This documents describes the MLIR bytecode format and its encoding.
MLIR uses the following four-byte magic number to indicate bytecode files:
‘[‘M’8, ‘L’8, ‘ï’8, ‘R’8]’
In hex:
‘[‘4D’8, ‘4C’8, ‘EF’8, ‘52’8]’
An MLIR Bytecode file is comprised of a byte stream, with a few simple structural concepts layered on top.
byte ::= `0x00`...`0xFF`
Fixed width integers are unsigned integers of a known byte size. The values are stored in little-endian byte order.
TODO: Add larger fixed width integers as necessary.
Variable width integers, or VarInts, provide a compact representation for integers. Each encoded VarInt consists of one to nine bytes, which together represent a single 64-bit value. The MLIR bytecode utilizes the “PrefixVarInt” encoding for VarInts. This encoding is a variant of the LEB128 (“Little-Endian Base 128”) encoding, where each byte of the encoding provides up to 7 bits for the value, with the remaining bit used to store a tag indicating the number of bytes used for the encoding. This means that small unsigned integers (less than 2^7) may be stored in one byte, unsigned integers up to 2^14 may be stored in two bytes, etc.
The first byte of the encoding includes a length prefix in the low bits. This prefix is a bit sequence of ‘0’s followed by a terminal ‘1’, or the end of the byte. The number of ‘0’ bits indicate the number of additional bytes, not including the prefix byte, used to encode the value. All of the remaining bits in the first byte, along with all of the bits in the additional bytes, provide the value of the integer. Below are the various possible encodings of the prefix byte:
xxxxxxx1: 7 value bits, the encoding uses 1 byte xxxxxx10: 14 value bits, the encoding uses 2 bytes xxxxx100: 21 value bits, the encoding uses 3 bytes xxxx1000: 28 value bits, the encoding uses 4 bytes xxx10000: 35 value bits, the encoding uses 5 bytes xx100000: 42 value bits, the encoding uses 6 bytes x1000000: 49 value bits, the encoding uses 7 bytes 10000000: 56 value bits, the encoding uses 8 bytes 00000000: 64 value bits, the encoding uses 9 bytes
Signed variable width integer values are encoded in a similar fashion to varints, but employ zigzag encoding. This encoding uses the low bit of the value to indicate the sign, which allows for more efficiently encoding negative numbers. If a negative value were encoded using a normal varint, it would be treated as an extremely large unsigned value. Using zigzag encoding allows for a smaller number of active bits in the value, leading to a smaller encoding. Below is the basic computation for generating a zigzag encoding:
(value << 1) ^ (value >> 63)
Strings are blobs of characters with an associated length.
section {
idAndIsAligned: byte // id | (hasAlign << 7)
length: varint,
alignment: varint?,
padding: byte[], // Padding bytes are always `0xCB`.
data: byte[]
}
Sections are a mechanism for grouping data within the bytecode. They enable delayed processing, which is useful for out-of-order processing of data, lazy-loading, and more. Each section contains a Section ID, whose high bit indicates if the section has alignment requirements, a length (which allows for skipping over the section), and an optional alignment. When an alignment is present, a variable number of padding bytes (0xCB) may appear before the section data. The alignment of a section must be a power of 2.
Given the generic structure of MLIR, the bytecode encoding is actually fairly simplistic. It effectively maps to the core components of MLIR.
The top-level structure of the bytecode contains the 4-byte “magic number”, a version number, a null-terminated producer string, and a list of sections. Each section is currently only expected to appear once within a bytecode file.
bytecode {
magic: "MLïR",
version: varint,
producer: string,
sections: section[]
}
strings {
numStrings: varint,
reverseStringLengths: varint[],
stringData: byte[]
}
The string section contains a table of strings referenced within the bytecode, more easily enabling string sharing. This section is encoded first with the total number of strings, followed by the sizes of each of the individual strings in reverse order. The remaining encoding contains a single blob containing all of the strings concatenated together.
The dialect section of the bytecode contains all of the dialects referenced within the encoded IR, and some information about the components of those dialects that were also referenced.
dialect_section {
numDialects: varint,
dialectNames: varint[],
opNames: op_name_group[]
}
op_name_group {
dialect: varint,
numOpNames: varint,
opNames: varint[]
}
Dialects are encoded as indexes to the name string within the string section. Operation names are encoded in groups by dialect, with each group containing the dialect, the number of operation names, and the array of indexes to each name within the string section.
Attributes and types are encoded using two sections, one section (attr_type_section) containing the actual encoded representation, and another section (attr_type_offset_section) containing the offsets of each encoded attribute/type into the previous section. This structure allows for attributes and types to always be lazily loaded on demand.
attr_type_section {
attrs: attribute[],
types: type[]
}
attr_type_offset_section {
numAttrs: varint,
numTypes: varint,
offsets: attr_type_offset_group[]
}
attr_type_offset_group {
dialect: varint,
numElements: varint,
offsets: varint[] // (offset << 1) | (hasCustomEncoding)
}
attribute {
encoding: ...
}
type {
encoding: ...
}
Each offset in the attr_type_offset_section above is the size of the encoding for the attribute or type and a flag indicating if the encoding uses the textual assembly format, or a custom bytecode encoding. We avoid using the direct offset into the attr_type_section, as a smaller relative offsets provides more effective compression. Attributes and types are grouped by dialect, with each attr_type_offset_group in the offset section containing the corresponding parent dialect, number of elements, and offsets for each element within the group.
In the abstract, an attribute/type is encoded in one of two possible ways: via its assembly format, or via a custom dialect defined encoding.
In the case where a dialect does not define a method for encoding the attribute or type, the textual assembly format of that attribute or type is used as a fallback. For example, a type of !bytecode.type would be encoded as the null terminated string “!bytecode.type”. This ensures that every attribute and type may be encoded, even if the owning dialect has not yet opted in to a more efficient serialization.
TODO: We shouldn't redundantly encode the dialect name here, we should use a reference to the parent dialect instead.
In addition to the assembly format fallback, dialects may also provide a custom encoding for their attributes and types. Custom encodings are very beneficial in that they are significantly smaller and faster to read and write.
Dialects can opt-in to providing custom encodings by implementing the BytecodeDialectInterface. This interface provides hooks, namely readAttribute/readType and writeAttribute/writeType, that will be used by the bytecode reader and writer. These hooks are provided a reader and writer implementation that can be used to encode various constructs in the underlying bytecode format. A unique feature of this interface is that dialects may choose to only encode a subset of their attributes and types in a custom bytecode format, which can simplify adding new or experimental components that aren't fully baked.
When implementing the bytecode interface, dialects are responsible for all aspects of the encoding. This includes the indicator for which kind of attribute or type is being encoded; the bytecode reader will only know that it has encountered an attribute or type of a given dialect, it doesn't encode any further information. As such, a common encoding idiom is to use a leading varint code to indicate how the attribute or type was encoded.
Resources are encoded using two sections, one section (resource_section) containing the actual encoded representation, and another section (resource_offset_section) containing the offsets of each encoded resource into the previous section.
resource_section {
resources: resource[]
}
resource {
value: resource_bool | resource_string | resource_blob
}
resource_bool {
value: byte
}
resource_string {
value: varint
}
resource_blob {
alignment: varint,
size: varint,
padding: byte[],
blob: byte[]
}
resource_offset_section {
numExternalResourceGroups: varint,
resourceGroups: resource_group[]
}
resource_group {
key: varint,
numResources: varint,
resources: resource_info[]
}
resource_info {
key: varint,
size: varint
kind: byte,
}
Resources are grouped by the provider, either an external entity or a dialect, with each resource_group in the offset section containing the corresponding provider, number of elements, and info for each element within the group. For each element, we record the key, the value kind, and the encoded size. We avoid using the direct offset into the resource_section, as a smaller relative offsets provides more effective compression.
The IR section contains the encoded form of operations within the bytecode.
op {
name: varint,
encodingMask: byte,
location: varint,
attrDict: varint?,
numResults: varint?,
resultTypes: varint[],
numOperands: varint?,
operands: varint[],
numSuccessors: varint?,
successors: varint[],
regionEncoding: varint?, // (numRegions << 1) | (isIsolatedFromAbove)
regions: region[]
}
The encoding of an operation is important because this is generally the most commonly appearing structure in the bytecode. A single encoding is used for every type of operation. Given this prevelance, many of the fields of an operation are optional. The encodingMask field is a bitmask which indicates which of the components of the operation are present.
The location is encoded as the index of the location within the attribute table.
If the operation has attribues, the index of the operation attribute dictionary within the attribute table is encoded.
If the operation has results, the number of results and the indexes of the result types within the type table are encoded.
If the operation has operands, the number of operands and the value index of each operand is encoded. This value index is the relative ordering of the definition of that value from the start of the first ancestor isolated region.
If the operation has successors, the number of successors and the indexes of the successor blocks within the parent region are encoded.
If the operation has regions, the number of regions and if the regions are isolated from above are encoded together in a single varint. Afterwards, each region is encoded inline.
region {
numBlocks: varint,
numValues: varint?,
blocks: block[]
}
A region is encoded first with the number of blocks within. If the region is non-empty, the number of values defined directly within the region are encoded, followed by the blocks of the region.
block {
encoding: varint, // (numOps << 1) | (hasBlockArgs)
arguments: block_arguments?, // Optional based on encoding
ops : op[]
}
block_arguments {
numArgs: varint?,
args: block_argument[]
}
block_argument {
typeIndex: varint,
location: varint
}
A block is encoded with an array of operations and block arguments. The first field is an encoding that combines the number of operations in the block, with a flag indicating if the block has arguments.