no two indicies represent structurally equivalent types).
The NUMENTRY record (code 1) contains a single value which indicates the total number of type code entries in the type table of the module. If present, NUMENTRY should be the first record in the block.
The VOID record (code 2) adds a void type to the type table.
The FLOAT record (code 3) adds a float (32-bit floating point) type to the type table.
The DOUBLE record (code 4) adds a double (64-bit floating point) type to the type table.
The LABEL record (code 5) adds a label type to the type table.
The OPAQUE record (code 6) adds an opaque type to the type table. Note that distinct opaque types are not unified.
The INTEGER record (code 7) adds an integer type to the type table. The single width field indicates the width of the integer type.
[POINTER, pointee type, address space]
The POINTER record (code 8) adds a pointer type to the type table. The operand fields are
[FUNCTION, vararg, ignored, retty, ...paramty... ]
The FUNCTION record (code 9) adds a function type to the type table. The operand fields are
[STRUCT, ispacked, ...eltty...]
The STRUCT record (code 10) adds a struct type to the type table. The operand fields are
[ARRAY, numelts, eltty]
The ARRAY record (code 11) adds an array type to the type table. The operand fields are
[VECTOR, numelts, eltty]
The VECTOR record (code 12) adds a vector type to the type table. The operand fields are
The X86_FP80 record (code 13) adds an x86_fp80 (80-bit floating point) type to the type table.
The FP128 record (code 14) adds an fp128 (128-bit floating point) type to the type table.
The PPC_FP128 record (code 15) adds a ppc_fp128 (128-bit floating point) type to the type table.
The METADATA record (code 16) adds a metadata type to the type table.
The CONSTANTS_BLOCK block (id 11) ...
The FUNCTION_BLOCK block (id 12) ...
In addition to the record types described below, a FUNCTION_BLOCK block may contain the following sub-blocks:
The TYPE_SYMTAB_BLOCK block (id 13) contains entries which map between module-level named types and their corresponding type indices.
[ENTRY, typeid, ...string...]
The ENTRY record (code 1) contains a variable number of values, with the first giving the type index of the designated type, and the remaining values giving the character codes of the type name. Each entry corresponds to a single named type.
The VALUE_SYMTAB_BLOCK block (id 14) ...
The METADATA_BLOCK block (id 15) ...
The METADATA_ATTACHMENT block (id 16) ...