This document describes the LLVM bytecode file format. It specifies the binary encoding rules of the bytecode file format so that equivalent systems can encode bytecode files correctly. The LLVM bytecode representation is used to store the intermediate representation on disk in compacted form.

The LLVM bytecode format may change in the future, but LLVM will always be backwards compatible with older formats. This document will only describe the most current version of the bytecode format. See Version Differences for the details on how the current version is different from previous versions.

This section describes the general concepts of the bytecode file format without getting into specific layout details. It is recommended that you read this section thoroughly before interpreting the detailed descriptions.

LLVM bytecode files consist simply of a sequence of blocks of bytes using a binary encoding Each block begins with an header of two unsigned integers. The first value identifies the type of block and the second value provides the size of the block in bytes. The block identifier is used because it is possible for entire blocks to be omitted from the file if they are empty. The block identifier helps the reader determine which kind of block is next in the file. Note that blocks can be nested within other blocks.

All blocks are variable length, and the block header specifies the size of the block. All blocks begin on a byte index that is aligned to an even 32-bit boundary. That is, the first block is 32-bit aligned because it starts at offset 0. Each block is padded with zero fill bytes to ensure that the next block also starts on a 32-bit boundary.

LLVM Bytecode blocks often contain lists of things of a similar type. For example, a function contains a list of instructions and a function type contains a list of argument types. There are two basic types of lists: length lists (llist), and null terminated lists (zlist), as described below in the Encoding Primitives.

Fields are units of information that LLVM knows how to write atomically. Most fields have a uniform length or some kind of length indication built into their encoding. For example, a constant string (array of bytes) is written simply as the length followed by the characters. Although this is similar to a list, constant strings are treated atomically and are thus fields.

Fields use a condensed bit format specific to the type of information they must contain. As few bits as possible are written for each field. The sections that follow will provide the details on how these fields are written and how the bits are to be interpreted.

To support cross-platform differences, the bytecode file is aligned on certain boundaries. This means that a small amount of padding (at most 3 bytes) will be added to ensure that the next entry is aligned to a 32-bit boundary.

Most of the values written to LLVM bytecode files are small integers. To minimize the number of bytes written for these quantities, an encoding scheme similar to UTF-8 is used to write integer data. The scheme is known as variable bit rate (vbr) encoding. In this encoding, the high bit of each byte is used to indicate if more bytes follow. If (byte & 0x80) is non-zero in any given byte, it means there is another byte immediately following that also contributes to the value. For the final byte (byte & 0x80) is false (the high bit is not set). In each byte only the low seven bits contribute to the value. Consequently 32-bit quantities can take from one to five bytes to encode. In general, smaller quantities will encode in fewer bytes, as follows:

Note that in practice, the tenth byte could only encode bit 63 since the maximum quantity to use this encoding is a 64-bit integer.

Signed VBR values are encoded with the standard vbr encoding, but with the sign bit as the low order bit instead of the high order bit. This allows small negative quantities to be encoded efficiently. For example, -3 is encoded as "((3 << 1) | 1)" and 3 is encoded as "(3 << 1) | 0)", emitted with the standard vbr encoding above.

Each field in the bytecode format is encoded into the file using a small set of primitive formats. The table below defines the encoding rules for the various primitives used and gives them each a type name. The type names used in the descriptions of blocks and fields in the Detailed Layoutnext section. Any type name with the suffix _vbr indicates a quantity that is encoded using variable bit rate encoding as described above.

In the detailed block and field descriptions that follow, a regex like notation is used to describe optional and repeated fields. A very limited subset of regex is used to describe these, as given in the following table:

So, for example, consider the following specifications:

with the following interpretations:

1. An optional string. Matches either nothing or a single string
2. One or more pairs of uint32_vbr.
3. Zero or more occurrences of either an unsigned followed by a uint32_vbr or just a uint32_vbr.
4. An optional length list of unsigned values.

The bytecode format uses the notion of a "slot" to reference Types and Values. Since the bytecode file is a direct representation of LLVM's intermediate representation, there is a need to represent pointers in the file. Slots are used for this purpose. For example, if one has the following assembly:

there are two definitions. The definition of %MyVar uses %MyType. In the C++ IR this linkage between %MyVar and %MyType is explicit through the use of C++ pointers. In bytecode, however, there's no ability to store memory addresses. Instead, we compute and write out slot numbers for every Type and Value written to the file.

A slot number is simply an unsigned 32-bit integer encoded in the variable bit rate scheme (see encoding). This ensures that low slot numbers are encoded in one byte. Through various bits of magic LLVM attempts to always keep the slot numbers low. The first attempt is to associate slot numbers with their "type plane". That is, Values of the same type are written to the bytecode file in a list (sequentially). Their order in that list determines their slot number. This means that slot #1 doesn't mean anything unless you also specify for which type you want slot #1. Types are always written to the file first (in the Global Type Pool) and in such a way that both forward and backward references of the types can often be resolved with a single pass through the type pool.

Slot numbers are also kept small by rearranging their order. Because of the structure of LLVM, certain values are much more likely to be used frequently in the body of a function. For this reason, a compaction table is provided in the body of a function if its use would make the function body smaller. Suppose you have a function body that uses just the types "int*" and "{double}" but uses them thousands of time. Its worthwhile to ensure that the slot number for these types are low so they can be encoded in a single byte (via vbr). This is exactly what the compaction table does.

In summary then, a slot number can be though of as just a vbr encoded index into a list of Type* or Value*. To keep slot numbers low, Value* are indexed by two slot numbers: the "type plane index" (type slot) and the "value index" (value slot).

This section provides the general structure of the LLVM bytecode file format. The bytecode file format requires blocks to be in a certain order and nested in a particular way so that an LLVM module can be constructed efficiently from the contents of the file. This ordering defines a general structure for bytecode files as shown below. The table below shows the order in which all block types may appear. Please note that some of the blocks are optional and some may be repeated. The structure is fairly loose because optional blocks, if empty, are completely omitted from the file.

Use the links in the table for details about the contents of each of the block types.

This section provides the detailed layout of the individual block types in the LLVM bytecode file format.

The signature occurs in every LLVM bytecode file and is always first. It simply provides a few bytes of data to identify the file as being an LLVM bytecode file. This block is always four bytes in length and differs from the other blocks because there is no identifier and no block length at the start of the block. Essentially, this block is just the "magic number" for the file.

The module block contains a small pre-amble and all the other blocks in the file. The table below shows the structure of the module block. Note that it only provides the module identifier, size of the module block, and the format information. Everything else is contained in other blocks, described in other sections.

The block header for the module block uses a longer format than the other blocks in a bytecode file. Specifically, instead of encoding the type and size of the block into a 32-bit integer with 5-bits for type and 27-bits for size, the module block header uses two 32-bit unsigned values, one for type, and one for size. While the 2²⁷ byte limit on block size is sufficient for the blocks contained in the module, it isn't sufficient for the module block itself because we want to ensure that bytecode files as large as 2³² bytes are possible. For this reason, the module block (and only the module block) uses a long format header.

The format information field is encoded into a uint32_vbr as shown in the following table.

Of particular note, the bytecode format number is simply a 28-bit monotonically increase integer that identifies the version of the bytecode format (which is not directly related to the LLVM release number). The bytecode versions defined so far are (note that this document only describes the latest version, 1.3):

• #0: LLVM 1.0 & 1.1
• #1: LLVM 1.2
• #2: LLVM 1.2.5 (not released)
• #3: LLVM 1.3
  

Note that we plan to eventually expand the target description capabilities of bytecode files to target triples.

The global type pool consists of type definitions. Their order of appearance in the file determines their type slot number (0 based). Slot numbers are used to replace pointers in the intermediate representation. Each slot number uniquely identifies one entry in a type plane (a collection of values of the same type). Since all values have types and are associated with the order in which the type pool is written, the global type pool must be written as the first block of a module. If it is not, attempts to read the file will fail because both forward and backward type resolution will not be possible.

The type pool is simply a list of type definitions, as shown in the table below.

Types in the type pool are defined using a different format for each kind of type, as given in the following sections.

Primitive Types

The primitive types encompass the basic integer and floating point types. They are encoded simply as their TypeID.

1. The values for the Type IDs for the primitive types are provided by the definition of the llvm::Type::TypeID enumeration in include/llvm/Type.h. The enumeration gives the following mapping:
   1. bool
   2. ubyte
   3. sbyte
   4. ushort
   5. short
   6. uint
   7. int
   8. ulong
   9. long
   10. float
   11. double

Function Types

Structure Types

Array Types

Pointer Types

Opaque Types

Packed Types

The module global info block contains the definitions of all global variables including their initializers and the declaration of all functions. The format is shown in the table below:

Global variables are written using an uint32_vbr that encodes information about the global variable and a list of the constant initializers for the global var, if any.

The table below provides the bit layout of the first uint32_vbr that describes the global variable.

The table below provides the format of the constant initializers for the global variable field, if it has one.

A constant pool defines as set of constant values. There are actually two types of constant pool blocks: one for modules and one for functions. For modules, the block begins with the constant strings encountered anywhere in the module. For functions, the block begins with types only encountered in the function. In both cases the header is identical. The tables that follow, show the header, module constant pool preamble, function constant pool preamble, and the part common to both function and module constant pools.

Common Block Header

Module Constant Pool Preamble (constant strings)

Function Constant Pool Preamble (function types)

The structure of the types for functions is identical to the Global Type Pool. Please refer to that section for the details.

Common Part (other constants)

Constants come in many shapes and flavors. The sections that follow define the format for each of them. All constants start with a uint32_vbr encoded integer that provides the number of operands for the constant. For primitive, structure, and array constants, this will always be zero since those types of constants have no operands. In this case, we have the following field definitions:

• Bool. This is written as an uint32_vbr of value 1U or 0U.
• Signed Integers (sbyte,short,int,long). These are written as an int64_vbr with the corresponding value.
• Unsigned Integers (ubyte,ushort,uint,ulong). These are written as an uint64_vbr with the corresponding value.
• Floating Point. Both the float and double types are written literally in binary format.
• Arrays. Arrays are written simply as a list of uint32_vbr encoded value slot numbers to the constant element values.
• Structures. Structures are written simply as a list of uint32_vbr encoded value slot numbers to the constant field values of the structure.

When the number of operands to the constant is non-zero, we have a constant expression and its field format is provided in the table below.

1. Both these fields are repeatable but only in pairs.

Function definitions contain the linkage, constant pool or compaction table, instruction list, and symbol table for a function. The following table shows the structure of a function definition.

1. Note that if the linkage type is "External" then none of the other fields will be present as the function is defined elsewhere.
2. Note that only one of the constant pool or compaction table will be written. Compaction tables are only written if they will actually save bytecode space. If not, then a regular constant pool is written.

Compaction tables are part of a function definition. They are merely a device for reducing the size of bytecode files. The size of a bytecode file is dependent on the values of the slot numbers used because larger values use more bytes in the variable bit rate encoding scheme. Furthermore, the compressed instruction format reserves only six bits for the type of the instruction. In large modules, declaring hundreds or thousands of types, the values of the slot numbers can be quite large. However, functions may use only a small fraction of the global types. In such cases a compaction table is created that maps the global type and value slot numbers to smaller values used by a function. Functions will contain either a function-specific constant pool or a compaction table but not both. Compaction tables have the format shown in the table below.

The type and length of a compaction table type plane is encoded differently depending on the length of the plane. For planes of length 1 or 2, the length is encoded into bits 0 and 1 of a uint32_vbr and the type is encoded into bits 2-31. Because type numbers are often small, this often saves an extra byte per plane. If the length of the plane is greater than 2 then the encoding uses a uint32_vbr for each of the length and type, in that order.

The instructions in a function are written as a simple list. Basic blocks are inferred by the terminating instruction types. The format of the block is given in the following table.

For brevity, instructions are written in one of four formats, depending on the number of operands to the instruction. Each instruction begins with a uint32_vbr that encodes the type of the instruction as well as other things. The tables that follow describe the format of this first part of each instruction.

Instruction Format 0

This format is used for a few instructions that can't easily be shortened because they have large numbers of operands (e.g. PHI Node or getelementptr). Each of the opcode, type, and operand fields is found in successive fields.

1. Note that if the instruction is a getelementptr and the type of the operand is a sequential type (array or pointer) then the slot number is shifted up two bits and the low order bits will encode the type of index used, as follows: 0=uint, 1=int, 2=ulong, 3=long.

Instruction Format 1

This format encodes the opcode, type and a single operand into a single uint32_vbr as follows:

Instruction Format 2

This format encodes the opcode, type and two operands into a single uint32_vbr as follows:

Instruction Format 3

This format encodes the opcode, type and three operands into a single uint32_vbr as follows:

Instructions encode an opcode that identifies the kind of instruction. Opcodes are an enumerated integer value. The specific values used depend on the version of LLVM you're using. The opcode values are defined in the include/llvm/Instruction.def file. You should check there for the most recent definitions. The table below provides the opcodes defined as of the writing of this document. The table associates each opcode mnemonic with its enumeration value and the bytecode and LLVM version numbers in which the opcode was introduced.

A symbol table can be put out in conjunction with a module or a function. A symbol table has a list of name/type associations followed by a list of name/value associations. The name/value associations are organized into "type planes" so that all values of a common type are listed together. Each type plane starts with the number of entries in the plane and the type slot number for all the values in that plane (so the type can be looked up in the global type pool). For each entry in a type plane, the slot number of the value and the name associated with that value are written. The format is given in the table below.

A symbol table type entry associates a name with a type. The name is provided simply as an array of chars. The type is provided as a type slot number (index) into the global type pool. The format is given in the following table:

A symbol table plane provides the symbol table entries for all values of a common type. The encoding is given in the following table:

A symbol table value entry provides the assocation between a value and the name given to the value. The value is referenced by its slot number. The format is given in the following table:

This section describes the differences in the Bytecode Format across LLVM versions. The versions are listed in reverse order because it assumes the current version is as documented in the previous sections. Each section here describes the differences between that version and the one that follows.

In version 1.2, the Type class in the LLVM IR derives from the Value class. This is not the case in version 1.3. Consequently, in version 1.2 the notion of a "Type Type" was used to write out values that were Types. The types always occuped plane 12 (corresponding to the TypeTyID) of any type planed set of values. In 1.3 this representation is not convenient because the TypeTyID (12) is not present and its value is now used for LabelTyID. Consequently, the data structures written that involve types do so by writing all the types first and then each of the value planes according to those types. In version 1.2, the types would have been written intermingled with the values.

In version 1.2, the getelementptr instruction required a ubyte type index for accessing a structure field and a long type index for accessing an array element. Consequently, it was only possible to access structures of 255 or fewer elements. Starting in version 1.3, this restriction was lifted. Structures must now be indexed with uint constants. Arrays may now be indexed with int, uint, long, or ulong typed values. The consequence of this was that the bytecode format had to change in order to accommodate the larger range of structure indices.

In version 1.2, block headers were always 8 bytes being comprised of both an unsigned integer type and an unsigned integer size. For very small modules, these block headers turn out to be a large fraction of the total bytecode file size. In an attempt to make these small files smaller, the type and size information was encoded into a single unsigned integer (4 bytes) comprised of 5 bits for the block type (maximum 31 block types) and 27 bits for the block size (max ~134MBytes). These limits seemed sufficient for any blocks or sizes forseen in the future. Note that the module block, which encloses all the other blocks is still written as 8 bytes since bytecode files larger than 134MBytes might be possible.

In version 1.2, the bytecode format does not store module's target triple or dependent. These fields have been added to the end of the module global info block. The purpose of these fields is to allow a front end compiler to specifiy that the generated module is specific to a particular target triple (operating system/manufacturer/processor) which makes it non-portable; and to allow front end compilers to specify the list of libraries that the module depends on for successful linking.

In version 1.2, type slot identifiers were written as 32-bit VBR quantities. In 1.3 this has been reduced to 24-bits in order to ensure that it is not possible to overflow the type field of a global variable definition. 24-bits for type slot numbers is deemed sufficient for any practical use of LLVM.

In version 1.1, the zero value for primitives was explicitly encoded into the bytecode format. Since these zero values are constant values in the LLVM IR and never change, there is no reason to explicitly encode them. This explicit encoding was removed in version 1.2.

In version 1.1, the Module Global Info block was not aligned causing the next block to be read in on an unaligned boundary. This problem was corrected in version 1.2.

None. Version 1.0 and 1.1 bytecode formats are identical.